Alma Mater Studiorum · Università di Bologna

Scuola di Scienze
Corso di Laurea Magistrale in Fisica

Image Quality and Dose Evaluation

of Filtered Back Projection Versus
Iterative Reconstruction Algorithm
in Multislice Computed
Tomography

Relatore: Presentata da:

Prof. Maria Pia Morigi Daniele Pesolillo

Correlatore:
Prof. Luisa Pierotti

Sessione III
Anno Accademico 2013/2014
Alla mia famiglia........che mi ha dato una seconda possibilità.
Contents

Abstract 2

Introduction 5

1 Basics of Computed-Tomography Technology 7

1.1 A brief history . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2 Fundamentals principles and Design . . . . . . . . . . . . . . . 9
1.3 Acquisition Modes . . . . . . . . . . . . . . . . . . . . . . . . 15
1.3.1 Configurations . . . . . . . . . . . . . . . . . . . . . . . 15
1.3.2 X-ray tube in various generations of CT . . . . . . . . 16
1.3.3 Axial CT Scanning vs Helical CT Scanning . . . . . . . 20
1.3.4 Difference between SSCT and MSCT . . . . . . . . . . 24

2 Reconstruction Algorithms 29
2.1 Theoretical background . . . . . . . . . . . . . . . . . . . . . . 29
2.1.1 Reconstruction Procedure . . . . . . . . . . . . . . . . 31
2.2 State of the art . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.2.1 GE Healthcare . . . . . . . . . . . . . . . . . . . . . . 41
2.2.2 Siemens Healthcare . . . . . . . . . . . . . . . . . . . . 43
2.2.3 Toshiba Medical System . . . . . . . . . . . . . . . . . 45
2.2.4 Philips Healthcare . . . . . . . . . . . . . . . . . . . . 46
2.2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . 48

3 Image Quality Assessment 49

3.1 Noise Power Spectrum Analysis . . . . . . . . . . . . . . . . . 49
3.1.1 Materials and Methods . . . . . . . . . . . . . . . . . . 50
3.2 Modulation transfer function analysis . . . . . . . . . . . . . . 75
3.2.1 Test Device and MTF processing . . . . . . . . . . . . 77

4 Low-Contrast Detectability 87
4.1 Catphan 600 phantom . . . . . . . . . . . . . . . . . . . . . . 87

i
ii CONTENTS

4.1.1 Contrast to noise ratio with Catphan 600 . . . . . . . . 89

4.1.2 CIRS 061 phantom . . . . . . . . . . . . . . . . . . . . 98
4.1.3 Contrast to noise ratio with CIRS 061 . . . . . . . . . 100

5 Dose Assessment 109

5.1 Computed Tomography Dose Index . . . . . . . . . . . . . . . 109
5.2 Dose Length Product . . . . . . . . . . . . . . . . . . . . . . . 111
5.3 CTDI and DLP Measurements . . . . . . . . . . . . . . . . . . 112
5.3.1 Protocols and Method . . . . . . . . . . . . . . . . . . 113

6 Conclusions 117

Appendix 118

A Images, tables and surface plot 119

A.1 Body Phantom images . . . . . . . . . . . . . . . . . . . . . . 119
A.2 Head Phantom images . . . . . . . . . . . . . . . . . . . . . . 121
A.3 NNPS tables both FBP and iterative algorithm, head and
body phantoms. 3D surface plot. . . . . . . . . . . . . . . . . 123

B CNR Plot → Catphan 600 127

C CNR plot → CIRS 061 133

Bibliography 133
List of Figures

1.1 Hounsfield’s sketch (left), Lithograph of Hounsfield Original Test Lathe, presented to
the author on the late 1970s (right). . . . . . . . . . . . . . .. . . . . . 8
1.2 A modern CT Scanner, Philips Brilliance 64 CT Scanner. . . . . . .. . . . . . 9
1.3 Sample of garnet-biotite-kyanite schist. . . . . . . . . . . . . .. . . . . . 11
1.4 Gantry virtual view. . . . . . . . . . . . . . . . . . . . .. . . . . . 12
1.5 Gantry External view. . . . . . . . . . . . . . . . . . . . .. . . . . . 13
1.6 Gantry Internal view. . . . . . . . . . . . . . . . . . . . .. . . . . . 13
1.7 Some of the most common configurations for CT scanners. . . . . .. . . . . . 16
1.8 A representation of first generation CT scanner (Parallel Beam, Translate-Rotate). . 17
1.9 A representation of second generation CT scanner (Fan Beam, Translate-Rotate). . . 18
1.10 A representation of third generation CT scanner (Fan Beam, Rotate only). . . . . . 19
1.11 A representation of fourth generation CT scanner (Fan Beam, stationary circular
detector). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.12 Artistic representation of axial CT. . . . . . . . . . . . . . . . . . . . . 21
1.13 Comparison between higher pitch and lower pitch[7]. . . . . . . . . . . . . . 23
1.14 Artistic representation of spiral CT. . . . . . . . . . . . . . . . . . . . . 23
1.15 Spiral Slice Sensitivity Profile (SSP) of SSCT in Spiral Mode (LEFT); As the pitch
increases, SSP curves deviate more and more from an ideal square wave (-0.5 to 0.5)
more similar to conventional (non-spiral) CT. Spiral Slice Sensitivity Profile of MSCT
in Spiral Mode (RIGHT); Fractional pitch of multislice leads to better approximation
of SSCT, more similar to ideal square wave (-0.5 to 0.5)[1]. . . . . . . . . . . . 26
1.16 SSCT arrays containing single, long elements along z-axis (Left). MSCT arrays with
several rows of small detector elements (Right)[8]. . . . . . . . . . . . . . . . 26
1.17 Diagrams of various 16-slice detector designs (in z-direction). Innermost elements can
be used to collect 16 thin slices or linked in pairs to collect thicker slices[8]. . . . . 27
1.18 Diagrams of various 64-slice detector designs (in z-direction). Most designs lengthen
arrays and provide all submillimeter elements. Siemens scanner uses 32 elements and
dynamic-focus x-ray tube to yield 2 measurements per detector[8]. . . . . . . . . 27

2.1 Radon Transform. . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

2.2 . . . . . . 30
The Shepp-Logan head phantom (left) and its Radon Transform (right).

iii
iv LIST OF FIGURES

2.3 The Fourier slice theorem. In the spatial domain, each view is found by integrating
the image along rays at a particular angle. In the frequency domain, the spectrum of
each view is a one dimensional slice of the two dimensional image spectrum. . . . . 33
2.4 Backprojection reconstructs an image by taking each view and smearing it along the
path it was originally acquired. The resulting image is a blurry version of the correct
image. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.5 Filtered backprojection reconstructs an image by filtering each view before backpro-
jection. This removes the blurring seen in simple backprojection, and results in a
. . . . . . . . . . . . . . . 35
mathematically exact reconstruction of the image.
2.6 The basic workflow of an FBP. . . . . . . . . . . . . . . . . . . . . . . 36
2.7 Schematic view of the iterative reconstruction process[13]. The volume estimated is
initiated either with an empty image or, if available, an FBP reconstruction. If a stop
criterion is matched,the loop is terminated and the current volumetric image becomes
the final volumetric image. . . . . . . . . . . . . . . . . . . . . . . . . 36
2.8 Selection of the most prominent iterative reconstruction algorithms. . . . . . . . 37
2.9 The basic workflow of an ASIR algorithm. . . . . . . . . . . . . . . . . . . 39
2.10 The basic workflow of an MBIR algorithm. . . . . . . . . . . . . . . . . . 39
2.11 Statistical and Model-based Iterative Reconstruction Algorithms Developed by Major
Computed Tomography Manufacturers[16]. . . . . . . . . . . . . . . . . . 40
2.12 In a 15-year-old patient presenting to the emergency department to rule out appen-
dicitis, low-dose scan with FBP reconstruction was noisier than follow-up imaging
using the same dose with ASiR reconstruction[18]. . . . . . . . . . . . . . . 41
2.13 Liver metastasis visualized with VEO. The right image is less noisy than other[19]. . 42
2.14 Comparison between standard protocol (FBP) and Iterative reconstruction in image
space [see www.healthcare.siemens.com]. . . . . . . . . . . . . . . . . . . 43
2.15 Comparison between FBP and Sinogram Reconstruction. Image noise decrease with-
out loss of resolution in the right image [see www.healthcare.siemens.com]. . . . . . 44
2.16 In the left image we see noise reduction with AIDR3D. In the right image is shown
the workflow for dose reduction AIDR 3D [see toshibamedicalsystems.com]. . . . . 45
2.17 Image enhancement of an abdomen using IMR [see www.healthcare.philips.com]. . . 46
2.18 Summary of noise reduction and artifact prevention capabilities provided by each
reconstruction generation (left). Adapting dose reduction and spatial resolution based
on the clinical indication (right) [see www.healthcare.philips.com]. . . . . . . . . 47

3.1 Philips Phantom used for acquisition. Body phantom (LEFT SIDE) and head phan-
tom (RIGHT SIDE). . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.2 Convolution kernel for body phantom. . . . . . . . . . . . . . . . . . . . 52
3.3 Spatial frequency (mm−1 ) and radially NNPS values (mm2 ) (LEFT). Body phantom
image and ROI utilized for calculate the Normalized Noise Power Spectrum (RIGHT). 53
3.4 Values of NNPS calculate for all seven slice with FBP algorithm and Convolution
kernel A. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
LIST OF FIGURES v

3.5 2D image of the NNPS, note the circular symmetry. . . . . . . . . . . . . . . 54

3.6 Normalized noise power spectrum for different iDose levels. An image reconstructed
using this filter, producing noise texture with low spatial frequency noise. See Ap-
pendix A for phantom’s image. . . . . . . . . . . . . . . . . . . . . . . 55
3.7 Normalized noise power spectrum for different iDose levels. First idose level will have
two peaks while iDose6 will have one peak shifted at lower frequencies. See Appendix
A for phantom’s image. . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.8 Normalized noise power spectrum for different iterative reconstruction levels. For
higher levels the peaks from 0.25 mm−1 to 0.35 mm−1 disappear. See Appendix A
for phantom’s image. . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.9 All curves shifted by high frequencies than other kernels. See Appendix A for phan-
tom’s image. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.10 Shape of Normalized Noise Power Spectrum of FBP and reconstruction levels. The
curves tends to zero more slowly than the others filter. See Appendix A for phantom’s
image. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.11 Trend of noise power spectrum for all reconstruction kernels. Radially averaged nor-
malized NPS curves show how noise texture is manifested in the NPS. . . . . . . 60
3.12 Trend of noise power spectrum for all reconstruction kernels. Note the differences at
∼ 0.45 mm−1 and ∼ 0.3 mm−1 . . . . . . . . . . . . . . . . . . . . . . 61
3.13 Trend of noise power spectrum for all reconstruction kernels. Note the lowest noise
power for kernel D (FBP) after 0.55 mm−1 . . . . . . . . . . . . . . . . . . 63
3.14 Noise texture fluctuations of Filtered Back Projection algorithm with convolution
kernel A (LEFT). Noise texture fluctuations of Iterative reconstruction algorithm
(iDose, level 4) with same convolution kernel of FBP (RIGHT). The teflon insert is
. . . . . . . . . . . 64
not affected by noise texture and reconstruction algorithm.
3.15 Convolution kernel for Head Phantom. . . . . . . . . . . . . . . . . . . . 65
3.16 Spatial frequency (mm−1 ) and radially NNPS values (mm2 ) (LEFT). Head phan-
tom image and region of interest utilized for calculate the Normalized Noise Power
Spectrum (RIGHT). . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.17 Values of NNPS calculate for all seven slice with FBP algorithm and Convolution
kernel A. Note the different values of spatial frequency for head phantom compared
to body phantom. . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.18 Trend of noise power spectrum for all reconstruction kernels, except kernels UB-EB.
Radially averaged normalized NPS curves show how noise texture is manifested in the
NPS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.19 Comparison between smooth convolution kernels for head acquisition. UB improves
bone-brain interface and no effect on HU values; EB head scans only and increased
to observed HU values (not shown here). . . . . . . . . . . . . . . . . . . 69
3.20 Trend of noise power spectrum for all reconstruction kernels, except kernels UB-EB.
Radially averaged normalized NPS curves show how noise texture is manifested in the
NPS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
vi LIST OF FIGURES

3.21 Comparison between smooth convolution kernels for head acquisition. UB improves
bone-brain interface and no effect on HU values; EB head scans only and increased
to observed HU values (not shown here). . . . . . . . . . . . . . . . . . . 71
3.22 Trend of noise power spectrum for all reconstruction kernels, except kernels UB-EB.
Radially averaged normalized NPS curves show how noise texture is manifested in the
NPS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.23 Comparison between smooth convolution kernels for head acquisition. UB improves
bone-brain interface and no effect on HU values; EB head scans only and increased
to observed HU values (not shown here). . . . . . . . . . . . . . . . . . . 74
3.24 Noise texture fluctuations of Filtered Back Projection algorithm with convolution
kernel DH (LEFT). Noise texture fluctuations of Iterative reconstruction algorithm
(iDose, level 4) with same convolution kernel of FBP (RIGHT). . . . . . . . . . 74
3.25 (a)Input images defining the point-spread function, the line-spread function and the
edge-spread function.(b) Simulated degraded-output images showing raw image data
used for the measurements of the PSF, LSF and ESF. The blurring seen in these
functions is due to the imperfect resolution properties of the imaging system being
characterized.(c) Graphs showing the actual PSF, LSF and ESF. The PSF is a 2D
function, and the LSF and ESF are 1D functions. . . . . . . . . . . . . . . . 77
3.26 Phantom image corresponding to the Philips head phantom using a typical adult
head protocol. Image window and level have been adjusted to show bead point source
within the ROI (LEFT). Modulation transfer function reconstructed with kernel A;
the spatial frequencies at 10% and 50% are shown (RIGHT). . . . . . . . . . . 79
3.27 MTF values with FBP and iterative algorithm using kernel A (LEFT). MTF plot and
values of spatial frequency at 10% and 50% (RIGHT.) . . . . . . . . . . . . . 80
3.28 MTF values with FBP and iterative algorithm using kernel EB (LEFT). MTF plot
and values of spatial frequency at 10% and 50% (RIGHT.) . . . . . . . . . . . 80
3.29 MTF values with FBP and iterative algorithm using kernel UB (LEFT). MTF plot
and values of spatial frequency at 10% and 50% (RIGHT.) . . . . . . . . . . . 80
3.30 MTF values with FBP and iterative algorithm using kernel C (LEFT). MTF plot and
values of spatial frequency at 10% and 50% (RIGHT.) . . . . . . . . . . . . . 81
3.31 MTF values with FBP and iterative algorithm using kernel DH (LEFT). MTF plot
and values of spatial frequency at 10% and 50% (RIGHT.) . . . . . . . . . . . 81
3.32 MTF for filtered back projection and level 1 of iterative reconstruction algorithm. . . 82
3.33 MTF for level 2, level 3 and level 4 of iterative reconstruction algorithm. . . . . . 83
3.34 MTF for level 5 of iterative reconstruction algorithm. . . . . . . . . . . . . . 84
3.35 Modulation transfer function at 10% compared standard deviation for all convolution
kernels. The filter DH has a value greater than other. . . . . . . . . . . . . . 86
3.36 Modulation transfer function at 50% compared standard deviation for all convolution
kernels. The filter DH has a greater value than other. . . . . . . . . . . . . . 86
LIST OF FIGURES vii

4.1 Catphan 600 phantom (LEFT). CTP515 low contrast module with supra-slice and
. . . . . . . . . . . . . . . . . . . . . 88
subslice contrast targets (RIGHT).
4.2 Catphan phantom analysis using nominal contrast of 1%. . . . . . . . . . . . . 91
4.3 Contrast to noise ratio for iterative and standard algorithm. We can see the differences
between the values; they are very similar between kernel UB and EB. . . . . . . . 92
4.4 Trends of contrast to noise ratio at varying levels of reconstruction. . . . . . . . . 93
4.5 Catphan phantom analysis using nominal contrast of 0.5%. . . . . . . . . . . . 94
4.6 Contrast to noise ratio for iterative and standard algorithm. We can see the small
differences between the filter C and EB. . . . . . . . . . . . . . . . . . . . 95
4.7 Trends of contrast to noise ratio at varying levels of reconstruction. . . . . . . . . 95
4.8 Catphan phantom analysis using nominal contrast of 0.3%. . . . . . . . . . . . 96
4.9 Contrast to noise ratio for iterative and standard algorithm. We can see the differences
between the filters C, EB and UB. . . . . . . . . . . . . . . . . . . . . . 97
4.10 Trends of contrast to noise ratio at varying levels of reconstruction. . . . . . . . . 98
4.11 Spiral CIRS phantom, internal view (LEFT). Phantom contains spherical objects;
these spheres are placed in three rows. Each row contains spheres that were origi-
nally designed to be 20, 10, and 5 HU below background (designed to equal liver; no
attenuation given (RIGHT). . . . . . . . . . . . . . . . . . . . . . . . 99
4.12 Cirs 061 phantom analysis using nominal contrast 2%. . . . . . . . . . . . . . 101
4.13 Contrast to noise ratio for iterative and standard algorithm. . . . . . . . . . . 102
4.14 Trends of contrast to noise ratio at varying levels of reconstruction. . . . . . . . . 103
4.15 Cirs 061 phantom analysis using nominal contrast 1%. . . . . . . . . . . . . . 103
4.16 Contrast to noise ratio for iterative and standard algorithm. . . . . . . . . . . 104
4.17 Trends of contrast to noise ratio at varying levels of reconstruction. . . . . . . . . 105
4.18 Cirs 061 phantom analysis using nominal contrast 0.5%. . . . . . . . . . . . . 105
4.19 Contrast to noise ratio for iterative and standard algorithm. . . . . . . . . . . 106
4.20 Trends of contrast to noise ratio at varying levels of reconstruction. . . . . . . . . 107

5.1 Illustration of the term ”Computed Tomography Dose Index”. . . . . . . . . . . 110

5.2 Illustration of the term ”Dose Length Product”. . . . . . . . . . . . . . . . 112
5.3 Phantom kit to evaluate CTDI (LEFT) and internal view with pencil chamber (RIGHT). 112

A.1 Filtered Back Projection reconstruction with kernel A (LEFT). Iterative reconstruc-
tion with kernel A (RIGHT). . . . . . . . . . . . . . . . . . . . . . . . 119
A.2 Filtered Back Projection reconstruction with kernel B (LEFT). Iterative reconstruc-
tion with kernel B (RIGHT). . . . . . . . . . . . . . . . . . . . . . . . 119
A.3 Filtered Back Projection reconstruction with kernel C (LEFT). Iterative reconstruc-
tion with kernel C (RIGHT). . . . . . . . . . . . . . . . . . . . . . . . 120
A.4 Filtered Back Projection reconstruction with kernel D (LEFT). Iterative reconstruc-
tion with kernel D (RIGHT). . . . . . . . . . . . . . . . . . . . . . . . 120
viii LIST OF FIGURES

A.5 Filtered Back Projection reconstruction with kernel DH (LEFT). Iterative reconstruc-
tion with kernel DH (RIGHT). . . . . . . . . . . . . . . . . . . . . . . 120
A.6 Filtered Back Projection reconstruction with kernel A (LEFT). Iterative reconstruc-
tion with kernel A (RIGHT). . . . . . . . . . . . . . . . . . . . . . . . 121
A.7 Filtered Back Projection reconstruction with kernel UB (LEFT). Iterative reconstruc-
tion with kernel UB (RIGHT). . . . . . . . . . . . . . . . . . . . . . . 121
A.8 Filtered Back Projection reconstruction with kernel EB (LEFT). Iterative reconstruc-
tion with kernel EB (RIGHT). . . . . . . . . . . . . . . . . . . . . . . 121
A.9 Filtered Back Projection reconstruction with kernel C (LEFT). Iterative reconstruc-
tion with kernel C (RIGHT). . . . . . . . . . . . . . . . . . . . . . . . 122
A.10 Filtered Back Projection reconstruction with kernel D (LEFT). Iterative reconstruc-
tion with kernel D (RIGHT). . . . . . . . . . . . . . . . . . . . . . . . 122
A.11 Filtered Back Projection reconstruction with kernel DH (LEFT). Iterative reconstruc-
tion with kernel DH (RIGHT). . . . . . . . . . . . . . . . . . . . . . . 122
A.12 Average NNPS both Filtered Back Projection and Iterative reconstruction algorithm,
head phantom. Spatial Frequency [mm−1 ], NNPS [mm2 ]. . . . . . . . . . . . 123
A.13 Average NNPS both Filtered Back Projection and Iterative reconstruction algorithm;
body phantom. Spatial Frequency [mm−1 ], NNPS [mm2 ]. . . . . . . . . . . . 123
A.14 Noise texture fluctuations of Filtered Back Projection algorithm with convolution
kernel B (LEFT). Noise texture fluctuations of Iterative reconstruction algorithm
(iDose, level 4) with same convolution kernel of FBP (RIGHT). The teflon insert is
not affected by noise texture and reconstruction algorithm. Body phantom. . . . . 124
A.15 Noise texture fluctuations of Filtered Back Projection algorithm with convolution
kernel C (LEFT). Noise texture fluctuations of Iterative reconstruction algorithm
(iDose, level 4) with same convolution kernel of FBP (RIGHT). The teflon insert is
not affected by noise texture and reconstruction algorithm. Body phantom. . . . . 124
A.16 Noise texture fluctuations of Filtered Back Projection algorithm with convolution
kernel D (LEFT). Noise texture fluctuations of Iterative reconstruction algorithm
(iDose, level 4) with same convolution kernel of FBP (RIGHT). The teflon insert is
not affected by noise texture and reconstruction algorithm. Body phantom. . . . . 125
A.17 Noise texture fluctuations of Filtered Back Projection algorithm with convolution
kernel DH (LEFT). Noise texture fluctuations of Iterative reconstruction algorithm
(iDose, level 4) with same convolution kernel of FBP (RIGHT). The teflon insert is
not affected by noise texture and reconstruction algorithm. Body phantom. . . . . 125
A.18 Noise texture fluctuations of Filtered Back Projection algorithm with convolution
kernel UB (LEFT). Noise texture fluctuations of Iterative reconstruction algorithm
(iDose, level 4) with same convolution kernel of FBP (RIGHT). Head phantom. . . . 126
A.19 Noise texture fluctuations of Filtered Back Projection algorithm with convolution
kernel EB (LEFT). Noise texture fluctuations of Iterative reconstruction algorithm
(iDose, level 4) with same convolution kernel of FBP (RIGHT). Head phantom. . . . 126
LIST OF FIGURES ix

B.1 CNR values for kernel UB; nominal contrast 1%. .

. . . . . . . . . . . . . . 127
B.2 CNR values for kernel EB; nominal contrast 1%. .
. . . . . . . . . . . . . . 128
B.3 CNR values for kernel A; nominal contrast 1%. . .
. . . . . . . . . . . . . . 128
B.4 CNR values for kernel C; nominal contrast 1%. . .
. . . . . . . . . . . . . . 129
B.5 CNR values for kernel UB; nominal contrast 0.5%. . . . . . . . . . . . . . . 129
B.6 CNR values for kernel EB; nominal contrast 0.5%. . . . . . . . . . . . . . . 130
B.7 CNR values for kernel A; nominal contrast 0.5%. . . . . . . . . . . . . . . . 130
B.8 CNR values for kernel C; nominal contrast 0.5%. . . . . . . . . . . . . . . . 131
B.9 CNR values for kernel UB; nominal contrast 0.3%. . . . . . . . . . . . . . . 131
B.10 CNR values for kernel EB; nominal contrast 0.3%. . . . . . . . . . . . . . . 132
B.11 CNR values for kernel A; nominal contrast 0.3%. . . . . . . . . . . . . . . . 132

C.1 CNR values for kernel A; nominal contrast 2%. . . . . . . . . . . . . . . . . 133

C.2 CNR values for kernel B; nominal contrast 2%. . . . . . . . . . . . . . . . . 134
C.3 CNR values for kernel C; nominal contrast 2%. . . . . . . . . . . . . . . . . 134
C.4 CNR values for kernel A; nominal contrast 1%. . . . . . . . . . . . . . . . . 135
C.5 CNR values for kernel B; nominal contrast 1%. . . . . . . . . . . . . . . . . 135
C.6 CNR values for kernel C; nominal contrast 1%. . . . . . . . . . . . . . . . . 136
C.7 CNR values for kernel A; nominal contrast 0.5%. . . . . . . . . . . . . . . . 136
C.8 CNR values for kernel B; nominal contrast 0.5%. . . . . . . . . . . . . . . . 137
C.9 CNR values for kernel C; nominal contrast 0.5%. . . . . . . . . . . . . . . . 137
Abstract

Il presente lavoro di tesi è stato svolto presso il servizio di Fisica Sanitaria

del Policlinico Sant’Orsola-Malpighi di Bologna.
Lo studio si è concentrato sul confronto tra le tecniche di ricostruzione
standard (Filtered Back Projection, FBP) e quelle iterative in Tomografia
Computerizzata.
Il lavoro è stato diviso in due parti: nella prima è stata analizzata la
qualità delle immagini acquisite con una CT multislice (iCT 128, sistema
Philips) utilizzando sia l’algoritmo FBP sia quello iterativo (nel nostro caso
iDose4 ). Per valutare la qualità delle immagini sono stati analizzati i seguenti
parametri: il Noise Power Spectrum (NPS), la Modulation Transfer Function
(MTF) e il rapporto contrasto-rumore (CNR). Le prime due grandezze sono
state studiate effettuando misure su un fantoccio fornito dalla ditta costrut-
trice, che simulava la parte body e la parte head, con due cilindri di 32 e 20
cm rispettivamente.
Le misure confermano la riduzione del rumore ma in maniera differente
per i diversi filtri di convoluzione utilizzati. Lo studio dell’MTF invece
ha rivelato che l’utilizzo delle tecniche standard e iterative non cambia la
risoluzione spaziale; infatti gli andamenti ottenuti sono perfettamente iden-
tici (a parte le differenze intrinseche nei filtri di convoluzione), a differenza di
quanto dichiarato dalla ditta. Per l’analisi del CNR sono stati utilizzati due
fantocci; il primo, chiamato Catphan 600 è il fantoccio utilizzato per carat-
terizzare i sistemi CT. Il secondo, chiamato Cirs 061 ha al suo interno degli
inserti che simulano la presenza di lesioni con densità tipiche del distretto
addominale. Lo studio effettuato ha evidenziato che, per entrambi i fantocci,
il rapporto contrasto-rumore aumenta se si utilizza la tecnica di ricostruzione
iterativa.
La seconda parte del lavoro di tesi è stata quella di effettuare una val-
utazione della riduzione della dose prendendo in considerazione diversi pro-
tocolli utilizzati nella pratica clinica, si sono analizzati un alto numero di
esami e si sono calcolati i valori medi di CTDI e DLP su un campione di
esame con FBP e con iDose4 . I risultati mostrano che i valori ricavati con

1
2 Abstract

l’utilizzo dell’algoritmo iterativo sono al di sotto dei valori DLR nazionali di

riferimento e di quelli che non usano i sistemi iterativi.
Introduction

Today, ionizing radiation from Computed Tomography (CT) scanners repre-

sents the greatest per capita medical exposure for the population of industri-
alized countries. Although this growth is mainly attributed to the increasing
number of CT examinations, CT dose per examination is still high and re-
mains an important worry.
Academia, industry, and government have responded with efforts to re-
duce the radiation dose required to obtain diagnostic-quality images. Re-
search has shown that some incident cancer cases may be associated with
CT scans. Although the risks for an individual are small, the rapid increase
of CT utilization has created some significant concern over the patient radi-
ation dose.
Automatic dose control comprimes all technical means to adapt the tube
current to the attenuation properties of individual patients. Dose modulation
is the adaptation of the tube current to varying attenuation by the patient
during one revolution of the x-ray source (circular dose modulation) or along
z-axis (longitudinal dose modulation). It results, if adequately designed, in
significantly reduced dose values depending on the body region. Longitudinal
dose modulation aims to ensure a constant noise level regardless of the local
attenuation properties. By doing so, dose will inevitably be increased when
proceeding from the upper abdomen to the pelvis in examinations of the
entire abdomen. Noise, however, is not the only characteristic related to
image quality; in the pelvis, the dose should instead be decreased owing to
the improved inherent contrast which permits an increased noise level.
With increasing recognition of the importance of radiation protection,
dose reduction has become an important issue in CT system development.
In the past decade, several techniques for reducing CT radiation dose have
been developed. The challenge of reducing dose is to maintain image quality
because noise is increased at decreasing exposure level.
Maintaining clinically acceptable image quality at low dose is the goal of
many techniques for reducing radiation dose.
Modern CT systems are equipped with several dose reduction techniques.

3
4 Introduction

These techniques range from hardware, such as a sliding collimator to elimi-

nate unnecessary radiation exposure due to overranging, to algorithms such
as improved filtered back projection (FBP) and iterative reconstruction (IR).
One step has been CT manufactures implementing iterative reconstruction
methods that for certain clinical tasks can improve dose efficiency over the
conventional reconstruction method, filtered back projection.
While analytical algorithms such as FBP are based on only a single recon-
struction, iterative algorithms use multiple repetitions in which the current
solution converges towards a better solution. As a consequence, the compu-
tational demands are much higher.
Due to the exponential growth of computer technology proposed by Moore’s
law, which is holding since the 1970s, and the computational capacities avail-
able in a modern processor or graphics adapter the usage of IR methods
has become a realistic option, with reconstruction times acceptable for clini-
cal workflow. Nevertheless, new algorithmic innovations are needed because
computational demands have increased due to the fact that image resolution
was improved, acquisition times were greatly reduced; CT scans became part
of the clinical routine and modern IR algorithms gained additional complex-
ity.
Iterative reconstruction algorithms may allow a notable dose reduction
due to a more precise modeling of the acquisition process. This is expected
to support the trend towards continued dose reduction, which is considered
a necessity in view of the increasing number of CT examinations in clinical
routine. In addition, iterative methods with the ability to include various
physical models represent a more intuitive and natural way of image re-
construction. Statistical reconstruction methods, for example, model the
counting statistics of detected photons by respective weighting of the mea-
sured rays. Other implementations include the modeling of the acquisition
geometry or incorporate further information on the x-ray spectra used for
improving the simulation of the acquisition process.
The performance of CT scanners is frequently measured using physical
phantoms targeting metrics that quantify radiation dose and image qual-
ity. These performance evaluations are used to perform quality control tests,
develop clinical protocols, accredit devices, or assess the utility of new scan-
ner designs and algorithms. Currently, a number of useful phantoms ex-
ist that are targeted to the measurements of image noise, spatial resolu-
tion, Hounsfield Unit accuracy, alignement, and detectability. Those include
the ACR Accreditation Phantom, the Catphan Phantom and manufacturer-
supplied quality control phantoms. Another industry standard phantom, the
CT dose index (CTDI) phantom, is used to parameterize CT dose.
Even though these phantoms are of significant value, they fail to capture
Introduction 5

the performance aspects of some of the key yet common technological at-
tributes of modern CT systems: image quality performance as a function of
body size, tube current modulation, and iterative reconstruction.
The purpose of this study is to evaluate the image quality and dose as-
sessment by using a filtered back projection and iterative reconstruction al-
gorithm. This thesis has been divided in two parts: initially, we analyze the
noise power spectrum and the modulation transfer function in both standard
and iterative reconstruction. Next we focus on low contrast detectability
by make use of two different phantoms. The second part allows us to ana-
lyze dose assessment in CT imaging and compare the obtained results with
national DLR. This thesis is organized as follows:

• Chapter 1 : A brief historical introduction to Computed Tomography;

fundamentals principles and design; acquisition modes and different
configurations; comparison between axial and helical scanning; single-
slice and multi-slice technologies.

• Chapter 2 : Theoretical background of reconstruction algorithms; re-

construction procedure; standard and iterative techniques; state of the
art of various manufacturers.

• Chapter 3 : Noise power spectrum analysis and noise reduction with IR;
phantoms and methods used; Modulation transfer function analysis;
test device and processing.

• Chapter 4 : Low-contrast analysis with Catphan 600 and Cirs 061 phan-
toms.

• Chapter 5 : Dose assessment; theoretical basis of Computed Tomogra-

phy Dose Index and Dose Length Product; comparison between FBP,
IR and national LDR.

• Chapter 6 : Conclusions and future evaluations.

Chapter 1

Basics of
Computed-Tomography
Technology

1.1 A brief history

For the 75 years of x-ray imaging, the detector used in diagnostic radiology,
such as radiographic film or image intensifiers, provided reasonably good
visualization of high-contrast objects.
However, their ability to record small differences in trasmitted x-ray sig-
nals was limited. Several factors contributed to the inability to resolve low-
contrast signals. First, large-area detectors record a large amount of scattered
radiation, making small differences in x-ray trasmission difficult to resolve.
Second, the superposition of the patient’s three dimensional information onto
a two-dimensional detector obscures low-contrast information.
Introduced clinically in the early 1970s, x-ray computed tomography (CT)
overcame many of the difficulties encountered in using large-area detectors.
First, the sequential irradiation of slabs of tissues and collimation at the
detector markedly reduced the amount of scattered radiation measured. Sec-
ond, the reconstruction of a tomographic image eliminated much of the prob-
lem of overlapping anatomy.
X-ray CT was the first imaging modality that allowed physicians to see the
internal structure of a three-dimensional object in cross-section1 [1].
CT differs from the more conventional x-ray tomography in that one uses
digital or computer techniques to restore the slice of interest rather than the
1
CT was rapidly accepted into clinical practice because of its tomographic nature and
superior contrast resolution.

7
8 Basics of Computed-Tomography Technology

analog techniques of deliberately casting unwanted information into out of

focus planes on a film moving in a complex prescribed geometrical pattern
with the x-ray tube.
The first clinically useful Computed Tomography system was pioneered
by Godfrey Hounsfield Fig.[1.1] of EMI Ltd. in England. This system was
installed in 1971 in the Atkinson Morley Hospital near London. The EMI
scanner arrived on the scene with an impact not unlike that of x-ray systems
following Roentgen’s discovery in 1895. The scanner developed by Hounsfield
in his laboratory took several hours to acquire the raw data for a single scan
or “slice”and took days to reconstruct a single image from this raw data.

Figure 1.1: Hounsfield’s sketch (left), Lithograph of Hounsfield Original Test Lathe, presented to the
author on the late 1970s (right).

By the 1975 EMI were marketing a body scanner, the CT5000, the first
of which was installed at Northwick Park Hospital in London. The first body
scanner in the USA was installed at the Mallinkrodt Institute and had its
first clinical use in October 1975. By this time, scan time had been reduced
to 20 seconds, for a 320x320 image matrix.
The mid-1970s were a time of rapid development in CT: 1976 saw 17
companies offering scanners, with scan times down to 5 seconds in some
cases. By 1978, there was an installed base of around 200 scanners in the
USA, image matrix size were up to 512x512 and some models of scanner
had the capability of ECG-triggered scans. By the end of the 1970s the
importance of CT scanning to medicine was clear: Hounsfield and Cormack
received the Nobel Prize for medicine in 1979.
The 1980s saw incremental development of CT scanner technology: short
scan times and matrix sizes, until by the late 1980s scan time were down to
only 3 seconds. Development continued through the 1990s, with the introduc-
tion of spiral scanning in the early 1990s and the development of multi-slice
scanners, with 4-slice scanners and 0.5 seconds scan times being ’state of the
art’ by the end of the century.
1.2 Fundamentals principles and Design 9

Development of CT scanner technology continued through the early years

of 21st century, particularly with multi-slice scanners. High-end scanners
were offering up to 320 slices, dual-source and dual-energy x-ray sources and
iterative reconstruction algorithm.
The latest multi-slice CT systems can collect up to 640 slices of data in
about 300 ms and reconstruct a 512× 512 matrix image from millions of data
points in less than a second. An entire chest can be scanned in five to ten
seconds using the most advanced multi-slice CT system.
During its 40-year history, CT has made great improvements in speed,
patient comfort, and resolution. A CT scan times have gotten faster, more
anatomy can be scanned in less time. Faster scanning helps to eliminate
artifacts from patient motion such as breathing or peristalsis. Tremendous
research and development has been made to provide excellent image quality
for diagnostic confidence at the lowest possible x-ray dose[2].

Figure 1.2: A modern CT Scanner, Philips Brilliance 64 CT Scanner.

1.2 Fundamentals principles and Design

Computed Tomography (CT) is a non invasive medical examination or pro-
cedure that utilized specialized x-ray equipment to produce cross-sectional
images of the body. Each cross-sectional images represents a “slices”of the
person being imaged.
These cross-sectional images are used for a variety of diagnostic and ther-
apeutic purposes. CT scans can be performed on every region of the body
for a variety of reasons (e.g., diagnostic, treatment planning, interventional).
10 Basics of Computed-Tomography Technology

CT images of internal organs, bones, soft tissue, and blood vessels provide
greater clarity and more details than conventional x-ray images, such as a
chest x-ray.
The value in a CT slice image correspond to x-ray attenuation, which
reflects the proportion of x-rays scattered or absorbed as they pass through
each voxel. X-ray attenuation is primarily a function of x-ray energy and the
density and composition of the material being imaged.
Tomographic imaging consists of directing x-rays at an object from mul-
tiple orientations and measuring the decrease in intensity along a series of
linear paths. This decrease is characterized by Lambert-Beer’s Law, which
describes intensity reduction as a function of x-ray energy, path length, and
material linear attenuation coefficient. A specialized algorithm [see Chapter
2] is then used to reconstruct the distribution of x-ray attenuation in the
volume being imaged[3][4].
The simplest form of Lambert-Beer’s law for a monochromatic x-ray beam
through a homogeneus material is

I = I0 exp[−µx] (1.1)

where I0 and I are the initial and the final x-ray intensity, µ is a material’s
linear attenuation coefficient and x is the length of the x-ray path. If there
are multiple materials, the equation becomes
hX i
I = I0 exp (−µi xi ) (1.2)
i

where each increment i reflects a single material with attenuation coefficient

µi with linear extent xi . In a well-calibrated system using a monochromatic
x-ray source (i.e. synchrotron or gamma-ray emitter) this equation can be
solved directly.
If a polychromatic x-ray source is used, to take into account the fact that
the attenuation coefficient is a strong function of x-ray energy, the complete
solution would require solving the equation over the range of the x-ray energy
(E) spectrum utilized
Z hX i
I = I0 (E) exp (−µi (E)xi ) dE (1.3)
i

However, such a calculation is usually problematic, as most reconstruction

strategies solve for a single µ value at each spatial position. In such cases, µ
is taken as an effective linear attenuation coefficient, rather than an absolute.
This complicates absolute calibration, as effective attenuation is a function of
1.2 Fundamentals principles and Design 11

both the x-ray spectrum and the properties of the scan object. It also leads
to beam-hardening artifacts: changes in image value caused by preferential
attenuation of low-energy x-rays[5].
There are a number of methods by which the x-ray attenuation data can
be converted into an image. The most frequent approach in CT imaging
is called “filtered backprojection”[see Chapter 2], in which the linear data
acquired at each angular orientation are convolved with a specially designed
filter and then backprojected across a pixel field at the same angle.
This principle is illustrated in Fig.[1.3]. A hand sample of garnet-biotite-
kyanite schist (top left) is rotated, and its midsection is imaged with a planar
fan beam (blue). The attenuation of x-rays by the sample as it rotates is
shown in the upper right; the more attenuation there is along a beam path
leading from the point source (bottom) to the linear detector (top), the fewer
x-rays reach the detector. The data collected at each angle are compiled in
the bottom right. In this image the horizontal axis corresponds to detector
channel, and the vertical axis corresponds to rotation angle (or time), and
brightness corresponds to the extent of x-ray attenuation. The resulting
image is called a sinogram, as any point in the original object corresponds to
a sine curve. After data acquisition is complete, reconstruction begins. Each
row of the sinogram is first convolved with a filter, and projected across the
pixel matrix (bottom right) along the angle at which it was acquired. Once
all angles have been processed, the image is complete.

Figure 1.3: Sample of garnet-biotite-kyanite schist.

CT-Scanner hardware is designed to determine effective x-ray attenua-

tion coefficients at each point within a volume of interest from transmission
12 Basics of Computed-Tomography Technology

measurements acquired at multiple angles through the object. A set of trans-

mission measurements through the object at a given angle is known as a
projection. This projection measurements are mathematically combined to
form a two-dimensional representation of a three-dimensional object.
So, while a typical digital image is composed of pixels (picture elements), a
CT slice image is composed of voxels (volume elements).
The scanner is made up of three primary systems, including the gantry,
the computer, and the operating console. Each of these is composed of
various subcomponents.

Figure 1.4: Gantry virtual view.

The gantry assembly is the largest of these systems. It is made up of

all the equipment related to the patient, including the patient support, the
positioning couch, the mechanical supports, and the scanner housing. It also
contains the heart of the CT scanner, the x-ray tube, as well as detectors
which respectively generate and detect x-rays.
The gantry is the ’donut’ shaped part of the CT scanner that houses the
components necessary to produce and detect x-rays to create a CT image.
The x-ray tube and detectors are positioned opposite each other and rotate
around the gantry aperture. Continous rotation in one direction without
cable wrap around is possible due to the use of slip rings.
The following images are of a Toshiba Aquilion 16 CT scanner with the
external and internal components of the gantry.
1.2 Fundamentals principles and Design 13

Figure 1.5: Gantry External view.

Figure 1.6: Gantry Internal view.

 14

NUM. Gantry External View Gantry Internal view

1 Gantry Aperture(720mm diameter) X-Ray tube
2 Microphone Filters, collimator, reference detector
3 Sagittal laser alignment light Internal Projector
4 Patient guide lines X-ray tube heat exchanger (oil cooler)
5 X-ray exposure indicator light High voltage generator (0-75 kV)
6 Emergency stop buttons Direct drive gantry monitor
7 Gantry control panels Rotation Control Unit
8 External laser alignment lights Data Acquisition system (DAS)
9 Patient couch Detectors
10 ECG gating monitor Slip rings
11 -None- Detector temperature controller
12 -None- High voltage generator (75-150 kV)
13 -None- Power unit
14 -None- Line noise filter

Table 1.1: Internal and External CT components (Toshiba Aquilion 16 CT scanner).

Basics of Computed-Tomography Technology
1.3 Acquisition Modes 15

1.3 Acquisition Modes

1.3.1 Configurations
Planar Fan Beam Configuration

The diagram in Fig.[1.7] illustrates some of the most common configurations

for CT scanners. In planar beam scanning, x-rays are collimated and mea-
sured using a linear detector array. Typically, slice thickness is determined
by the aperture of the linear array. Collimation is necessary to reduce the
influence of X-ray scatter, which results in spurious additional x-rays reach-
ing the detector from locations not along the source-detector path. Linear
arrays can generally be configured to be more efficient than planar ones, but
have the drawback that they only acquire data for one slice image at a time.

Cone Beam Configuration

In cone-beam scanning, the linear array is replaced by a planar detector,

and the beam is no longer collimated. Data for an entire object, or a con-
siderable thickness of it, can be acquired in a single rotation. The data are
reconstructed into images using a cone-beam algorithm. In general, cone-
beam data are subject to some blurring and distortion the further one goes
from the central plane that would correspond to single-slice acquisition. They
are also more subject to artifacts stemming from scattering if high-energy x-
rays are utilized. However, the advantage of obtaining data for hundreds or
thousands of slices at a time is considerable, as more acquisition time can be
spent at each turntable position, decreasing image noise. In this thesis we
used this configuration to do our acquisitions.

Parallel Beam Configuration

Parallel-beam scanning is done using a specially configured synchrotron beam

line as the x-ray source. In this case, volumetric data are acquired and there
is no distortion. However, the object size is limited by the width of the
x-ray beam; depending on beam line configuration, objects up to 6 cm in
diameter may be imaged. Synchrotron radiation generally has very high
intensity, allowing data to be acquired quickly, but the x-rays are generally
low-energy (< 35 keV), which can preclude imaging samples with extensive
high-Z materials.
16 Basics of Computed-Tomography Technology

Figure 1.7: Some of the most common configurations for CT scanners.

1.3.2 X-ray tube in various generations of CT

The great majority of CT systems use x-ray tubes, although tomography can
also be done using a synchrotron or gamma-ray emitter as a monochromatic
x-ray source. Important tube characteristics are the target material and peak
x-ray energy, which determine the x-ray spectrum that is generated; current,
which determines x-ray intensity; and the focal spot size, which impacts
spatial resolution.
Most CT x-ray detectors utilize scintillators. Important parameters are
scintillator material, size and geometry, and the means by which scintillation
events are detected and counted. In general, smaller detectors provide bet-
ter image resolution, but reduced count rates because of their reduced area
compared to larger ones. To compensate, longer acquisition times are used
to reduce noise levels.
1.3 Acquisition Modes 17

First Generation
CT scanners used a pencil-thin beam of radiation. The images were acquired
by a ”translate-rotate” method in which the x-ray source and the detector
in a fixed relative position move across the patient followed by a rotation of
the x-ray source/detector combination (gantry) by 1 for 180. The thickness
of the slice, typically 1 to 10 mm, is generally defined by pre-patient collima-
tion using motor driven adjustable wedges external to the x-ray tube. This
generation used axial platforms.

Figure 1.8: A representation of first generation CT scanner (Parallel Beam, Translate-Rotate).

Second Generation
The x-ray source changed from the pencil-thin beam to a fan shaped beam.
The ”translate-rotate” method was still used but there was a significant
decrease in scanning time. Rotation was increased from one degree to thirty
degrees. Because rotating anode tubes could not withstand the wear and tear
of rotate-translate motion, this early design required a relatively low output
stationary anode x-ray tube.
The power limits of stationary anodes for efficient heat dissipation were
improved somewhat with the use of asymmetrical focal spots (smaller in the
18 Basics of Computed-Tomography Technology

scan plane than in the z-axis direction), but this resulted in higher radiation
doses due to poor beam restriction to the scan plane. Nevertheless, these
scanners required slower scan speeds to obtain adequate x-ray flux at the
detectors when scanning thicker patients or body parts. This generation
used axial platforms.

Figure 1.9: A representation of second generation CT scanner (Fan Beam, Translate-Rotate).

Third Generation
Designers realized that if a pure rotational scanning motion could be used
rather than the slam-bang translational motion, then it would be possible to
use higher-power (output), rotating anode x-ray tubes and thus improve scan
speeds in thicker body parts in which the 3rd generation become a Rotate-
Rotate geometry.
A typical machine employs a large fan beam such that the patient is com-
pletely encompassed by the fan, the detector elements are aligned along the
arc of a circle centered on the focus of the x-ray tube. The x-ray tube and
detector array rotate as one through 360 degrees, different projections are
obtained during rotation by pulsing the x-ray source, and bow-tie shaped
filters are chosen to suit the body or head shape by some manufacturers to
control excessive variations in signal strength.
1.3 Acquisition Modes 19

Such filters generally attenuate the peripheral part of the divergent fan beam
to a greater extent than the central part. It also helps overcome the effects
of beam hardening and to minimize patient skin dose in the peripheral part
of the field of view.
A number of variants on this geometry have been developed, which in-
clude those based on offsetting the centre of rotation and the use of a flying
focus x-ray tube. This generation used axial/helical platforms.

Figure 1.10: A representation of third generation CT scanner (Fan Beam, Rotate only).

Fourth Generation
Fourth generation of CT scanner uses Rotate-Fixed Ring geometry where
a ring of fixed detectors completely surrounds the patient. The X-ray tube
rotates inside the detector ring through a full 360 degrees with a wide fan
beam producing a single image. Due to the elimination of translate-rotate
motion the scan time is reduced comparable with third generation scanner,
initially, to 10 seconds per slice but the radiographic geometry is poor be-
cause the X-ray tube must be closer to the patient than the detectors, i.e.
the geometric magnification is large also scatter artifact is more than third
generation since they cannot use anti-scatter grid.
20 Basics of Computed-Tomography Technology

The disadvantages of poor geometry noted above have been alleviated

very neatly by the so called nutating geometry. The X-ray tube is external to
the detector ring but slightly out of the detector plane, this change resulted in
increasing both the acquisition speed, and image resolution[6]. The method
of scanning was still slow, because the X-ray tube and control components
interfaced by cable, limiting the scan frame rotation. Further, they were
more sensitive to artifacts because the non-fixed relationship to the x-ray
source made it impossible to reject scattered radiation. This generation use
axial/helical platforms.

Figure 1.11: A representation of fourth generation CT scanner (Fan Beam, stationary circular detec-
tor).

Several other CT scanner geometries which have been development (fifth

and sixth generation) and marketed do not precisely fit the above categories.
In the next section we will see the two basic modes for CT acquisition.

1.3.3 Axial CT Scanning vs Helical CT Scanning

After the third generation, CT technology remained stable until 1987. By
then, CT examinations times were dominated by interscan delays. After each
360 rotation, cables connecting rotating components to the rest of the gantry
required that rotation stop and reverse direction (Slip Ring).
1.3 Acquisition Modes 21

Scanning, breaking and reversal required at least 8-10 s, of which only 1-2
were spent acquiring data. The result were poor temporal resolution and
long procedure times.

Axial (sequential) scanning

In this scan mode, the patient table remains stationary while the tube and
detector array rotate once around the patient, collecting the necessary data
for image recontruction. After one rotation, the patient table is moved along
the z axis to the next position and another set of scan data are acquired.
If projection through the entire organ of interest can be acquired in one
rotation, such as with 16 cm wide detector arrays, then no table translation
is required.
In single detector row, the image thickness is determined primarly by the
collimation of the x-ray beam along the z axis, and one wide detector array
was used to acquire different slice thicknesses.
In multi detector scanners (MDCT), the image thickness is determined
by the detector element dimensions; the data from adjacent detector rows
can be added together to give wider image thickness and a range of different
slice thickness can be acquired simultaneously.

Figure 1.12: Artistic representation of axial CT.

22 Basics of Computed-Tomography Technology

Helical (spiral) scanning

Spiral scanning involves continuous translation of the patient table with con-
tinuous x-ray rotation and data collection. This decreases overall scan time,
and can allow scanning of the entire adult torso within a breath hold. The
major advantage of spiral scanning is the volume of coverage for a given
rotation of x-ray exposure.
With the introduction of spiral scanning, the slice is not so simply defined
by the x-ray collimation; rather, the nature of the moving table requires
interpolation schemes to provide estimates of information within a given slice.
This information acquired in an acquisition which includes information from
the slice above and below the slice interest and then interpolates the data to
establish an effective slice at a given position.
In helical scanning, extra rotations of data acquisition are required at the
beginning and end of the scan in order to provide sufficient data for image
reconstruction at the edges of the prescribed scan range.
Eliminating interscan delays required continuous rotation and the strat-
egy is to continuously rotate and acquire data as the table moving though
the gantry. The resulting trajectory of the tube and detectors relative to the
patient traces out a helical or spiral path. This powerful concept allows for
rapid scans of entire z-axis regions of interest.
Certain concepts associated with helical CT are fundamentally different
from those of axial scanning. One such concept is how fast the table slides
through the gantry relative to the rotation time and slice thicknesses being
acquired. This aspect is referred to as the helical pitch and is defined as the
table movement per rotation divided by the slice thickness[7].
To understand this concept we consider an MSCT scanner with n arrays
that have a thickness T (at isocenter), the beam width a as measured at the
isocenter is given by
a = nT + η (1.4)
where η is the over-beaming that is necessary in MSCT systems. The η
portion of the beam corrsponds to the width of the penumbra on both sides
of the active beam, which extends beyond the edges of the active detector
arrays (nT) to reduce artifacts. Then, the pitch is defined by

b
p= [mm] (1.5)
nT
where b is the ratio of the table feed.
The choice of pitch is examination dependent, involving a trade-off be-
tween coverage and accuracy.
1.3 Acquisition Modes 23

Figure 1.13: Comparison between higher pitch and lower pitch[7].

In single detector row CT, as the pitch is increased, the data sampling
along z is more sparse, and the result image is wider2 [Fig. 1.13]. Image noise
is not affected, however, as the same number of projections is always used to
form an image.
In multi detector row CT, scanners use spiral interpolation algorithms
that are different than those in single detector, and take advantage of the
multiple rings of transmission data. For MDCT, the width of the section
sensitivity profile remains relatively constant as the pitch changes.

Figure 1.14: Artistic representation of spiral CT.

2
If the slice thickness is 10 mm and the table moves 15 mm during one tube rotation,
then the pitch = 1510 = 1.5.
24 Basics of Computed-Tomography Technology

1.3.4 Difference between SSCT and MSCT

The principal difference between single-slice CT (SSCT) and multi-slice CT
(MSCT) are:

• Primary difference is in the design of the detector arrays, as illustrated

in Fig.[1.16].

• Secondary difference is that MSCT offer a potentially thinner slice that

can be achieved by physical and/or electronic collimation.

• MSCT offer a better defined sensitivity profile Fig.[1.15].

Single-Slice CT
SSCT detector arrays are one dimensional; that is, they consist of a large
number (typically 750 or more) of detector elements in a single row across
the irradiated slice to intercept the x-ray fanbeam. In the slice thickness
direction (z-direction), the detectors are monolithic, that is, single elements
long enough (typically about 20 mm) to intercept the entire x-ray beam
width, including part of the penumbra.
In SSCT, slice thickness is determined by prepatient and possibly post-
patient x-ray beam collimators. Generally, the x-ray beam collimation was
designed such that the z-axis width of the x-ray beam at the isocenter (i.e., at
the center of rotation) is the same as the desired slice thickness. (The x-ray
beam width, usually defined as the full width at half maximum (FWHM) of
the z-axis x-ray beam intensity profile).
The interpolation process tends to create slice where the FWHM is often
matched to the nominal slice width, but the area tails of the slice extend the
sensitivity profile significantly into the neighboring slices, and much beyond
the normal slice width.

Multi-Slice CT
In a multi slice CT, the key factor is that the x-ray collimation allows simul-
taneous radiation of several adjoining z-axis slices at the same time. This
significantly enhances x-ray tube utilization. In MSCT, each of the indi-
vidual, monolithic SSCT detector elements in the z-direction is divided into
several smaller detector elements, forming a 2-dimensional array. Rather
than a single row of detectors encompassing the fan beam, there are now
multiple, parallel rows of detectors.
In MSCT, however, slice thickness is determined by detector configura-
tion and x-ray beam collimation. Because it is the length of the individual
1.3 Acquisition Modes 25

detector (or linked detector elements) acquiring data for each of the simulta-
neously acquired slices that limits the width of the x-ray beam contributing
to that slice, this length is often referred to as detector collimation.
The installation of MSCT scanners providing 16 data channels for 16
simultaneously acquired slices began in 2002. In addition to simultaneously
acquiring up to 16 slices, the detector arrays associated with 16-slice scanners
were redesigned to allow thinner slices to be obtained as well.
One potential problem for the multi-slice system is that a wider area
is scanned at one time, and therefore more scattered radiation per slice is
generated affecting deleteriously both image quality and radiation dose. The
collimator and detector design must be optimized for MSCT and may need
to compensate for x-ray movements in the longitudinal direction by allowing
a wider beam than the actual slice thickness, thus impacting deleteriously
on dose buildup from neighboring slices.
Detector arrays for various 16-slice scanner models are illustrated in
Fig.[1.17]. Note that in all of the models, the innermost 16 detector ele-
ments along the z-axis are half the size of the outermost elements, allowing
the simultaneous acquisition of 16 thin slices (from 0.5 mm thick to 0.75
mm thick, depending on the model). When the inner detectors were used to
acquire submillimeter slices, the total acquired z-axis length and therefore
the total width of the x-ray beam ranged from 8 mm for the Toshiba version
to 12 mm for the Philips and Siemens versions. Alternatively, the inner 16
elements could be linked in pairs for the acquisition of 16 thicker slices.
By 2005, 64 slice scanners were announced, and installations by most
manufacturers began. Detector array designs used by several manufacturers
are illustrated in Fig.[1.18]. The approach used by our manufacturer (Philips)
for 64 slice detector array designs was to lengthen the arrays in the z-direction
and provide all submillimeter detector elements: 64 × 0.625 mm (total z-axis
length of 40 mm).
In addition to the simultaneous acquisition of more slices, MSCT x-ray
beam widths can be considerably wider than those for SSCT. Sixteen slice
MSCT beam widths are up to 32 mm; 64-slice beams can be up to 40 mm
wide; and even wider beams are used in systems currently under develop-
ment or in clinical evaluation. A possible consequence is that more scatter
may reach the detectors, compromising low-contrast detection. Generally,
however, the antiscatter septa traditionally used with third generation CT
scanners can be made sufficiently deep to remain effective with MSCT.
26 Basics of Computed-Tomography Technology

Figure 1.15: Spiral Slice Sensitivity Profile (SSP) of SSCT in Spiral Mode (LEFT); As the pitch
increases, SSP curves deviate more and more from an ideal square wave (-0.5 to 0.5) more similar to
conventional (non-spiral) CT. Spiral Slice Sensitivity Profile of MSCT in Spiral Mode (RIGHT); Fractional
pitch of multislice leads to better approximation of SSCT, more similar to ideal square wave (-0.5 to 0.5)[1].

Figure 1.16: SSCT arrays containing single, long elements along z-axis (Left). MSCT arrays with
several rows of small detector elements (Right)[8].
1.3 Acquisition Modes 27

Figure 1.17: Diagrams of various 16-slice detector designs (in z-direction). Innermost elements can
be used to collect 16 thin slices or linked in pairs to collect thicker slices[8].

Figure 1.18: Diagrams of various 64-slice detector designs (in z-direction). Most designs lengthen
arrays and provide all submillimeter elements. Siemens scanner uses 32 elements and dynamic-focus x-ray
tube to yield 2 measurements per detector[8].
Chapter 2

Reconstruction Algorithms

2.1 Theoretical background

X-ray CT imaging is a procedure to get images of thin slice of an unknown
object, such as biological tissue, from the projection data collected by illumi-
nating the object from many different directions using x-ray. The object can
be represented by its distribution of x-ray attenuation coefficient1.2. When a
parallel beam of x-rays propagates through the object, the total attenuation
of the beam can be expressed by a line integral, which is the well-known
Radon transform.

Figure 2.1: Radon Transform.

The Radon transform(2.1) of an image represented by the function f(x,y)

can be defined as a series of line integrals trough f(x,y) at different offsets
from the origin. It is defined as
Z ∞
R(p, τ ) = f (x, px + τ )dx (2.1)
−∞

29
30 Reconstruction Algorithms

where p and τ are the slope and intercepts of the line. A more directly
applicable form of the transform can be defined by using a delta function
Z ∞ Z ∞
R(r, θ) = f (x, y)δ(x cos θ + y sin θ − r)dxdy (2.2)
−∞ −∞

where f(x,y) denotes the object, R(r,θ) denotes the projection data when
the scanning angle is θ and the distance between the projection line and the
origin is r (the perpendicular offset of the line); δ denotes the Dirac delta
function, the term between brackets represents a projection line of x-rays.
The acquisition of data in medical imaging techniques such as MRI, CT and
PET scanners involves a similar method of projecting a beam through the
object, and the data is in a similar form to that described in the eq.(2.2).
The plot of the Radon transform, or scanner data, is referred to as a
sinogram due to its characteristic sinusoid shape. Next figure shows a simple
head phantom and the sinogram created by taking the Radon transform at
intervals of one degree from 0 to 180 degrees.

Figure 2.2: The Shepp-Logan head phantom (left) and its Radon Transform (right).

Unfortunately, the actual data detected by a medical imaging system

does not correspond exactly to the Radon transform of the “true ”image.
In any imaging system, projection data will be corrupted by noise, and the
projections are measured with only limited resolution. The geometry of the
imaging system may differ from the ideal, particularly in transmission tomog-
raphy, where a fan beam imaging system is more easily implemented than a
parallel-beam system.
2.1 Theoretical background 31

2.1.1 Reconstruction Procedure

Computed tomography (CT) reconstruction is computationally demanding,
but by applying the latest high performance processors and advanced soft-
ware programming techniques, it’s possible to reign in processing times. With
the resulting performance gains, CT scanners operate faster while also en-
hancing image quality and increasing acquisition flexibility. These advances
in CT imaging enable radiologists and department managers to improve pa-
tient care, reduce the time to diagnosis and boost the department’s produc-
tivity.
This procedure is very important for CT imaging. The properties of the
final reconstructed image strongly depends upon reconstruction algorithm
used[9]. Image reconstruction has a fundamental impact on image quality
and therefore on radiation dose.
For a given radiation dose it is desirable to reconstruct images with the
lowest possible noise without sacrificing image accuracy and spatial resolu-
tion. Reconstruction algorithms that improve image quality can be translated
into a reduction of radiation dose because images of acceptable quality can
be reconstructed at lower dose[18].
There are many approch to reconstruction algorithms; in the literature a
large number of acronyms can be found. Unfortunately these acronyms are
not always used consistently in the literature and a number of variations and
combinations of different concepts exist.
We divide these in four approaches to calculating the slice image given
the set of its views. These are called CT reconstruction algorithms.

• Solving many simultaneous linear equations.

• Fourier Reconstruction.

• Filtered Back Projection (FBP).

• Iterative Techniques.

Solving many simultaneous linear equations

One equation can be written for each measurement. That is, a particular
sample in a particular profile is the sum of a particular group of pixels in
the image. To calculate N 2 unknown variables (i.e., the image pixel values),
there must be N 2 independent equations, and therefore N 2 measurements.
Most CT scanners acquire about 50% more samples than rigidly required by
this analysis.
32 Reconstruction Algorithms

For example, to reconstruct a 512× 512 image, a system might take 700
views with 600 samples in each view. By making the problem overdetermined
in this manner, the final image has reduced noise and artifacts. The problem
with this first method of CT reconstruction is computation time. Solving
several hundred thousand simultaneous linear equations is a crazy thing.

Fourier Reconstruction
In the spatial domain, CT reconstruction involves the relationship between
a 2D image and its set of 1D views. By taking the 2D Fourier transform of
the image and the one-dimensional Fourier transform of each of its views,
the problem can be examined in the frequency domain. As it turns out, the
relationship between an image and its views is far simpler in the frequency
domain than in the spatial domain. The frequency domain analysis of this
problem is a milestone in CT technology called the Fourier slice theorem 1 .
Fig.[2.3] shows how the problem looks in both the spatial and the fre-
quency domains. In the spatial domain, each view is found by integrating
the image along rays at a particular angle. In the frequency domain, the
image spectrum is represented in this illustration by a two dimensional grid.
The spectrum of each view (a one dimensional signal) is represented by a
dark line superimposed on the grid. As shown by the positioning of the lines
on the grid, the Fourier slice theorem states that the spectrum of a view is
identical to the values along a line (slice) through the image spectrum. For
instance, the spectrum of view 1 is the same as the center column of the
image spectrum, and the spectrum of view 3 is the same as the center row of
the image spectrum. Notice that the spectrum of each view is positioned on
the grid at the same angle that the view was originally acquired. All these
frequency spectra include the negative frequencies and are displayed with
zero frequency at the center.
Fourier reconstruction of a CT image requires three steps. First, the one
dimensional FFT is taken of each view. Second, these view spectra are used
to calculate the two dimensional frequency spectrum of the image, as outlined
by the Fourier slice theorem. Since the view spectra are arranged radially,
and the correct image spectrum is arranged rectangularly, an interpolation
routine is needed to make the conversion. Third, the inverse FFT is taken of
the image spectrum to obtain the reconstructed image[11].
Unfortunately this method suffer from artifacts due to interpolation in
Fourier plane and aliasing2 .
1
The Fourier Slice Theorem describes the relationship between an image and its views
in the frequency domain.
2
All the frequency domain information of a band-limited function is contained in an
2.1 Theoretical background 33

Figure 2.3: The Fourier slice theorem. In the spatial domain, each view is found by integrating the
image along rays at a particular angle. In the frequency domain, the spectrum of each view is a one
dimensional slice of the two dimensional image spectrum.

Filtered Backprojection

FBP is the most important analytical scheme for image recontruction that is
currently widely used on clinical CT scanners because of their computational
efficiency and numerical stability. Many FBP-based methods have been de-
veloped for different generations of CT data-acquisition geometries, from
axial parallel and fan beam CT in the 1970s and 1980s to current multi-slice
helical CT and cone-beam CT with large area detectors.
This method is a modification of an older technique, called backprojection
or simple backprojection. Fig.[2.4] shows that simple backprojection is a
common sense approach, but very unsophisticated. An individual sample
is backprojected by setting all the image pixels along the ray pointing to
the sample to the same value. In less technical terms, a backprojection is
formed by smearing each view back through the image in the direction it was
originally acquired. The final backprojected image is then taken as the sum
of all the backprojected views.
While backprojection is conceptually simple, it does not correctly solve
the problem. As shown in Fig.[2.4b], a backprojected image is very blurry.
A singlepoint in the true image is reconstructed as a circular region that
decreases in intensity away from the center. In more formal terms, the point
spread function of backprojection is circularly symmetric, and decreases as
the reciprocal of its radius.

interval. If this interval is not satisfied, the transform in this interval is corrupted by
contributions from adjacent periods.
34 Reconstruction Algorithms

Figure 2.4: Backprojection reconstructs an image by taking each view and smearing it along the path
it was originally acquired. The resulting image is a blurry version of the correct image.

Filtered backprojection (FBP) is a technique to correct the blurring en-

countered in simple backprojection. As illustrated in Fig.[2.5], each view is
filtered before the backprojection to counteract the blurring PSF. That is,
each of the one-dimensional views is convolved with a one-dimensional filter
to create a set of filtered views. These filtered views are then backprojected
to provide the reconstructed image, a close approximation to the ”correct”
image. In fact, the image produced by filtered backprojection is identical to
the correct image when there are an infinite number of views and an infinite
number of points per view.
Notice how the profiles have been changed by the filter. The image in this
example is a uniform white circle surrounded by a black background. Each of
the acquired views has a flat background with a rounded region representing
the white circle. Filtering changes the views in two significant ways. First,
the top of the pulse is made flat, resulting in the final backprojection creating
a uniform signal level within the circle. Second, negative spikes have been
introduced at the sides of the pulse. When backprojected, these negative
regions counteract the blur.
FBP algorithm and its modified versions for 2D and 3D projection recon-
struction, such as FDK (Feldkamp-Davis-Krees) algorithm have been used
in almost all the fields of straight ray tomography. The projection can be
classified into two types: parallel and fan beam projection.
Since FBP for fan beam tomography is usually obtained by modifying
that for parallel beam tomography. The derivation of FBP algorithm for par-
2.1 Theoretical background 35

Figure 2.5: Filtered backprojection reconstructs an image by filtering each view before backprojec-
tion. This removes the blurring seen in simple backprojection, and results in a mathematically exact
reconstruction of the image.

allel beam tomography is rather simple. The Fourier slice theorem links 1D
Fourier transform (FT) of the projection data collected at angle θ,Sθ (w)[12],
with 2D FT at the frequency samples. We consider the Radon transform,
namely
Z ∞
Sθ (ω) = R(r, θ) exp (−i2πωr)dr
−∞
Z ∞Z ∞
(2.3)
= f (x, y) exp [−i2πω(x cos θ + y sin θ)]dxdy
−∞ −∞
= F (ω cos θ, ω sin θ)

Then, the unknown f (x, y) can be reconstructed by the Inverse Fourier

Transform (IFT) or the dual Radon transform as following
Z πZ ∞
ˆ
f (x, y) = F (ω cos θ, ω sin θ)|ω| exp [i2πω(x cos θ + y sin θ)]dωdθ
0 −∞
Z πZ ∞
= Sθ (ω)|ω| exp (i2πωr)dωdθ
0 −∞
(2.4)

where fˆ(x, y) denotes the reconstructed image; r = x cos θ + y sin θ; |ω|

is known as “ramp filter ”in the frequency domain. fˆ(x, y) will be identical
with f (x, y) almost everywhere according to the properties of FT and IFT. In
36 Reconstruction Algorithms

practice, the projection data and reconstructed images have to be discretized

to record, calculate and display.

Figure 2.6: The basic workflow of an FBP.

Iterative Techniques
Iterative reconstruction (IR) and algorithms such as projection or image
based noise reduction are currently being introduced to help users to reduce
absorbed dose in CT.
In iterative reconstruction, an initial estimate of the object being imaged
is generated from the acquired projection data. This is typically done using
FBP, which are very fast. IR do not reduce absorbed dose in CT scanning.
Rather, by improving the image quality through noise reduction, the tech-
nical factors that affect absorbed dose can be adjusted to realize excellent
image quality at reduced noise levels.

Figure 2.7: Schematic view of the iterative reconstruction process[13]. The volume estimated is initi-
ated either with an empty image or, if available, an FBP reconstruction. If a stop criterion is matched,the
loop is terminated and the current volumetric image becomes the final volumetric image.

This section will give an overview of available iterative reconstruction

methods and the range of physical models which can be applied in these
algorithms. We distinguish between:
2.1 Theoretical background 37

• Pure iterative methods without any modeling.

• Statistical methods with modeling of the photon counting statistics.

• Model-based methods which go beyond statistical modeling.

Figure 2.8: Selection of the most prominent iterative reconstruction algorithms.

Pure iterative methods without any modeling

All iterative reconstruction consist of three major steps Fig.[2.7]; first, a

forward projection of the current volumetric image is necessary to create
artificial raw data. Then artificial and measured raw data are compared and
an update image is computed which subsequently is backprojected to the
current volumetric image.
The simplest form of iterative reconstruction is the algebraic reconstruc-
tion technique, namely ART. This is based on Kaczmarz method for solving
linear system of equations Ax = b, where in terms of image reconstruction x
are the voxels of the volume to be reconstructed, A is system matrix used for
producing the raw data and b are the pixels of the measured raw data. The
entries of the matrix A correspond to rays from the x-ray source through the
volume to the detector pixels.
To start the ART algorithm, all the pixels in the image array are set to some
arbitrary value. An iterative procedure is then used to gradually change the
image array to correspond to the profiles. An iteration cycle consists of loop-
ing through each of the measured data points. For each measured value, the
following question is asked: how can the pixel values in the array be changed
to make them consistent with this particular measurement? In other words,
38 Reconstruction Algorithms

the measured sample is compared with the sum of the image pixels along
the ray pointing to the sample. If the ray sum is lower than the measured
sample, all the pixels along the ray are increased in value. Likewise, if the
ray sum is higher than the measured sample, all of the pixel values along the
ray are decreased. After the first complete iteration cycle, there will still be
an error between the ray sums and the measured values. This is because the
changes made for any one measurement disrupts all the previous corrections
made. The idea is that the errors become smaller with repeated iterations
until the image converges to the proper solution[14].

Statistical methods with modeling of the photon counting statistics

The key idea of statistical methods is to incorporate counting statistics of

the detected photons into the reconstruction process. In transmission CT the
number of photons leaving the x-ray tube as well as the measured photons
at the detector, which passed through the patient or object, are assumed to
be Poisson distributed[15]. Statistical reconstruction methods can work in
three different domains:
1. Methods in the raw domain (sinogram domain): the meth-
ods working in the raw data apply an adaptive or iterative 2D edge-
preserving denoising algorithm directly on the projection raw data.
2. Methods in the image domain: Algorithm in the volumetric image
domain are denoising the volume after the reconstruction and they are
aiming for noise reduction while preserving spatial resolution. All filters
which are applied on the reconstructed volumetric data without using
any information on the acquisition process are not considered statistical
recostruction methods, but post-processing filters.
3. Full statistical methods: These algorithms can be roughly divided
into two groups: methods based on correlated to the maximum likeli-
hood principle and those based or correlated to the least square prin-
ciple. By maximizing the likelihood, the algorithm tries to find the
set of parameters that makes the measurements the most probable. In
maximization methods the natural algorithm of the likelihood is of-
ten more convenient, which fortunately has its maximum values at the
same points as the function itself.
Next figure shows an example of a statistical IR algorithm, sometimes known
as a hybrid IR algorithm because of its ability to blend with FBP, this com-
mercial manufacturer is called adaptive statistical iterative reconstruction
(ASIR).
2.1 Theoretical background 39

Figure 2.9: The basic workflow of an ASIR algorithm.

Model-based methods which go beyond statistical modeling

The term model-based classifies all methods which go beyond modeling

statistics of the detected photons as Poisson distributed. These methods try
to model the acquisition process as accurately as possible. The acquisition is
a physical process in which photons with a spectrum of energies are emitted
by the focus area on the anode of an x-ray tube; they travel through the
object and are either registered within the area of a detector pixel, scattered
outside of the detector or are absorbed by the object[16].
Model based iterative algorithm (MBIR) is an advanced CT algorithms
that incorporates modeling of several key parameters that were omitted in
earlier algorithms to reduce computational requirement and speed up scans.
More recently, model-based iterative reconstruction (MBIR), also known as
pure IR algorithm, has been shown to significantly improve image quality
while reducing noise and artifacts in multislice CT scans during initial tests.
In MBIR, images are reconstructed by minimizing the objective function
incorporated with an accurate system model, a statistical noise model, and
a prior model. The system model deals with the nonlinear, polychromatic
nature of x-ray tubes by modeling the photons in the measured data set. The
statistical noise model takes into consideration the size of an x-ray tube focal
spot and the three-dimensional shape of detectors. The prior model is a regu-
larization algorithm that corrects unrealistic situations during reconstruction
to speed up the process[17].

Figure 2.10: The basic workflow of an MBIR algorithm.

40 Reconstruction Algorithms

2.2 State of the art

With increasing recognition of the importance of radiation protection, dose
reduction has become an important issue in CT system development. Mod-
ern CT systems are equipped with several dose reduction techniques. These
techniques range from hardware, such as a sliding collimator to eliminate
unnecessary radiation exposure due to overranging, to algorithms such as
improved filtered back projection (FBP) and iterative reconstruction, as pre-
viously seen.
Newer iterative reconstruction techniques are promising for effectively
reducing dose. Studies have shown that with iterative techniques, image noise
(measured as the SD of attenuation) decreases and contrast-to-noise ratio
increases compared with the values with FBP at similar dose levels. Clinical
observer studies have confirmed that with iterative techniques, substantial
reduction in radiation dose can be achieved without loss of image quality.
A limitation of iterative reconstruction is that the images have a different
appearance resulting from a change in shape of noise power spectrum (NPS).
This change in NPS shape hampers objective comparisons of image quality
by assessment of noise expressed as Standard Deviation of attenuation, which
is the general method for measuring noise and its relation to radiation dose.
Each CT manufacturer has developed its own iterative reconstruction
technique. Consequently, results may not be generalized for all iterative
reconstruction techniques. Studies have been performed with ASiR (GE
Healthcare), Veo (GE Healthcare), IRIS (Siemens Healthcare), Safire (Siemens
Healthcare), iDose (Philips Healthcare), and AIDR (Toshiba Medical Sys-
tems).

Figure 2.11: Statistical and Model-based Iterative Reconstruction Algorithms Developed by Major
Computed Tomography Manufacturers[16].

Objective measurements in most of the previous studies have been limited

to measurements of noise levels as Standard Deviation. In only a few studies
did investigators support their results with NPS measurements, and some
2.2 State of the art 41

investigators added an objective evaluation of spatial resolution. In general,

study results show a significant decrease in noise and potential dose reduction.
Now we see the main differences between these manufacturer, finally ex-
plains in detail the algorithm used in this thesis (iDose4 ).

2.2.1 GE Healthcare
ASiR

ASiR, the industry’s first iterative reconstruction technology, may improve

low contrast detectability (LCD) allow for reduced noise, and reduced mA all
without degrading anatomical integrity. ASiR overcomes the limitations of
the conventional CT reconstruction approach (FBP) arriving at an optimal
image using an advanced iterative computation. This computation may dra-
matically reduce the amount of dose needed to acquire a high-performance
image. In early studies, the use of ASiR has been found to enable dose
reductions of one-third to one-half while improving image quality.

Figure 2.12: In a 15-year-old patient presenting to the emergency department to rule out appendicitis,
low-dose scan with FBP reconstruction was noisier than follow-up imaging using the same dose with ASiR
reconstruction[18].

In clinical practice, the use of ASiR may reduce CT patient dose de-
pending on the clinical task, patient size, anatomical location and clinical
practice. A consultation with a radiologist and a physicist should be made
to determine the appropriate dose to obtain diagnostic image quality for the
particular clinical task.
42 Reconstruction Algorithms

VEO
Enabling imaging under 1 mSv with profound clarity; the world’s first model-
based CT iterative reconstruction product, Veo combines sophisticated algo-
rithms with advanced computing power. This breakthrough is changing the
way physicians use CT imaging, delivering a unity of high-performance im-
ages and low dose that was previously unthinkable. VEO establishes new
rules in the relationship between image quality and dose reduction, open-
ing up new possibilities for challenging cases and sensitive patients. VEO
can give the clinicians the diagnostic information they need at previously
unthinkable ultra low dose levels.

Figure 2.13: Liver metastasis visualized with VEO. The right image is less noisy than other[19].

Veo is the worlds first model-based iterative reconstruction product. Ra-

diologists and clinicians can operate under a new set of rules. Lower noise
and higher resolution can be achieved within a single image. At the same
time, significant dose reduction capabilities are now available, opening up
new possibilities for challenging cases and sensitive patients.
The benefits of this capability are significant, especially for the most
radio-sensitive patients including pediatric cases and young women, or those
requiring regular follow-up and monitoring as it dramatically reduces cumu-
lative dose in patients who require regular follow-up exams. Veo is able to
improve image resolution by up to 50% and reduce noise while using as little
as one eighth of the normal dose as compared to a standard CT image. In
traditional CT reconstruction approaches, higher spatial resolution is typi-
cally accompanied by higher image noise. VEO’s CT reconstruction platform
2.2 State of the art 43

challenges this trade-off by improving resolution while reducing noise.

2.2.2 Siemens Healthcare

IRIS
Instead of reducing the amount and complexity of corrective models to gain
reconstruction speed Siemens has developed a new method for iterative re-
construction which maintains the image correction quality of theoretical it-
erative reconstruction. To accelerate the convergence of the reconstruction
and to avoid long reconstruction times the new IRIS applies the raw data
reconstruction only once.
During this newly developed initial raw data reconstruction a so called
master image is generated that contains the full amount of raw data infor-
mation yet at the expense of significant image noise. The following iterative
corrections known from theoretical iterative reconstruction are consecutively
performed in the image space. They clean up the image and remove the
image noise without degrading image sharpness. Therefore, a time consum-
ing repeated projection and corresponding back projection can be avoided.
In addition, the noise texture of the images is comparable to standard well
established convolution kernels.

Figure 2.14: Comparison between standard protocol (FBP) and Iterative reconstruction in image
space [see www.healthcare.siemens.com].

This technique results in artifact and noise reduction, increased image

sharpness, and dose savings of up to 60% for a wide range of clinical appli-
44 Reconstruction Algorithms

cations. IRIS allows to enhance spatial resolution and to reduce image noise
by introducing multiple iteration steps in the reconstruction process.

SAFIRE

SAFIRE (Sinogram Affirmed Iterative Reconstruction) is considered innova-

tive compared with IRIS, because it works not only in image space but also
in raw data domain. First of all, an anisotropic noise model is applied to
images reconstructed with FBP in order to reduce the variance of the signal.
After each iteration, data are reprojected in sinogram space to validate the
images with measurement data, and the detected deviations are corrected,
yielding an updated image.

Figure 2.15: Comparison between FBP and Sinogram Reconstruction. Image noise decrease without
loss of resolution in the right image [see www.healthcare.siemens.com].

SAFIRE is an advanced IR technique that utilizes both projection space

data and image space data, with the number of iterations in each space
dependent on the needs of a specific scan. In contrast to other pure raw
data based IR algorithms, SAFIRE is available right on the scanner and can
reconstruct up to 20 images per second. Therefore, SAFIRE can easily be
used in routine clinical workflow, with well established reconstruction kernels,
providing up to 60% reduction in dose.
2.2 State of the art 45

2.2.3 Toshiba Medical System

AIDR 3D

AIDR (Adaptive Iterative Dose Reduction) is Toshiba’s latest evolution of

iterative reconstruction technology; it has been fully integrated and auto-
mated into the imaging chain to improve dose reduction and image quality.
It is designed to reduce dose by reducing the magnitude of image noise while
preserving image detail, allowing for lower dose acquisitions.
The reconstruction process incorporates raw data and image space noise
optimization. During AIDR 3D reconstruction, a scanner model and a sta-
tistical noise model are taken into account to minimize the effects of photon
starvation and electronic and statistical noise in projection space while an
image-based anatomical model minimizes quantum noise magnitude.

Figure 2.16: In the left image we see noise reduction with AIDR3D. In the right image is shown the
workflow for dose reduction AIDR 3D [see toshibamedicalsystems.com].

This advanced iterative reconstruction algorithm works in two parts. The

first part adaptively removes photon noise in the 3D raw data domain. This
is followed by the second part, model-based iterative noise reduction in the
reconstruction process. AIDR 3D reconstruction has therefore been system-
atically optimized to minimize the impact on patient throughput. As a result,
advanced iterative reconstruction with AIDR 3D adds mere seconds to the
total reconstruction time.
With each interation, AIDR 3D fine-tunes these results for the particular
exam, balancing the relationship between noise magnitude reduction and
spatial resolution preservation. The original FBP image is used as an input
for each iteration and is adaptively blended with the processed data to create
the final image, helping to ensure the noise texture is well maintained for the
viewer.
46 Reconstruction Algorithms

2.2.4 Philips Healthcare

IMR
The low required dose is the result of Philips’ proprietary Iterative model
reconstruction (IMR) system which has set a new direction in CT image
quality with industry leading low-contrast resolution and virtually noise-
free images. Long associated with MR, this improvement in low-contrast
resolution is a made possible through IMR, the first technique to be built on a
knowledge-based model. Enabled by hardware innovation, its reconstruction
speed allows IMR to be used in even the most demanding applications. The
new capabilities also mark the first time a knowledge-based solution takes
benefits previously realised only in routine body imaging to advanced neuro
and cardiovascular applications.

Figure 2.17: Image enhancement of an abdomen using IMR [see www.healthcare.philips.com].

With Philips IMR, low dose and enhanced image quality can be delivered
together in CT imaging. With IMR, clinicians acquiring CT images can
simultaneously lower radiation dose by 60-80%, with 43-80% improvement
in low contrast detectability and 70-83% less image noise, relative to standard
reconstruction techniques (FBP).

iDOSE4
iDose4 is a 4th generation reconstruction technique that provides significant
improvements in image quality and radiation dose reduction. The [Fig. 2.18]
summarizes the advantages of this generation reconstruction techniques in
terms of artifact prevention and the efficiency of quantum mottle noise re-
duction across all frequencies.
2.2 State of the art 47

Figure 2.18: Summary of noise reduction and artifact prevention capabilities provided by each recon-
struction generation (left). Adapting dose reduction and spatial resolution based on the clinical indication
(right) [see www.healthcare.philips.com].

Optimizing the implementation of iDose4 on the Philips CT scanner plat-

forms has enabled the additional clinical benefit of being able to adapt the
spatial resolution and dose reduction benefits to the specific clinical indi-
cation. For pediatric imaging where radiation dose reduction is paramount,
iDose4 enables significantly lower radiation dose while maintaining diagnostic
image quality. In other cases, where image quality (e.g., spatial resolution)
is of higher priority than the dose reduction, such as in the assessment of
coronary stent patency, iDose4 enables significantly improved image quality.
In FBP, the reconstruction filter amplifies any noise present in the projec-
tions proportional to the spatial resolution characteristics of the filter. High
spatial resolution (sharp filter) reconstructions amplify image noise levels to
clinically unacceptable levels and make them suboptimal for low contrast as-
sessments. Hence, the need to perform two reconstructions (i.e., soft filter
and sharp filter).
With iDose4 , the noise in sharp reconstructions can be maintained at a
sufficiently low level to permit soft tissue and detailed, high contrast assess-
ment from a single reconstruction that provides improved image quality over
either individual FBP reconstruction.
This algorithm provides an innovative solution in which iterative pro-
cessing is performed in both the projection and image domains. The recon-
struction algorithm starts first with projection data where it identifies and
corrects the noisiest CT measurements, those with very poor signal to noise
ratio, or very low photon counts. Each projection is examined for points
that have likely resulted from very noisy measurements using a model that
includes the true photons statistics. Through an iterative diffusion process,
the noisy data is penalized and edges are preserved.
This process ensures that the gradients of underlying structures are re-
tained, thus preserving spatial resolution while allowing a significant noise
48 Reconstruction Algorithms

reduction. Since the corrections are performed on the acquisition data; this
method successfully prevents bias error. The noise that remains after this
stage of the algorithm is propagated to the image space; however, the propa-
gated noise is now highly localized and can be effectively removed to support
the desired level of dose reduction.

2.2.5 Summary
After summarizing these iterative reconstruction algorithms, we can say:

• All IR algorithms improved general image quality compared with FBP.

• The statistical IR algorithms improves image quality proportional in-

dependent of radiation dose.

• The model-based IR algorithms improves image quality progressive

with reducing radiation dose.

• ASiR, VEO, IMR, iDOSE4 maintained or improved spatial resolution

at the same time reducing the noise.

• All IR algorithms affected the NPS, especially the model-based.

The application of the iterative process is weighed on scales that change as

a function of the vendors (GE, Siemens, Toshiba, Philips). General Electric
uses percentages (10%, 50%, 90%), Siemens uses nominal values (strength
1, strength 3, strength 5), Toshiba uses standard names (mild, standard,
strong).
In our case (iDose4 ) the scale is numerical and varies from level 1 to level
7, in relation to the anatomical district acquired.
Chapter 3

Image Quality Assessment

3.1 Noise Power Spectrum Analysis

In this section we describe and study a powerful metric for analyzing CT
performance, called Noise Power Spectrum (NPS). In diagnostic CT imaging,
noise is mainly caused by photon detection statistics and cannot be avoided
due to the desire to limit radiation dose. The standard deviation (SD) of
noise, provides no information about noise spatial characteristics and thus
has only a gross predictive value for object detectability.
While the image SD is relatively easy to calculate, it may not fully reflect
the impact that the noise will have on a diagnostic task. In this study, the
NPS will be used to described the noise including correlation introduced by
filtering, remembering that NPS characterizes both the magnitude and spa-
tial frequency distribution of image noise; the magnitude reflects the degree
of randomness at each spatial frequency.
The shape of the NPS reveals where the noise power is concentrated in
frequency space:

1. Low-frequency noise power concentration means noise will be signifi-

cantly mottled in appearance.

2. High-frequency noise power will result in finer grain noise.

Furthermore, the sum of noise power over all non zero frequencies in the
NPS yield the noise variance. By taking into account both the variance and
spatial characteristics of the image noise, the NPS is a more thorough noise
descriptor than pixel standard deviation.
The NPS calculated in this work is divided by the square of the mean
value of the pixels used for analysis. This ratio is referred to as the normalized

49
50 Image Quality Assessment

noise power spectrum (NNPS)[20]:

N P S(u, v)
N N P S(u, v) = [mm2 ] (3.1)
(large area signal)2
The large are signal is basically a measure of the gain of the system. It
is measured in the same units as the data used to determine NPS.
The noise power spectrum is related to the frequency decomposition of
the image noise. First, an image containing nothing but the noise to be ana-
lyzed is decomposed into its two-dimensional frequency components. These
components, called the Fourier amplitudes, are then squared to obtain the
power contribution at each frequency.
This result is averaged with the results obtained from other similar images
to arrive finally at the average noise power for each frequency, or, in other
words, the noise power spectrum. It is therefore proportional to the average
square value of the frequency amplitudes of the noise. The square of the
amplitude is used instead of the amplitude itself, since the average value of
the amplitude is zero while the average value of the squared amplitude is not.
Often the response of the system is assumed to be circularly symmetric[21]
(as we will see later).

3.1.1 Materials and Methods

In this work we acquired some phantoms normally used to investigate im-
age quality in CT. A simple approach, ensuring spectral response similar to
the clinical response, is the use of a homogeneous phantom simulating the
spectral response of the real tissue. The acquisitions were made on a Philips
phantom with different scan, recostruction data, iterative reconstruction lev-
els and different convolution kernel (or filters).
The main phantom used for testing, is composed of two part: head and
body, which are mounted on a support before each measurement.

Figure 3.1: Philips Phantom used for acquisition. Body phantom (LEFT SIDE) and head phantom
(RIGHT SIDE).
3.1 Noise Power Spectrum Analysis 51

The measurements were performed using iCT128 MSCT (Philips Health-

care) installed at Radiology Department of Sant’Orsola-Malpighi Hospital,
in Bologna. The images were acquired in axial and spiral scan mode, with
reconstruction matrices 512 × 512 (both head and body phantom).

Body Phantom

The first phantom analyzed was the body unit, which is made from a single
polymethylmethacrylate (PMMA) container of 320 mm in diameter, filled
with water and with a teflon (PTFE) pin inside.
This phantom is used for calibration and quality assurance in body modal-
ity. It is mounted on the bed of the system.
First of all, several images are acquired for each set analyzed. To study
the noise power spectrum, we used the method suggested in [22][23][24],
therefore we selected the first seven slice reconstructions. The reason for
this is because if there are large exposure variations over the measurements
area, computing the NNPS separately for each ROI and then averaging will
tend to cancel some (but not all) the variations in noise power, resulting
from regional variations in x-ray exposure. Namely decrease the statistics
fluctuations of noise texture, in such a way that the shape of NNPS will have
less peaks.
The next table shows the acquisition and reconstruction parameters for
the Noise Power Spectrum study.

Voltage 120 kVp

Exposure 200 mAs
Beam Collimation 64×0.625
Convolution Kernel A, B, C, D, D(H)
Slice Thickness 2.5 mm
Pixel Sixe 0.6835×0.6835 mm
FOV 350×350 mm
Rotation Time 0.75 s
Matrix Size 512 × 512

Table 3.1: List of parameters utilized.

52 Image Quality Assessment

Figure 3.2: Convolution kernel for body phantom.

The images were studied by a set of plugin applications for ImageJ soft-
ware, called QA-distribution. The Noise Power Spectrum was calculated
using this software in the following way.
The plugin can handle both linear and square root compressed images
without signal offset. The NNPS is calculated from a uniformly exposed
image. The image is divided in image fragments with 2N × 2N , in this case
64 × 64 because this ROI provide a bin frequency around 0.03 lp/mm, a good
compromise to get a proper definition without excessive noise[22][23][24].
The noise analysis is performed on these individual subregions and the
final NNPS is calculated as the average of all individual noise spectra, ob-
tained from each subregion. We assume that a flatfield corrected linear image
is used[25].
At this time, we select the parameters for calculate NNPS.

• Select actual pixel size.

• Select NNPS fragment size.

• Choose the number of lines around the central axis not be used for the
evaluation of the NNPS.

• Choose whether you want to show the 2D NNPS image on the screen.

• Select image type.

Taking account that the size for the NNPS calculation should be a power
of 2, and a 2D FFT-algorithm is used, we choose a uniformly exposed area and
select a ROI excluding the teflon insert. The radially averaged NNPS is cal-
culated by radial averaging the NNPS over all pixels of the 2D NNPS[26][27].
3.1 Noise Power Spectrum Analysis 53

Figure 3.3: Spatial frequency (mm−1 ) and radially NNPS values (mm2 ) (LEFT). Body phantom
image and ROI utilized for calculate the Normalized Noise Power Spectrum (RIGHT).

Figure 3.4: Values of NNPS calculate for all seven slice with FBP algorithm and Convolution kernel
A.
54 Image Quality Assessment

Figure 3.5: 2D image of the NNPS, note the circular symmetry.

Since the 2D Fourier spectrum is a circular symmetry we select column

H in the Fig.[3.2], this process was repeated for each slices of acquisition and
NNPS was calculated for all seven slices Fig.[3.4]. Finally this values are
averaged and utilized for the final NNPS[28][29].
The reconstruction kernel changes in relation to the characteristics of
spatial resolution and contrast of the images that they determine. These
feature of image quality are in opposition to each other. Then, they can be
classified in two categories:

1. Smoothing kernel, improve contrast resolution, low frequency: to high-

light elements that have little contrast to the background (e.g. ab-
domen, skull). It can be advantageous to employ these kernels in obese
patients, where the signal-to-noise ratio can be diminished secondary
to attenuation from adipose tissue.

2. Edge kernel, improve spatial resolution, high frequency: edge enhance-

ment of anatomic structure (e.g. lung).

First of all we plot each convolution filter with different iDose levels to see
how the curves change the shape of NNPS[30][31]. Observing this plot, NNPS
measured at a fixed radius in the uniform region showed a dependency on the
reconstruction method. For iterative reconstruction we found that different
levels of iDose reduce the noise magnitude and shifted the noise texture.

Convolution kernel A

For increasing levels of iDose, the magnitude of the NNPS curve decreases
while the peak is shifted to low frequencies. For iDose6, the spatial frequency
peak is reduced by ∼ 0.1 mm−1 respect to iDose1.
Compared to convolution kernel D (or C or DH), image reconstructed
with kernel A will have worse spatial resolution, but will also have less noise
at lower frequencies.
3.1 Noise Power Spectrum Analysis 55

Figure 3.6: Normalized noise power spectrum for different iDose levels. An image reconstructed using
this filter, producing noise texture with low spatial frequency noise. See Appendix A for phantom’s image.

Convolution kernel B

Figure 3.7: Normalized noise power spectrum for different iDose levels. First idose level will have two
peaks while iDose6 will have one peak shifted at lower frequencies. See Appendix A for phantom’s image.

For this kernel we observe that the first iDose level has two peaks at ∼ 0.15
56 Image Quality Assessment

and ∼ 0.25 mm−1 . The positive slope region is caused by the ramp filtering,
and the subsequent fall-off at higher frequencies is due to the mathematical
reconstruction kernel that is used.
In this case the fall-off for iDose1 is past 0.3 mm−1 and for iDose6 is
beyond the first peak at 0.15 mm−1 . Note the plateau between 0.05 mm−1
and 0.15 mm−1 .

Convolution kernel C

Figure 3.8: Normalized noise power spectrum for different iterative reconstruction levels. For higher
levels the peaks from 0.25 mm−1 to 0.35 mm−1 disappear. See Appendix A for phantom’s image.

By studying this filter, we note that the negative slope at high spatial
frequencies fall more in subsequent zone (beyond 0.7 mm−1 ) than kernel A
(∼ 0.6 mm−1 ) and B (∼ 0.7 mm−1 ).
This convolution kernel shows two peculiarities.

• First, iDose1 has more noise fluctuations than iDose6.

• Second, last reconstruction level has two peaks at 0.1 mm−1 and ∼
0.15 mm−1 similarly to iDose6 of kernel A.

In this case the difference is a plateau in this range of frequencies and the
fall-off is beyond ∼ 0.45 mm−1 .
The positive slope region of the NNPS is lower than kernel A, B, D, DH;
the plateau has an higher NNPS, however, is shorter than filter B.
3.1 Noise Power Spectrum Analysis 57

Since this kernel has a wide range in all iterative levels, it can be used to
study NNPS in medium spatial frequencies, even though it is a sharp filter.

Convolution kernel D

Different kernels are used that represent different trade-offs between spa-
tial resolution and image noise, but all clinical kernels produce some roll-off
in response at high frequencies. Roll-off refers to the progressive reduction
in the filter function at higher spatial frequencies, to reduce the impact of
quantum noise on the image[32].
This filter works at higher frequencies than the others. The plot shows
that NNPS varies more than in the previous cases and for iDose1 between
0.32 mm−1 to 0.5 mm−1 will form a wide variation of NNPS values.
This feature vanishes with higher iterative levels and for iDose7 will form
a plateau and also the noise power decreases slower after ∼ 0.5 mm−1 . Unlike
the previous case, the shape of noise power spectrum is shifted towards higher
frequencies than filter C.
Note that for iDose1, iDose2, iDose3 and iDose4 the curves overlap at
low frequencies (∼ 0.06 mm−1 ); iDose4 and iDose5 plots are separated and
then again overlaps for iDose5, iDose6 and iDose7 for very low values (< 0.02
mm−1 ).

Figure 3.9: All curves shifted by high frequencies than other kernels. See Appendix A for phantom’s
image.
58 Image Quality Assessment

Convolution kernel DH

This filter is a sharp kernel that passes more higher-frequency noise. Com-
pared with kernel A,B,C and D will have better spatial resolution but will
also have more noise at higher frequencies.
This filter stress fine parts, in the image. Unfortunately, while kernel A,
B smooths out noise, sharp filtering does the exact inverse: it enhances noise.
One can skip this, if the original image is not excessively noisy; generally the
noise will overpower the image. We can high-pass filter only the brightest
parts of the image, where the signal-to-noise ratio is most noteworthy.
In Fig.[3.10] the shapes of noise power spectrum are very similar up to ∼
0.45 mm−1 . Beyond this value we note that the peak at ∼ 0.5 mm−1 , with
increasing iterative levels, disappears and the NNPS will have little variations
respect to lower frequencies.
Secondly, for values under ∼ 0.15 mm−1 there are no changes; furthermore
for iDose7 the slope remains unchanged. For other reconstruction levels there
are variations from 0.02 to ∼ 0.15 mm−1 , especially for the first four levels.

Figure 3.10: Shape of Normalized Noise Power Spectrum of FBP and reconstruction levels. The
curves tends to zero more slowly than the others filter. See Appendix A for phantom’s image.

Next study focuses on compared Filtered Back Projection and Iterative

reconstruction algorithm using three iterative levels, which are level 1, level
3 and level 4. The iDose level is selectable, and it reflects a scale of how
aggressively the noise reduction is performed relative to noise that would be
obtained if reconstructed with FBP. The purpose is to examine the effects of
3.1 Noise Power Spectrum Analysis 59

reconstruction filters on NNPS using traditional geometry (i.e. non-helical).

In particular, the effects of the traditional roll-off smooth filter (A,B) and
the enhancing filter (C, D, DH) will be studied. To improve accuracy, the
result was averaged over all seven images acquired previously.
Remember that, while NNPS have much potential in the way of standard-
izing protocols across institutions, certain protocols, such as cardiac imaging,
in which the number of projections contributing to the final image at a par-
ticular gantry angle can vary, are not emanable for NNPS analysis.
For these iDose levels and convolution filters we calculate the normalized
noise power spectrum [see Appendix A for others tables] and we compared
each reconstruction level with FBP. The iterative algorithm levels utilized
for this study are the ones used in clinic:

• Level 1 is more similar than Filtered Back Projection (10.6% noise

reduction).

• Level 3 has a medium reconstruction level (22.5% noise reduction).

• Level 4 has a reconstruction level greater than previous (29.3% noise

reduction).

iDose1 vs FBP

Convolution kernels are designed to weight the frequency content of the pro-
jection. Therefore, the specific design of the kernel greatly influences the
shape of the NPS and the quality of the final images. It following that
choosing different convolution kernels allows the user to somewhat control
noise texture. Also, the ability to choose a specific kernel allows the user to
exploit a fundamental tradeoff in CT between noise and resolution.
In frequency space, most filters follow the same basic ramp-up roll-off
design. The ramp is designed to reduce blurring and roll-off portion is highly
variable and controls the high-frequency content of the image. As mentioned
previously, sharp kernels preserve more high frequency content and thus re-
sult in better spatial resolution. The limit for high resolution is increased
noise at high spatial frequencies. Soft kernels suppress high frequency content
and offer reduced noise at the cost of lower spatial resolution.
We analyze [Fig. 3.11] and focus on different aspects such as the shape
at low frequencies, the slope at high frequencies and noise reduction for all
kernels, starting from Filtered Back Projection.
60 Image Quality Assessment

Figure 3.11: Trend of noise power spectrum for all reconstruction kernels. Radially averaged normal-
ized NPS curves show how noise texture is manifested in the NPS.

• Shape analysis at low frequencies: for range between 0 and 0.1 mm−1 ,
kernel DH has less noise than kernel D (both iterative and standard)
while kernel DH FBP has the same noise texture of iterative kernel
D. Instead, for other kernels the NNPS is overlying and impossible to
study.
For range between 0.1 mm−1 and 0.2 mm−1 the sharp filters tends to
have different trends, the NNPS varies slightly therefore can be studied.
The smooth filters (A,B) don’t allow us to study the noise texture. The
use of smooth filters does not make much sense for the previous reason,
however a comparison of sharp filters is useful to understand when to
use them.
• Slope analysis at high frequencies: for values higher than 0.5 mm−1 ,
the trend of sharper filters drops more quickly than smooth filters.
However, kernel A,B and C go down from 0.25 mm−1 and lowest values
of the curves overlap.
The convolution kernels D,DH improve spatial resolution but, as we
can see from the plot, the noise values are higher and are unable to
discern with higher intensity than the ones of other filters.
• Noise reduction: By using iterative algorithms the noise power is much
3.1 Noise Power Spectrum Analysis 61

lower than standard reconstruction, for sharp filter the difference is

more marked in a range between 0.2 and 0.6 mm−1 for kernel DH, while
it is remarkable between 0.2 and 0.55 mm−1 for kernel D. The difference
is much lower for smooth filters, especially for kernel A there is little
difference between 0.2 and 0.3 mm−1 . Images with coarse texture have
an NNPS concentrated at low frequencies (kernel A,B) while images
with finer texture have an NNPS concentrated at higher frequencies
(kernel D,DH)[33].

iDose3 vs FBP

Level 3 of iterative algorithm shows further differences than level 1. The

curves are similar but at some frequency range they overlap, and we must be
careful to select which filters are used for analysis.
We repeat the previous procedure to analyze the noise power spectrum.

Figure 3.12: Trend of noise power spectrum for all reconstruction kernels. Note the differences at ∼
0.45 mm−1 and ∼ 0.3 mm−1 .

• Shape analysis at low frequencies: for ∼ 0.2 mm−1 , iDose3 kernel DH

has less noise than kernel D (FBP) even if the first one is high resolu-
tion. The sharper filters (with iterative reconstruction) shows oscillat-
ing curves which intersect in two points (∼ 0.08 and ∼ 0.18 mm−1 ).
62 Image Quality Assessment

The difference shown is greater than iDose1 analysis. The noise texture
for sharp kernels (both FBP and iterative) is impossible to study for
values < 0.1, but between 0.1 and 0.2 the trends can be studied (even
if the differences are minimal) for standard versus iterative algorithm.
The smooth filters (A,B) for values < 0.1 cannot be studied but between
0.1 and 0.2 the shape and the slope is similar. In this case the difference
among two algorithms is lower than sharp filters.

• Slope analysis at high frequencies: for values higher than 0.5 mm−1 , the
sharper filters are steeper than smooth filters. However, smooth kernels
(A,B) for iterative and standard reconstruction begin to go down from
0.25 mm−1 but for kernel C the curve start to decline for ∼ 0.45 mm−1 .
The convolution kernels D,DH improve spatial resolution but, aa can be
seen from the plot, the noise values are higher and are unable to discern
NNPS variations principally between 1.00 × 10−3 and 3.00 × 10−3 .

• Noise reduction: By using iterative algorithms the noise power is much

lower than standard reconstruction (this is true for all iDose levels
studied), the difference is more marked for the sharp filter in range
between 0.1 and 0.7 mm−1 . The variation is much lower for smooth
filter, starting from ∼ 0.1 for kernel C, and from ∼ 0.25 for kernel A and
B. The main difference with [Fig. 3.11] is that kernel DH (iterative)
and kernel D (FBP) are crossing at ∼ 0.42; before this value the NNPS
for kernel DH is lower than kernel D, while after it becomes higher.
Then, we must be careful about which filter use to study a specific
spatial frequency range.

iDose4 vs FBP

Level 4 of iterative algorithm shows some peculiarities compared to other

reconstruction levels, mainly at high frequency.
Once again, we analyze the three main features which characterize the
noise power spectrum analysis. We repeat the same method used previously.

• Shape analysis at low frequencies: for values < 0.1 mm−1 , both kernel D
and kernel DH follow a pattern that respects the characteristics of those
filters; for values of ∼ 0.05 mm−1 , standard and iterative have a similar
NNPS and then we cannot observe a good texture noise variation. For
∼ 0.2 mm−1 the sharper filters (both iterative and standard), are very
similar to each other.
3.1 Noise Power Spectrum Analysis 63

The smooth filters (A,B) for values < 0.1 mm−1 cannot be studied
because the variations of NNPS are too small, whereas between 0.1
and 0.2 trends are beginning to take shape allowing to distinguish the
values of noise although very small. In this case the difference among
two algorithms is similar than sharp filters.

Figure 3.13: Trend of noise power spectrum for all reconstruction kernels. Note the lowest noise
power for kernel D (FBP) after 0.55 mm−1 .

• Slope analysis at high frequencies: for values higher than 0.5 mm−1 ,
iterative reconstruction for smooth kernel tends go down more quickly
than FBP. The difference is less remarkable for kernel A, while for
kernel B the curve is slightly different although beyond this value is
difficult to make an accurate assessment. It is clear from the graph
that NNPS do not ramp up as to the traditional roll-offs shown for
smooth kernel (this for all reconstruction levels). Sharper filters has
a similar slope, especially iterative reconstruction with kernel DH and
standard reconstruction with kernel D after 0.55 mm−1 .
• Noise reduction: In Fig.[3.11], we show the NNPS associated to the
standard roll-off filters (kernel A,B,C), ranging from more smooth to
the sharper in their effects. Remembering that these noise power spec-
trum represent a constant dose condition (200 mAs) for unsubtracted
64 Image Quality Assessment

axial images, the noise power shifts to higher frequencies and alias-
ing of noise does not occur with the sharper filters in the NNPS. In
this case, unlike the previous, one can see two differences: first, at
0.5 mm−1 kernel DH (iterative reconstruction) and D (standard recon-
struction) overlap; after ∼0.55 mm−1 NNPS is lower for FBP but for
values <0.5 mm−1 the iterative becomes smaller. Second, the smooth
filters have the same trend as previously cases except kernel C. Note
the low-frequency structured noise is not present. Because most ob-
jects are low-frequency dominated, this noise could have a potentially
deleterious effect on the object detectability.

Next step is to study and visualize the noise fluctuations within body
phantom. This process has been done with all convolution kernels both FBP
and iterative reconstruction (See appendix A for other plots), by analyzing
all slice acquired. Internally the image is scaled to a square image using
nearest neighbor sampling. For selections, the bounding box of the selection
is used for the surface plot.
In the left image we have a FBP reconstruction that represent the noise
fluctuations (light blue), where x-axis and y-axis is the FOV (field of view).
The luminance of an image is interpreted as height for the plot. Internally
the image is scaled to a square image using nearest neighbor sampling. Right
image represent the iterative reconstruction, it can be noted that the noise
fluctuations are less pronounced than filtered back projection.

Figure 3.14: Noise texture fluctuations of Filtered Back Projection algorithm with convolution kernel
A (LEFT). Noise texture fluctuations of Iterative reconstruction algorithm (iDose, level 4) with same
convolution kernel of FBP (RIGHT). The teflon insert is not affected by noise texture and reconstruction
algorithm.
3.1 Noise Power Spectrum Analysis 65

Head Phantom
Head phantom unit is composed by three sections. Section 1 is a PMMA
transparent box, with a diameter of 225 mm, filled by air. Section 2 and 3
are enclosed in a PVC box with a diameter of 200 mm, filled with distilled
water. The head unit is mounted outside, free in air.

• Section 1 is a physical layer and this part is equipped with guides, used
to measure the slice width and the impulse response.

• Section 2 is a water layer, used for head calibration and to measure

uniformity and noise. Furthermore, the rapid test on the image quality
use this section to measure and calculate the low contrast resolution.

• Section 3 is a multiple pins layer, equipped with pins of various ma-

terials, used to measure the linearity of brain mode and the contrast
scale. This section has a pin used to measure the spatial resolution in
standard mode.

Voltage 120 kVp

Exposure 200 mAs
Beam Collimation 64×0.625
Convolution Kernel A, EB, UB, C, D, D(H)
Slice Thickness 2.5 mm
Pixel Sixe 0.4882×0.4882 mm
FOV 250×250 mm
Rotation time 0.75 s
Matrix Size 512 × 512

Table 3.2: List of parameters utilized.

Figure 3.15: Convolution kernel for Head Phantom.

For noise characterization, to generate an image noise similar to a clini-

cal head scan, an homogeneous phantom filled with water, which mimics a
66 Image Quality Assessment

normo-type head, was used. To reproduce the clinical conditions, the FOV
of 250 mm was used for the homogeneous phantom acquisition.

For each reconstruction stack of 16 images, a subset of 7 images was

processed with QA-distribution (imageJ plugin). For each image, the NNPS
from a 64 × 64 region of interest (ROI) fragments at the image center was
calculated and subtracted pixel-by-pixel from the same ROI to cancel the
Fourier Transform dc component. The dc component is not useful to a human
observer because, the human visual response is low at low frequencies and the
absolute image brightness cannot be used as an image feature for detection.

The noise power spectrum was produced by squaring the FFT magnitude
image and, finally, the NPS stack was averaged using seven images to generate
a single NNPS image with better accuracy. The ROI used was large enough
to accurately identify the low frequencies in the NNPS.

To compare and present the results, the mono-dimensional NNPS were

extracted from the 2D images with a radial average of the data points (exactly
as for the other phantom).

Figure 3.16: Spatial frequency (mm−1 ) and radially NNPS values (mm2 ) (LEFT). Head phantom
image and region of interest utilized for calculate the Normalized Noise Power Spectrum (RIGHT).
3.1 Noise Power Spectrum Analysis 67

Figure 3.17: Values of NNPS calculate for all seven slice with FBP algorithm and Convolution kernel
A. Note the different values of spatial frequency for head phantom compared to body phantom.

iDose1 vs FBP

The reconstruction kernel is defined as the image processing filter applied

to the raw data to yield a final scan image. The sharpness of the final image
is most directly influenced by the type of filter employed. The kernels usually
have more straightforward names such as smooth kernels (A, EB, UB).
These filters will tend to smooth edges and reduce the amount of image
noise. In the head are used to study the soft part of the brain that has low
contrast.
Sharp kernels (C, D, DH) tend to enhance edges at the cost of increased
overall image noise.
The predominant source of noise is the x-ray photon concentration fluc-
tuation, called quantum noise. The prevalence of noise makes convolving our
signal with a filter necessary.
We analyze Fig.[3.18-3.19] and focus on different aspects such as the shape
at low frequencies, the slope at high frequencies and noise reduction for all
kernels, starting from Filtered Back Projection (exactly as for body phan-
tom). Kernels UB and EB have been studied separately, because these filters
are very similar to each other.
68 Image Quality Assessment

Figure 3.18: Trend of noise power spectrum for all reconstruction kernels, except kernels UB-EB.
Radially averaged normalized NPS curves show how noise texture is manifested in the NPS.

• Shape analysis at low frequencies: These curves are very similar to each
other for values less than 0.2 mm−1 . Smooth filter A and sharp filter
C are overlap for iterative and filtered back projection reconstruction,
then both can be used to improve contrast resolution in this range of
frequencies. The shape of NNPS for ∼ 0.2 mm−1 begins to change,
and for values greater than 0.2, we must be careful which filter to use
if we want to make a thorough analysis.
For sharper filters (D and DH), NNPS overlaps at ∼ 0.1 mm−1 . Com-
pared to iDose1 for body phantom, we note that the curves are similar
and from this value onwards they take slightly different trends.
• Slope analysis at high frequencies: There are many differences in com-
parison to body phantom. First, the peaks of kernel DH (both iterative
and standard) are shifted beyond 0.6 mm−1 compared to [Fig. 3.11]
and this affects the texture of the image. For kernel D at spatial fre-
quency of ∼ 1 mm−1 , both curves overlap and it is difficult to estimate
the correct measure for this range of frequency. Not only that, but
in this case, using the iterative or FBP will not change the result of
reconstruction. Kernels A and C are impossible to study because the
differences are virtually undetectable.
• Noise reduction: For head phantom, the values of NNPS are lower than
body phantom (for all kernels). For kernel DH the curve is less noisy
3.1 Noise Power Spectrum Analysis 69

and steep than FBP; there is a plateau at ∼ 0.6 mm−1 , which means
that the texture fluctuations are less noticeable compared to [Fig. 3.11].
For kernel DH the greater difference is between 0.5 and 0.7 mm−1 ,
and accordingly towards the higher frequencies compared to kernel DH
for body phantom. For kernel D is ∼ 0.38 mm−1 , therefore tends
to go to lower frequencies compared to kernel D for body phantom.
Kernel A goes towards zero at 0.5 mm−1 , i.e. for the same value of
Fig.[3.11]; iterative and FBP reconstruction have the same NNPS (same
shape). For iterative algorithm, kernel C is slightly different than FBP
reconstruction as it varies only in a small range of frequencies (from
0.2 to 0.4 mm−1 ).

Figure 3.19: Comparison between smooth convolution kernels for head acquisition. UB improves
bone-brain interface and no effect on HU values; EB head scans only and increased to observed HU values
(not shown here).

• Comparison between kernel UB and EB : These two filters are very sim-
ilar; they both work at lower spatial frequencies, which is a sign that
these kernels are more suited for the study of contrast resolution (e.g.
abdomen, skull). For values beyond 0.5 mm−1 , these filters are com-
pletely unnecessary as there are no variations in the values of NNPS.
The major changes occur at lower frequencies, in fact we observe sev-
eral things: first of all, the curve of kernel EB is too smooth compared
to the kernel UB, therefore is subject to minor fluctuations in the tex-
ture of the image. However, filter EB presents a level of NNPS greater
70 Image Quality Assessment

than filter UB, even if it comes to very small values.

iDose3 vs FBP

Remember that iDose4 is a hybrid iterative reconstruction algorithm that

provides reduced image noise compared with the conventional filtered back
projection. Level 3 of reconstruction for the head phantom has several simi-
larities with level 3 of body phantom indeed even here the noise is reduced,
in a slightly better way than body. We now repeat the same considerations
made for all other cases.

Figure 3.20: Trend of noise power spectrum for all reconstruction kernels, except kernels UB-EB.
Radially averaged normalized NPS curves show how noise texture is manifested in the NPS.

• Shape analysis at low frequencies: This small range is identical to the

smooth kernel A (both iterative and standard reconstruction) compared
to body phantom. For ∼ 0.2 mm−1 kernel A (FBP) and kernel C
(iterative) overlap, and then their use to analyze the texture in this
range is the same as before. Sharper filters (D and DH) have similar
trends and begin to diversify for values > 0.2 mm−1 . We note that,
in this case, there are no differences between iDose3 for head phantom
and iDose3 for body phantom.
• Slope analysis at high frequencies: For higher frequencies the curves
for kernel DH go down for values > 0.6 mm−1 . Kernel D for standard
3.1 Noise Power Spectrum Analysis 71

reconstruction go down for values > 0.4 mm−1 ; instead kernel D for
iterative decreases after 0.5 mm−1 . The other two filters analyzed,
at frequencies > 0.7 mm−1 do not give contribution. This is correct
because most of the information is given at lower frequencies.

• Noise reduction: Also here, the variations of noise power spectrum are
greater than level 1 of reconstruction, which is observed in particular
for sharper kernels. The main changes are found around 0.4 mm−1 for
kernel D, around 0.6 mm−1 for kernel DH, around 0.3 mm−1 for kernel
C and around 0.25 mm−1 for kernel A (although, in the latter case, the
variations are too small). Since the amplitude of the NNPS is greater
than iDose1 (especially for kernel DH), the trends of the curves here
overlap. For example, all kernels will tend to overlap at lower values of
the NNPS.

Figure 3.21: Comparison between smooth convolution kernels for head acquisition. UB improves
bone-brain interface and no effect on HU values; EB head scans only and increased to observed HU values
(not shown here).

• Comparison between kernel UB and EB : The NNPS is lower than the

previous case (for iterative algorithm). In this level of reconstruction,
kernel EB and UB show differences respect to level 1 of reconstruction.
First, kernel EB go down from 6.5 × 10−5 mm2 to 5.5 × 10−5 mm2 and
shows a fall around the value of 0.1 mm−1 , afterwards it goes up imme-
diately to 0.15 mm−1 then decreases steadily to zero at ∼ 0.7 mm−1 .
72 Image Quality Assessment

Kernel UB shows a less relevant peak and has a value of NNPS lower
than level 1; further between 0.35 and 0.4 mm−1 it does not overlap
with the other filter, whereas previously the overlap is more marked for
values > 0.25 mm−1 .

iDose4 vs FBP

Compared to pixel standard deviation, NPS has greater clinical potential

for task-based image quality assessment, describing both the magnitude and
spatial frequency characteristics of image noise. While iDose4 reduces the
NPS magnitude more dramatically than FBP, further studies are needed to
assess the impacts of noise texture alteration on its clinical usage. Compared
to FBP, the iDose4 algorithm reduced the NPS magnitude while preferentially
reducing noise at mid-range spatial frequencies, altering noise texture. This
reduction was more significant with increasing iDose4 noise reduction level.

Figure 3.22: Trend of noise power spectrum for all reconstruction kernels, except kernels UB-EB.
Radially averaged normalized NPS curves show how noise texture is manifested in the NPS.

• Shape analysis at low frequencies: In this range the situation can be

summarized as follows. The shape is identical to the previous cases,
moreover for values < 0.1 mm−1 the curves are indistinguishable and
for range between 0.1 and 0.2 slight differences starts to be seen. The
3.1 Noise Power Spectrum Analysis 73

iterative reconstruction allows a better noise reduction than filtered

back projection, and the smooth filters have a similar behavior if we
want to analyze the contrast resolution.

• Slope analysis at high frequencies: Kernel A and C do not give contri-

bution at high frequencies (> 0.8 mm−1 ). Kernel D and DH behave
differently than the two iterative levels previously analyzed. In this
case, the high resolution filter tends to almost overlap filter D (FBP);
at a closer look, the iterative has a value of NNPS which is slightly
higher, and this is due to the shift of the curve towards the higher
frequencies.

• Noise reduction:We remember that the NPS can be thought as the

variance associated with a particular frequency component of an image.
For a flat-field image, NPS is the variance associated with a particular
frequency component of the noise in that image. In most cases, noise
may be represented or approximated, as a stationary Gaussian random
process with zero-mean.
Even in this situation, the NNPS magnitude is reduced (especially for
level 4) through the use of an iterative reconstruction. Kernel DH
(iterative) and kernel D (FBP) have the same value of NNPS (3 ×
10−4 mm2 ), but have different texture because the first kernel has a
peak shifted toward higher frequencies compared to kernel D (FBP).
Therefore, the filter DH (iterative) has a spatial resolution better than
filter D (FBP), in fact the peak is shifted at high frequencies. Farther,
both kernels have the same NNPS amplitude. Unlike the level 3 of
reconstruction, at 0.2 mm−1 kernel A (FBP) has NNPS slightly higher
than kernel C (iterative), but at 0.4 mm−1 the situation is reversed.

• Comparison between kernel UB and EB : This level of reconstruction

has some differences compared to the previous cases. Iterative kernel
EB has one peak and does not present changes in the curve of NNPS
up to 0.12 mm−1 . After this value, it begins to fall steadily reaching
zero for spatial frequency of ∼ 0.7 mm−1 .
Unlike the previous cases (levels 1 and 3), kernel UB presents two peaks
at the same value of NNPS (∼ 4 × 10−5 mm2 ) for spatial frequencies 0.1
and 0.2 mm−1 . Then this filter has a trend less uniform than kernel
EB, but as before it overlap from 0.25 to 0.5 mm−1 showing an identical
behavior to filter EB.
74 Image Quality Assessment

Figure 3.23: Comparison between smooth convolution kernels for head acquisition. UB improves
bone-brain interface and no effect on HU values; EB head scans only and increased to observed HU values
(not shown here).

As for the case of body phantom, here also it is possible to observe and
analyze noise fluctuations. The teflon pin is not present, and the fluctuations
are similar to those already observed in the body phantom. Both images
represent the convolution kernel DH. On the x-y plane there is the field of
view (FOV), which is smaller than Fig.[3.14].

Figure 3.24: Noise texture fluctuations of Filtered Back Projection algorithm with convolution kernel
DH (LEFT). Noise texture fluctuations of Iterative reconstruction algorithm (iDose, level 4) with same
convolution kernel of FBP (RIGHT).
3.2 Modulation transfer function analysis 75

3.2 Modulation transfer function analysis

Another common metric of image quality is the MTF, which US FDA (Food
and Drug Administration) accepts as one component of performance valida-
tion for solid state digital detectors. The MTF, related to image resolution,
is the Fourier transform of the point spread function (PSF), the systems re-
sponse to a point object. The MTF and PSF are useful quantities for linear,
shift-invariant systems, where the imaging systems response to an arbitrary
object can be determined by convolving the true object and the point spread
function.
This is not true for nonlinear IR algorithms or even for FBP when non-
linear filters are used. For nonlinear systems, the MTF whose definition
assumes independence of location in the image, dose, contrast, and algo-
rithm parameters-acquires dependence on these quantities. As the MTF is
not well defined for images reconstructed by IR, it is of limited utility in
assessing the quality of these images[34].
Spatial resolution is an important attribute of any radiological imaging
system. In CT, spatial resolution depends not only upon physical parame-
ters such as the focal-spot size and detector element dimensions similar to
projection radiography, but because all CT images are reconstructed math-
ematically, the resolving power of a CT image is fundamentally linked to
the image reconstruction methods as well. In addition to traditional filtered
back projection (FBP) reconstruction algorithms, which utilize a variety of
reconstruction kernels that have a profound impact on spatial resolution, it-
erative CT image reconstruction techniques using statistical or model based
constructs are also used clinically. These non linear, adaptive algorithms
create a challenging mathematical environment for objectively characteriz-
ing spatial resolution[35].
The spatial resolution of a CT measurement is affected by a multitude of
factors. Substantial contributions are the properties of the CT device, namely
the x-ray source (focal spot size) and the detector (pixel size, scattering),
but also the used magnification in cone-beam geometry. The measurement
strategy, especially the number of projections per full rotation, and the re-
construction process (voxel size, interpolation, filtering) have a big influence
too.
Spatial resolution refers to the ability of an image to convey detail; med-
ical imaging systems produce images that are usually degraded in detail
compared with the actual object being imaged.
Fourier based metrology is a prevalent approach to characterizing med-
ical imaging performance. This includes the MTF, which characterizes the
resolution of the imaging system. The output signal I(x, y) (i.e., image)
76 Image Quality Assessment

is related to the input signal f (x, y) (i.e., object) by the impulse response
function as
Z +∞ Z +∞
I(x, y) = f (x − x0 , y − y 0 )P SF (x0 , y 0 )dx0 dy 0 (3.2)
y=−∞ x=−∞

where P SF (x, y) is defined as the PSF such that

Z Z
P SF (x, y)dxdy = 1 (3.3)

The Fourier representation of the PSF is the optical transfer function (OTF)

OT F (u, v) = F {P SF (x, y)} (3.4)

where F{ } denotes the Fourier transform. The MTF is the modulus of the
OTF as
M T F (u, v) = |OT F (u, v)| (3.5)
describing the magnitude of transferred signal at each spatial frequency. Note
that (3.3) implies that M T F (0) = 1 (dc signal magnitude is unchanged by
the transfer function); and an ideal system has M T F = 1 at all spatial
frequencies.
The MTF can be employed to characterize the resolution of an imaging
system in terms of M T F (f ) or to reflect the resolution in a scalar form, for
example, the frequency (f50 ) at which M T F (f ) reduces to 0.50[36].
Experimentally, it is impractical to measure the PSF directly. That would
require imaging of an infinitesimally small object, a 2D delta function, so that
I(x, y) = P SF (x, y). If one is to do so, the mean signal magnitude would
be correspondingly small and the resulting image would be dominated by
noise. More convenient impulse function techniques are typically used for
such measurements, such as the LSF, which is the Radon transform of the
PSF Z
LSF (x) = P SF (x, y)dxdy (3.6)

The edge-spread function can be determined by the integration of the line-

spread function Z x
ESF (x) = LSF (x0 )dx0 (3.7)
x=−∞

which can be measured from the image of a sharp edge. This can be rewritten
as
∂
LSF (x) = ESF (x) (3.8)
∂x
3.2 Modulation transfer function analysis 77

Hence, the PSF, LSF, ESF, and MTF are all related. Equations (3.6)(3.7)(3.8)
illustrate that the family of spread functions, are related and that any one of
them can be used to assess the resolution of a CT system in the axial plane.
The PSF, LSF and ESF are functions that describe resolution in the spatial
domain. However, it is common to transform these functions into the spatial
frequency domain, to obtain the modulation transfer function[37].
For a radially symmetric MTF, equations from (3.3) to (3.8) together
with the Fourier slice theorem suggest that the one-dimensional MTF (i.e.,
a slice of the full 2D MTF) is given by
R +∞
| −∞ LSF (x)e−2πif x dx|
M T F (f ) = R +∞ (3.9)
−∞
LSF (x)dx

where here f represents spatial frequency. The integral in the denominator

normalizes the MTF to unity at f = 0. If the PSF or ESF is measured
directly in CT, these functions can be transformed into the LSF using Eqs.
(3.6) or (3.8), respectively, for subsequent assessment of the MTF. Although
analytical computation of the MTF from the LSF is possible in some cases.
Eqs. (3.9) is evaluated in most cases by a computer subroutine.

Figure 3.25: (a)Input images defining the point-spread function, the line-spread function and the edge-
spread function.(b) Simulated degraded-output images showing raw image data used for the measurements
of the PSF, LSF and ESF. The blurring seen in these functions is due to the imperfect resolution properties
of the imaging system being characterized.(c) Graphs showing the actual PSF, LSF and ESF. The PSF
is a 2D function, and the LSF and ESF are 1D functions.

3.2.1 Test Device and MTF processing

For this study we utilize the Philips phantom (head unit) used in section 3.
78 Image Quality Assessment

The next table shows the acquisition and reconstruction parameters for
the Modulation Transfer Function analysis.

Voltage 120 kVp

Exposure 200 mAs
Beam Collimation 64×0.625
Convolution Kernel A, EB, UB, C, DH
Slice Thickness 1.25 mm
Pixel Sixe 0.3515×0.3515 mm
FOV 180×180 mm
Rotation Time 0.75 s

Table 3.3: List of parameters utilized.

In order to characterize the reconstruction filters used in this study, the

MTF of each reconstruction filter was measured. The methodology used
to generate these MTF was a wire technique: a wire perpendicular to the
scan plane was scanned axially on a multislice CT scanner (ICT 128, Philips
Healtcare). The wire was located to the left of the image and acquired with
a multiple axial scan using 120 kVp, 200 mAs, 1.25 mm slice thickness and
180 mm reconstruction FOV.
The wire was acquired once and then reconstructed with each filter as
indicated in table 3.3. The reconstructed field of view (FOV) was chosen to
be small enough to ensure that all frequencies up to the Nyquist frequency (as
dictated by the machine sampling rate) would not be aliased by the display
grid. A wire image was then generated for each reconstruction algorithm.
For each of the resulting wire images, the center of the wire was identified. A
spatially invariant linear system is assumed, implying that there is no overlap
of system response in frequency space.
However, inside the phantom, there is a slanted wire (50-80 µm tungsten)
that reproduced a point spread function witch is suspended in air; since
the wire is subpixel sized it is not usually necessary to compensate for its
size. Then, we use the impulse source to estimate the point source response
function of the CT system. From the print out of a digitized image of the area
surrounding the impulse source, we used the numerical data to determine the
two-dimensional array of the CT values arising from the impulse source[38].
The FWHM of the point spread function is determined from the best-fit
curve of the point spread function numerical data. The average of several
different arrays of impulse response functions is calculated to obtain the
average point spread function of the system. The point spread function
3.2 Modulation transfer function analysis 79

(PSF) was measured, and then the 2D Fourier transformation of the PSF
was performed to obtain a MTF curve. The spatial frequencies at 10% and
50% MTFs were calculated to evaluate the axial resolution[39][40].
All this has been performed through the use of a software called IQ-
WORKS. The procedure performed by the software is the following: the
software did automatically search the wire and analyzed a squared ROI that
contained it. The array analyzed had dimension of 16 × 16 pixel and from
it, central profile was taken and the offset has been eliminated. The software
applied a Gaussian filter to reduce the noise, and then has performed a
zero-filling Fourier Transform1 . Finally, from pixel and array dimensions
was calculated the correct scale of the spatial frequencies. MTF data are
normalized to zero spatial frequency.

Figure 3.26: Phantom image corresponding to the Philips head phantom using a typical adult head
protocol. Image window and level have been adjusted to show bead point source within the ROI (LEFT).
Modulation transfer function reconstructed with kernel A; the spatial frequencies at 10% and 50% are
shown (RIGHT).

From here on begins the actual analysis. For each MTF was taken one
1
It turns out that the Fourier transform algorithm used by computer programs is most
suited to a number of data points which is a power of 2. So,for example, 214 = 16384 is
a suitable number of data points to transform, but 15000 is not. In practice, therefore, it
is usual to zero fill the time domain data so that the total number of points is a power
of 2. Zero filling costs nothing in the sense that no extra data is required; it is just
a manipulation in the computer. Of course, it does not improve the resolution as the
measured signal remains the same, but the lines will be better defined in the spectrum.
This is desirable, at least for aesthetic reasons if nothing else.
80 Image Quality Assessment

image with the bead point source that reproduce the point spread function.
The IQ WORKS software generates a plot which represents the modulation
transfer function with correct spatial frequency; from it were extrapolated
values that will be used later to compare the graphs obtained with different
convolution kernels (both standard and iterative reconstruction).

Figure 3.27: MTF values with FBP and iterative algorithm using kernel A (LEFT). MTF plot and
values of spatial frequency at 10% and 50% (RIGHT.)

Figure 3.28: MTF values with FBP and iterative algorithm using kernel EB (LEFT). MTF plot and
values of spatial frequency at 10% and 50% (RIGHT.)

Figure 3.29: MTF values with FBP and iterative algorithm using kernel UB (LEFT). MTF plot and
values of spatial frequency at 10% and 50% (RIGHT.)
3.2 Modulation transfer function analysis 81

Figure 3.30: MTF values with FBP and iterative algorithm using kernel C (LEFT). MTF plot and
values of spatial frequency at 10% and 50% (RIGHT.)

Figure 3.31: MTF values with FBP and iterative algorithm using kernel DH (LEFT). MTF plot and
values of spatial frequency at 10% and 50% (RIGHT.)

The high contrast or spatial resolution of the system is often determined

using objects having a large signal to noise ratio. Spatial resolution is influ-
enced by factors including:

• System geometric resolution limits-focal spot size, detector width and

ray sampling.

• Pixel size.

• Properties of the convolution kernel/mathematical reconstruction filter.

In the previous figures [3.27-3.31] we note a slightly difference between MTFs

obtained with all convolution kernels, especially for filter DH. The MTF curve
is shifted by higher spatial frequencies, which means that this filter gives a
better spatial resolution than others. Typically, these reconstruction filters
that enhance or preserve the higher spatial frequencies do so at the cost of
increased noise in the image. For high signal to noise objects, this tradeoff
is usually acceptable.
82 Image Quality Assessment

The method described here can be used for all images from multislice
sequential scans and also for images from helical acquisitions. Scan plane
spatial resolution is not generally dependent on helical scan parameters, al-
though the values may be slightly different from those taken with sequential
scans due to small differences in acquisition and reconstruction techniques.

Now we plot all modulation transfer functions to show how the recon-
struction algorithms behave with different filters. We use five levels of recon-
struction to see if there are differences between the trends of curves.

Figure 3.32: MTF for filtered back projection and level 1 of iterative reconstruction algorithm.
3.2 Modulation transfer function analysis 83

Figure 3.33: MTF for level 2, level 3 and level 4 of iterative reconstruction algorithm.
84 Image Quality Assessment

Figure 3.34: MTF for level 5 of iterative reconstruction algorithm.

The iterative reconstructions are identical to the Filtered Back Projection.

Fig.[3.32-3.34] illustrate the results of the MTF calculation with the phantom.
The five curves are very similar in shape and position. The iDose algorithm
does not decrease the spatial resolution. This is an important observation,
because in this case the iDose algorithm reduces noise (as seen previously),
whereas the spatial resolution is maintained. Then, a key component of
the iDose4 algorithm, is the preservation of the spatial resolution among the
different strengths (levels).
The above results can be summarized as follows: high MTF values in
the low frequency range is needed to outline the coarse details of the image
and are important for presentation and detection of relatively large but low
contrast lesions. Consequently increased MTF values in the high frequency
range are necessary to portray fine details and sharp edges. This is of obvious
importance for small objects but also sometimes for larger objects due to the
importance of edges and sharp borders for detection of low contrast objects
and for accurate assessment of their size and shape.
The final analysis of this chapter is to assess the values of MTF50% ,
MTF10% and standard deviation (in HU) of FBP and iterative reconstruc-
tion algorithm, thous to observe what kind of pattern has the variation of
the modulation transfer function.
All MTF data was averaged into 0.05 mm−1 bins as per IEC specification
to facilitate a comparison of the data, as shown in[39].
3.2 Modulation transfer function analysis 85

FBP
MTF10% MTF50% SD[HU]
kernel A 0.5 0.3 43.4
kernel UB 0.5 0.3 46.7
kernel EB 0.5 0.2 47.6
kernel C 0.6 0.3 57.3
kernel DH 1.0 0.5 137.1
Table 3.4: Values of MTF at 10%, 50% and standard deviation (noise) of filtered back projection for
all convolution kernels.

iDose1
MTF10% MTF50% SD[HU]
kernel A 0.5 0.3 43.4
kernel UB 0.5 0.3 46.7
kernel EB 0.5 0.2 47.6
kernel C 0.6 0.3 57.3
kernel DH 1.0 0.5 137.0
Table 3.5: Values of MTF at 10%, 50% and standard deviation (noise) of idose level 1 for all convolution
kernels.

iDose5
MTF10% MTF50% SD[HU]
kernel A 0.5 0.3 43.4
kernel UB 0.5 0.3 46.7
kernel EB 0.5 0.2 47.6
kernel C 0.6 0.3 57.3
kernel DH 1.0 0.5 136.9
Table 3.6: Values of MTF at 10%, 50% and standard deviation (noise) of idose level 5 for all convolution
kernels.
86 Image Quality Assessment

Figure 3.35: Modulation transfer function at 10% compared standard deviation for all convolution
kernels. The filter DH has a value greater than other.

Figure 3.36: Modulation transfer function at 50% compared standard deviation for all convolution
kernels. The filter DH has a greater value than other.
Chapter 4

Low-Contrast Detectability

The low contrast detectability (LCD) represents the ability of CT scanner to

distinguish between objects that have similar x-ray attenuation coefficients.
In cases of other diagnostic imaging modalities it means the ability of imag-
ing equipment to differentiate between objects that have similar properties.
LCD is one of the important performance parameters for Computed Tomog-
raphy (CT). Often LCD is determined by visual inspection of low contrast
phantom images. The testing process is gated not only by limitations of
existing LCD phantoms, but also by the subjective nature of the evaluation
methodology[41].
In general, LCD aims to describe the performance of a CT system in
detecting objects of low contrast against the background. Ideally, one would
define an objective test method, for example, using phantoms, to assess LCD
for scanner characterization. In current practice, LCD is typically specified
by measuring a low contrast phantom with objects of different sizes and
different densities.
The crux lies in the presumed ability to see a certain low contrast struc-
ture, since for an individual this is a highly subjective task. It makes difficult
to obtain statistically objective data with some level of confidence using vi-
sual methods.

4.1 Catphan 600 phantom

Our study is based on images acquired with the Catphan 600 phantom (The
Phantom Laboratory Incorporated, Salem, NY). The phantom, cylindrical
in shape, is constructed of PMMA and consists of 5 modules designed to
perform various quality tests in tomographic images. The phantom long axis
(z-axis) has to be placed longitudinal to the CT table, the modules are in

87
88 Low-Contrast Detectability

transverse planes to the phantom z-axis (x-y plane).

The Catphan 600 phantom has multiple test modules for assessing various
image quality parameters. Low contrast detectability was measured using the
CTP515 low contrast module. The ability to distinguish between the object
and the background is limited by object size and image noise.

Figure 4.1: Catphan 600 phantom (LEFT). CTP515 low contrast module with supra-slice and subslice
contrast targets (RIGHT).

Supra-slice targets have diameters from 2.0mm to 15.0mm. Subslice tar-

gets have diameters fro 3.0mm to 9.0mm. In low contrast resolution module
we can find three areas with different nominal contrast levels: 1%, 0.5% and
0.3% (supra-slice targets). All of the targets in each contrast group are cast
from a single mix to assure that the contrast levels will be the same for all
targets.
Along with the supra-slice (targets with z axis dimension longer than
most maximum slice width) the CTP515 low contrast module includes sub-
slice targets (targets with z axis length smaller than some of the usual slice
width). The subslice targets are arranged in the inner circle of tests in the
module. The subslice targets are cast from the same mix as the 1.0% supra-
slice targets. Because they are from the same mix in the evaluation of the
actual subslice target contrast the supra-slice targets can be used to estab-
lish contrast values. The subslice targets have z axis lengths of 3, 5, and 7
mm and diameters of 3, 5, 7, and 9 mm. The evaluation of subslice target
readability is helpful in understanding the scanner’s different spiral imaging
settings and how the settings will affect the ability to visualize small objects
with low contrasts from their background.
4.1 Catphan 600 phantom 89

4.1.1 Contrast to noise ratio with Catphan 600

Low-contrast resolution was assessed in two steps using the low-contrast
module of the phantom Fig.[4.2]. The module contains three sets of outer
supraslice cylinders, with a z-axis dimension of 40 mm, which according to
the manufacturer have a nominal contrast of 1%, 0.5% and 0.3%. Objective
assessment was done by measuring the CT numbers (HU) within the ROI and
in the background selected, the noise was measured as the standard deviation
within the same region of interest (ROI) of background selected previously.
Contrast-to-noise ratios (CNRs) were calculated using the formula:

HUobject − HUbackground
CN R = (4.1)
σbackground

where HUobject is the average pixel value of the low contrast object in a 15mm
diameter region of interest, HUback. is the average pixel value of the back-
ground region of interest and σ is the standard deviation of the attenuation
values of the background (in HU). The parameter σback. not only includes
photon statistics and electronic noise in the results but also structural noise
than can obscure the object.
The CNR is a useful metric for describing the signal amplitude relative to
the ambient noise for simple and largely homogeneous objects. However, the
CNR depends only on contrast and noise. Actual signal detectability also
depends on factors including signal size, shape, and density distribution;
background level, variability, and correlation; the variance and covariance
of measurement noise; spatial resolution; and the observer and detection
strategy used. The CNR can be useful in some simple situations, e.g., de-
termining thresholds of contrast agents at which signals on a test phantom
become visible.
Since the target contrasts are nominal, the actual target contrasts need
to be determined before testing specific contrast performance specifications.
The limiting detectability should be measured with the reconstruction al-
gorithm of the scanner which is routinely used, as well as other clinically
relevant reconstruction algorithms. The baseline performance level must
be stated for a given phantom at specific scan conditions, including radi-
ation dose, viewing conditions, and visualization criteria. It should be noted
that this visual test for establishing LCD is subjective since it depends on a
number of factors including the visual acuity of the observers and ambient
lightning conditions.
The next table shows the acquisition and reconstruction parameters for
the CNR analysis.
90 Low-Contrast Detectability

Voltage 120 kVp

Exposure 300 mAs
Beam Collimation 64×0.625
Convolution Kernel A, EB, UB, C
Slice Thickness 5 mm
Pixel Sixe 0.4901×0.4901 mm
FOV 250×250 mm
Rotation Time 0.75 s

Table 4.1: List of parameters utilized.

The actual contrast levels are measured by making region of interest mea-
surements over the three larger targets, for all contrast and in the local back-
ground area. To determine actual contrast levels, average the measurements
made from several scans. For this study we use the first five reconstruction
slice using kernel A, EB, UB and C (both standard and iterative algorithm).
It is important to measure the background area adjacent to the measured
target because ”cupping” and ”capping” effects cause variation of CT num-
bers from one scan region to another. Position the region of interest to avoid
the target edges. The region of interest should be at least 4 × 4 pixels in
diameter.
Because low contrast measurements are ”noisy” it is advisable to calculate
the average of the multiple measurements made from several scans. Carefully
monitor the mAs setting because the photon flux will improve with increased
x-ray exposure. Use the size of the targets visualized under various noise
levels to estimate information on contrast to noise ratio.
Although decreasing tube current is the most means of reducing CT ra-
diation dose, this alteration also reduces the contrast-to-noise ratio, which
may affect the diagnostic outcome of the examination. This is especially
true in abdominal studies, where the low-contrast area are severely affected
by CNR. Some studies suggest that scanning with low tube voltage is possi-
ble to reduce dose without markedly affecting image quality; however, there
are few reports on the effect of low tube voltage on abdominal image quality
and low-contrast detectability[42][43][44].
We selected only the first big target (15mm) for all nominal contrasts.
This is because the other smaller targets are very difficult to study on a
conventional displays (contrast and resolution limited). All measurements
were made using the same region of interest (ROI), stored using imageJ
software with reconstruct image sets using 5mm slice thickness.
4.1 Catphan 600 phantom 91

First and foremost we have done a window-level to improve the image display
and to perform all measurements correctly. Subsequently, the measurements
were made using the slice centered on the image plane and the mean values
of the object and background are acquired obtaining the contrast to noise
ratio. The same method was applied to all reconstructed images with all
convolution kernels.
Then, the software allowed to store a circular ROI, accordingly obtain
the values used for the calculation of CNR, this for all analyzed kernels and
for each phantom insert (0.3%, 0.5%, 1%). The procedure and the analysis
are shown in Fig.[4.2-4.5-4.8].

CNR =⇒ Nominal Contrast 1%.

The detectability of detail increases with object size and/or contrast be-
tween object and background. For example, the detectability of objects with
the same contrast will increase in line with the object size. Similarly, when
object size is maintained, detectability will increase with increasing contrast.
Hence, small objects can have higher contrast than larger objects for the
same detectability.

Figure 4.2: Catphan phantom analysis using nominal contrast of 1%.

The CNR was obtained by subtracting the mean CT number measured

in the 15mm diameter object from the mean CT number measured nearby
(phantom background) and by dividing the result by the standard deviation
of the pixel values of the phantom background.
92 Low-Contrast Detectability

The region of interest used to perform the measurements was slightly

smaller than the 15mm diameter of the object in order to include only it in the
measurements. The measurement was repeated for filtered back projection
and iterative algorithm (for the first five reconstruction levels).
Remembering that, the larger the CNR, the less noise is viewed on an
image and the quality is improved; we can summarize the results shown in
the table below as follows: looking at this data, the images with the highest
CNR are, for all convolution kernels, those using the iterative algorithm with
the highest level of reconstruction.
The convolution kernel A has CNR higher than the others, which means
that its use allows us to obtain images (and therefore reconstructions) with
better contrast for the same pin with nominal contrast of 1%.

Nominal contrast 1%
Kernel A Kernel UB Kernel EB Kernel C
FBP 6.2± 0.2 4.6± 0.3 4.5± 0.3 3.6± 0.3
idose 1 6.4± 0.2 4.8± 0.3 4.9± 0.3 4.0± 0.2
idose 2 7.6± 0.2 5.3± 0.3 5.3± 0.3 4.2± 0.2
idose 3 8.1± 0.2 5.7± 0.2 5.7± 0.2 4.5± 0.2
idose 4 8.6± 0.1 6.1± 0.2 6.1± 0.2 4.9± 0.1
idose 5 9.4± 0.1 6.6± 0.1 6.7± 0.1 5.5± 0.1
Table 4.2: CNR values both standard and iterative reconstruction; nominal contrast 1%.

Figure 4.3: Contrast to noise ratio for iterative and standard algorithm. We can see the differences
between the values; they are very similar between kernel UB and EB.
4.1 Catphan 600 phantom 93

Fig.[4.3] shows the variations of the CNR, compared to the different it-
erative levels used. The images which show the greatest amount of noise are
the images with the lowest CNR, which are those related to the convolution
kernel C (i.e. sharp filter). This data also reflects the results seen in the low
contrast detectability exam, where only a few targets are visible with images
that have low CNR.

Figure 4.4: Trends of contrast to noise ratio at varying levels of reconstruction.

Fig.[4.4] shows the trend of the various filters by varying the algorithm
and its levels of reconstruction. We observe that filter A does not vary much,
however it rises between iterative level 1 and level 2; while tends to have a
linear trend from the second level onwards.
UB and EB filters overlap, then they give the same contrast variation
using Filtered Back Projection and iterative algorithm. These two filters
have, in fact, very similar features; the only thing that distinguishes them is
the type of acquisition, axial for UB and helical for EB.
The filter C instead has the lowest value of CNR; this is because, being a
moderate sharp filter, allows the noise to affect more the image. In this way
the calculated contrast will be lower than the previous filters.

CNR =⇒ Nominal Contrast 0.5%

Low contrast objects with a high noise level will generate a low CNR
value which results in reduced visibility of the object as marked in the yellow
ROI. CNR plays a certain role when it comes to detecting objects of inherent
94 Low-Contrast Detectability

low contrast. With high inherent contrast, the object is clearly visible also
with a high noise level.
The samples and the background material have equivalent effective atomic
numbers. Only the densities is varied to produce changes in effective atten-
uation coefficients.
Unlike previous measures, the ROI was placed on the medium low con-
trast object and mean CT number was recorded. In addition a ROI with the
same size was selected from the background and mean CT number values and
SD were stored. The CNR values were calculated according to the equation
(4.1).

Figure 4.5: Catphan phantom analysis using nominal contrast of 0.5%.

Nominal contrast 0.5%

Kernel A Kernel UB Kernel EB Kernel C
FBP 4.5± 0.3 2.6± 0.3 2.5± 0.3 2.4± 0.3
idose 1 4.8± 0.2 2.7± 0.3 2.7± 0.3 2.7± 0.3
idose 2 5.4± 0.2 3.0± 0.2 2.9± 0.2 2.9± 0.2
idose 3 5.8± 0.2 3.2± 0.2 3.1± 0.2 3.1± 0.2
idose 4 6.3± 0.2 3.4± 0.2 3.3± 0.2 3.4± 0.2
idose 5 7.0± 0.1 3.8± 0.2 3.5± 0.2 3.7± 0.1
Table 4.3: CNR values both standard and iterative reconstruction; nominal contrast 0.5%.

The table 4.3 shows the values obtained using a pin with nominal contrast
0.5%. Even here, values of CNR increase with the use of iterative algorithm;
4.1 Catphan 600 phantom 95

however, the values are generally lower than those obtained previously for all
convolution kernels.

Figure 4.6: Contrast to noise ratio for iterative and standard algorithm. We can see the small
differences between the filter C and EB.

Figure 4.7: Trends of contrast to noise ratio at varying levels of reconstruction.

Fig.[4.6] shows that for filter A the variation between FBP and level 5 of
iDose is greater than the other filters. This means that the iterative algorithm
allows to reconstruct the image with better contrast.
96 Low-Contrast Detectability

For filter C we immediately noticed that for standard reconstruction the

value of CNR is lower than filter EB; while for the first iterative level, the
CNR is almost equal. Instead, for higher levels of reconstruction its values
are even higher (the latter compared with the filter EB). The other values
of CNR have a performance similar to the case designed for the pin with
nominal contrast 1%.
Fig.[4.7] allows to observe how the CNR varies depending on the levels of
reconstruction. Unlike the previous case, UB, EB and C filters behave in a
different way: all three trends are superimposed, especially for the first three
levels of iDose4 . Instead as regards the filtered back projection, there is a
slight difference in the values for filter C; while in the case of the pin with
nominal contrast 1% variability was more pronounced.
The last two levels of reconstruction show different features. The CNR
for the filter EB has an even lower value (even if slightly) compared to the
sharp filter and standard filter, which is not observed in the case discussed
previously.

CNR =⇒ Nominal Contrast 0.3%

Also in this case, the procedure for the calculation of the CNR is similar
to that used in the two previous cases. We made sure to select the region
correctly, given the very low visibility. The target chosen is the biggest one
with the lowest nominal contrast. The results showed that increasing level
of iDose4 improved the low contrast resolution.

Figure 4.8: Catphan phantom analysis using nominal contrast of 0.3%.

4.1 Catphan 600 phantom 97

In this case, from the carried analysis, the variations of the CNR can be
observed, compared to the previous analysis. For the filtered back projection,
the difference with level 5 of iDose is less marked than nominal contrasts 1%
and 0.5%. All filters have low values, this in agreement with the fact of
having considered the pins with lower contrast, thus more difficult to detect.
In the filter A there is a greater variation, this is because the iterative
algorithm influence to a greater extent the smooth filter. Instead, the filters
EB and UB have almost the same values, instead the filter C has values even
higher. We will immediately notice the difference in the values of this filter,
via the analysis made in the case of the nominal contrast 0.5%.

Nominal contrast 0.3%

Kernel A Kernel UB Kernel EB Kernel C
FBP 2.9± 0.3 1.6± 0.2 1.6± 0.2 1.7± 0.2
idose 1 3.0± 0.3 1.7± 0.2 1.7± 0.2 1.9± 0.2
idose 2 3.6± 0.2 1.9± 0.2 1.9± 0.2 2.0± 0.2
idose 3 3.9± 0.2 2.0± 0.2 2.0± 0.2 2.1± 0.1
idose 4 4.1± 0.2 2.2± 0.1 2.2± 0.1 2.3± 0.1
idose 5 4.4± 0.1 2.4± 0.1 2.3± 0.1 2.5± 0.1
Table 4.4: CNR values both standard and iterative reconstruction; nominal contrast 0.3%.

Fig.[4.9] shows the values of the contrast-to-noise ratio at varying levels

of reconstruction. Also here, the filter A is one that has the highest contrast
(both standard and iterative algorithm), for the same considered pins.

Figure 4.9: Contrast to noise ratio for iterative and standard algorithm. We can see the differences
between the filters C, EB and UB.
98 Low-Contrast Detectability

Instead, what we see is that the other filters have almost similar values,
and filter C has even higher values than standard filters (EB and UB).
The figure below describes the trends of the CNR, from filtered back
projection to level 5 of iterative reconstruction. The trend of the smooth
filter is very similar to the previous one, while the others tend to have the
same variation of the CNR. But here we have a peculiarity: the sharp filter
has a value very close to the one of standard filters UB and EB, for FBP.
While for all iDose levels, this filter has higher values than the other two
filters. This feature is found only by analyzing the low contrast referred to
pin with nominal contrast 0.3%.

Figure 4.10: Trends of contrast to noise ratio at varying levels of reconstruction.

4.1.2 CIRS 061 phantom

The CIRS Helical CT Phantom is designed to test scanning protocols to ver-
ify that small, low contrast lesions will be detected. The phantom permits
complete testing of low contrast lesion detection when scan parameters are
varied. These parameters include collimation, pitch, reconstructed field of
view, reconstruction algorithms, z-axis interpolators, kVp, mA and rotation
time. Testing can be applied to protocols designed for head and abdomen.
The phantom is manufactured from a proprietary epoxy material which ac-
curately mimics liver tissue.
The CT Lesion Detectability Phantom is particularly useful to physicians,
CT technologists, and medical physicists who design scanning protocols for
abdominal, pelvic, and brain CT. It allows users to test various scanning
4.1 Catphan 600 phantom 99

protocols to verify that small low contrast lesions will be detected. This is
the only way to be sure that a CT scanner is ”seeing” tumors that are known
to be present. The use of this phantom removes many doubt as to the limit
of low contrast spherical lesion detectability for various scan protocols, it
is used for 3D analysis. This lesion detectability testing can be applied to
protocols designed for imaging of the liver, spleen, pancreas, kidneys, and
adrenal glands. It can also be used for mass detection in the brain.
The CT lesion detectability Phantom is a tissue-equivalent test object
that consists of an 18 cm diameter right circular cylinder with a CT value of
50 HU at 120 kVp. Within the phantom is an 18 cm diameter, 4 cm deep right
circular void in which a soft-tissue equivalent disk (containing low contrast
spheres) can be placed. The cylindrical void is in a plane containing the
z-axis of the scanner. The soft-tissue-equivalent disk also has a background
CT value of 50 HU. Embedded within the disk are three sets of simulated
spherical lesions.
There are three rows of spherical targets that are 5, 10 and 20 HU below
the liver equivalent background matrix. This phantoms is designed to assist
technical and clinical staffs in the selection of optimal spiral/helical scanning
parameters.
One set is 5 HU below background, a second set is 10 HU below back-
ground, and the last set is 20 HU below background. Each set contains one
sphere each of the following diameters: 2.4, 3.2, 4.0, 4.8, 6.3 and 9.5 mm.
These diameters were chosen to encompass the full range of clinically signif-
icant lesions. The disk can also be placed at the end of the phantom when
axial scanning detectability testing is desired.

Figure 4.11: Spiral CIRS phantom, internal view (LEFT). Phantom contains spherical objects; these
spheres are placed in three rows. Each row contains spheres that were originally designed to be 20, 10,
and 5 HU below background (designed to equal liver; no attenuation given (RIGHT).
100 Low-Contrast Detectability

4.1.3 Contrast to noise ratio with CIRS 061

The phantom body serves a dual purpose. One, it enables the test disk to
be positioned parallel to the CT couch in the z-axis without artifacts and
with appropriate surrounding attenuation. Two, it can be used as a support
device to position the test disk at 90 degrees to the CT couch in the x, y
axis. The testing disk has exterior markings to enable location and orienta-
tion of the embedded targets. The disk has a scribe line which indicates the
midline of the test disk and the targets. The disk is also marked top, left
and right to correspond with the Fig.[4.11].
The main features for this phantom are:
• Incorporates clinically-relevant lesion shape (spherical) and size.

• Provides clinically-relevant absolute HU values for soft tissue.

• Provides a clinically-relevant HU differential (i.e. tumors have a slightly

lower HU than backround).

• Designed for use on all conventional and spiral (helical) CT scanners.

• Valid for x-ray energies from 80 to 140 kVp.

• Background Hounsfield Units (HU) approximate liver tissue.

Next table shows the acquisition and reconstruction parameters for the
CNR analysis.

Voltage 120 kVp

Exposure 300 mAs
Beam Collimation 64×0.625
Convolution Kernel A, B, C
Slice Thickness 5 mm
Pixel Sixe 0.4901×0.4901 mm
FOV 250×250 mm
Rotation Time 0.75 s
Pitch 1.067

Table 4.5: List of parameters utilized.

In the abdomen, the most significant challenge related to low dose is low-
contrast object detection, especially in the liver, where neoplastic disease
is commonly manifest as a low-attenuation object within a background of
4.1 Catphan 600 phantom 101

slightly higher attenuation normal liver. Thus, any dose-reduction strategy

must address and maintain low-contrast object detection.
In this case, since the targets are below the liver equivalent background
matrix, the equation to calculate the contrast-noise ratio can be written as:
 
 HUobject − HUbackground 
CN R =   (4.2)
σbackground 

Now, we proceed in a similar manner to the case of Catphan phantom.

The only difference will be the nominal contrast analyzed.

CNR =⇒ Nominal Contrast 2%

A low-contrast phantom (CIRS Helical CT Phantom, Computerized Imag-

ing Reference Systems), composed of three rows of cylinders was scanned.
The raw data were reconstructed using FBP and iDose4 with kernel A, B
and C. All reconstructions were performed using the same size FOV at the
same table position.

Figure 4.12: Cirs 061 phantom analysis using nominal contrast 2%.

In the phantom, by using imageJ software, same size regions of interest

(ROIs) in the same position and belonging to the same slice were placed on
the large cylinder, and the adjacent background used is the one belonging to
the right of the sphere. The contrast was calculated as the difference between
the mean CT numbers of the ROI from the low contrast cylinder and the
phantom background. The noise was the SD of the CT numbers inside the
102 Low-Contrast Detectability

ROI of the phantom background. Then, the CNR was calculated using (4.2)
and the results are shown in the next table.
Nominal contrast 2%
Kernel A Kernel B Kernel C
FBP 5.8± 0.4 3.4± 0.3 2.2± 0.2
idose 1 6.5± 0.4 3.8± 0.3 3.3± 0.2
idose 2 7.2± 0.4 4.1± 0.3 3.5± 0.2
idose 3 7.7± 0.3 4.3± 0.2 3.8± 0.1
idose 4 8.5± 0.3 4.7± 0.2 4.2± 0.1
idose 5 9.6± 0.3 5.2± 0.1 4.7± 0.1
Table 4.6: CNR values both standard and iterative reconstruction; nominal contrast 2%.

These values reflect the assessment made for the other phantom. Here
too, the filter A gives a greater value than other filters. Even, the FBP is
almost half of the level 5 of iDose4 (for smooth filter) and less than half
respect to sharp filter. This is due to the fact that, probably, the nominal
contrast is different from Catphan 600, furthermore, Hounsfield unit varies
significantly between the pin and background.

Figure 4.13: Contrast to noise ratio for iterative and standard algorithm.

Next figure shows the curves obtained using kernel A, B and C. The
smooth filter has a values larger than other two, then maintains the char-
acteristic of being a filter that improves the contrast resolution. Instead,
4.1 Catphan 600 phantom 103

kernels B and C have very similar values, except in the range between FBP
and level 1 of iterative algorithm.

Figure 4.14: Trends of contrast to noise ratio at varying levels of reconstruction.

CNR =⇒ Nominal Contrast 1%

Next analysis leads us to consider the pin with nominal contrast of 10 HU

(1%). For the phantom that we are considering, for each object imaged below
background attenuation (10 HU), there is an observation at low contrast
cylinder both in FBP and iDose4 .

Figure 4.15: Cirs 061 phantom analysis using nominal contrast 1%.
104 Low-Contrast Detectability

The simulated lesions usually have fixed location and are framed by sur-
rounding structures, which are often more conspicuous than the lesions of
interest. The routine use of thinner collimation with MSCT has resulted in
more frequent detection of very small lesions, and the detection and char-
acterization of small lesions can be important for patients with history of
malignancy[47]. Now, we repeat the same procedure by positioning the ROI
above the cylinder.
Nominal contrast 1%
Kernel A Kernel B Kernel C
FBP 2.7± 0.3 1.9± 0.2 1.6± 0.2
idose 1 2.9± 0.3 2.1± 0.2 1.8± 0.2
idose 2 3.1± 0.3 2.2± 0.2 1.9± 0.2
idose 3 3.3± 0.2 2.4± 0.2 2.0± 0.1
idose 4 3.7± 0.2 2.5± 0.2 2.2± 0.1
idose 5 4.0± 0.2 2.8± 0.1 2.4± 0.1
Table 4.7: CNR values both standard and iterative reconstruction; nominal contrast 1%.

The table above shows the values obtained by using pin (lesions) with
density 10 HU lower than that of the background; nevertheless, while lesions
with great difference in density or with higher density to background might
present different results, a difference in the density of 10 HU represents the
minimum lesion to liver contrast necessary for detection.

Figure 4.16: Contrast to noise ratio for iterative and standard algorithm.

Here, we note a small differences between nominal contrast 2% and nom-

inal contrast 1%.
4.1 Catphan 600 phantom 105

In the nominal contrast of 2%, kernels B and C have a variation in the

CNR higher than nominal contrast of 1%. For Filtered Back Projection the
range of variation is 1.2, while for iterative reconstruction is 0.5 (level 5). In
the nominal contrast 1%, the range is 0.3 (FBP) and 0.4 (iDose, level 5).

Figure 4.17: Trends of contrast to noise ratio at varying levels of reconstruction.

CNR =⇒ Nominal Contrast 0.5%

Figure 4.18: Cirs 061 phantom analysis using nominal contrast 0.5%.
106 Low-Contrast Detectability

Last analysis leads us to consider the pin with nominal contrast of 5 HU

(0.5%). Our results suggest that the use of a phantom with fixed lesion
location may therefore be adequate for optimization of routine liver proto-
cols when the detection of small lesions is required. We repeat the same
procedure, by positioning the ROI above the insert.
We note that it is very difficult to place the region of interest, because
the pin has a very limited visibility. We must be careful do not incorporate
the background when we calculate the CT mean values.

Nominal contrast 0.5%

Kernel A Kernel B Kernel C
FBP 2.4± 0.3 1.8± 0.2 1.4± 0.2
idose 1 2.5± 0.2 2.0± 0.2 1.6± 0.2
idose 2 2.7± 0.2 2.1± 0.2 1.6± 0.1
idose 3 2.8± 0.2 2.3± 0.2 1.8± 0.1
idose 4 3.0± 0.1 2.4± 0.2 1.9± 0.1
idose 5 3.2± 0.1 2.6± 0.1 2.1± 0.1
Table 4.8: CNR values both standard and iterative reconstruction; nominal contrast 0.5%.

In this case, the variations for kernel B and C are 0.4 (FBP) and 0.5
(iterative algorithm).
Obviously, also in this case, the filter A maintains an higher value, respect-
ing the intrinsic characteristics of contrast resolution (i.e, this filter allows to
study better low contrast lesions).

Figure 4.19: Contrast to noise ratio for iterative and standard algorithm.
4.1 Catphan 600 phantom 107

The standard deviation obtained when smaller pin sizes were used, were
significantly increased compared to a nominal contrast of 1% and 2%. The
increase of SD as resulted in a decrease of the CNR of the images acquired
when using filtered back projection compared to iterative reconstruction.

Figure 4.20: Trends of contrast to noise ratio at varying levels of reconstruction.

In this plot, the trends of CNRs are almost linear, except for the range
between FBP and level 2 of iDose (kernel B and C).
If we consider the sharp filter, we notice slight differences from linearity
also in the transition from level 4 to level 5 of iterative reconstruction.
Chapter 5

Dose Assessment

CT and dose, far too often a book of mystery for many of those who have to
deal with this imaging modality. When asking somebody for the radiation
exposure from a given CT examination, the casual answer is: ”so and so
many mAs”. Even in scientific publications the applied current-time product
is used as a synonym for radiation exposure.
This point of view is not completely wrong as there is a linear relationship
between the applied tube current-time product and radiation dose. However,
it is often not recognized that this relationship differs depending on the type
of scanner. Dose comparisons in terms of mAs statements are therefore
not appropriate in the field of CT and are far from allowing a reasonable
indication of the radiation exposure relative to that from conventional x-ray
projection techniques.
In this context it has turned out as very useful to distinguish between
local and integral dose quantities. Local dose quantities are indicators of the
intensity of the irradiation inside the limits of the irradiated body region[48].
Computed tomography Dose Index (CTDI), dose free-in-air on the axis
of rotation and organ dose are members of this group. In contrast, integral
dose quantities. such as dose-lengh product (DLP) and effective dose, are
descriptors of the total amount of radiation absorbed by taking into account
also the extent of the body region being irradiated.

5.1 Computed Tomography Dose Index

The CTDI has for many years commonly been used as the most-specific dose
quantity for CT. Whenever, as usually occurs in practice, several adjacent
slices are scanned instead of a single slice, the dose for a particular slice is
increased due to the contributions from slices in its neighbourhood.

109
110 Dose Assessment

If the examination is performed with overlapping slices, i.e. by using a

table feed smaller than the slice thickness, the increase in dose becomes even
larger. The packing factor is used as an indicator of the degree of overlap,
which is roughly given by the ratio of slice thickness and table feed, i.e. pitch.

Figure 5.1: Illustration of the term ”Computed Tomography Dose Index”.

Fig.[5.1] illustrates the meaning of this term: the CTDI is the equivalent
of the dose value inside the irradiated slice that would result if the absorbed
radiation dose profile were entirely concentrated to a rectangular profile of
width equal to the nominal slice thickness. All dose contributions from out-
side the nominal slice width are added to the area inside the slice[49].
The corresponding mathematical definition of CTDI therefore describes
the summation of all dose contributions along a line which is parallel to the
axis of rotation for the scanner
1 +∞
Z
CT DI = D(z)dz [mGy] (5.1)
h −∞
where D(z) is the value of the dose at a given location z, and h is the
nominal slice thickness. CTDI is therefore equal to the area of the dose
profile (the DLP) divided by the nominal slice thickness.
Weighted CTDI is a weighted mixture of the pair of CTDI100 values, with
weightings of 31 for the central CTDI100,c and 23 for the peripheral CTDI100,p :
1 2
CT DIw = CT DI100,c + CT DI100,p (5.2)
3 3
CTDIw must be calculated separately for both the head and body phan-
toms. It is important to differentiate between absolute and normalized values
5.2 Dose Length Product 111

of CTDIw . The only advantage is that it enables the use of a single number
instead of two, particularly in the case of the body region where the central
and peripheral values are not of the same magnitude.

5.2 Dose Length Product

CTDI, weighted CTDI and axial dose free-in-air are by definition only indi-
cators of the level of local dose in the irradiated slice. Dose is an indicator
of the intensity of irradiation inside the irradiated part of the body, but only
here. If an organ, such as the liver, is already completely situated inside
the scan range, then the dose to the liver remains the same even if the scan
range is further extended. Only in those cases where an organ was partially
irradiated, will the organ dose grow with the increasing scan length, but only
once again until the organ is fully irradiated[50].
In conventional projection radiography, the quantity dose area product
(DAP) is used ti express both aspects of an irradiation, intensity and ex-
tent. The analogy for CT, where the diameter of body is always completely
irradiated, is dose length product (DLP).
DLP is obtained by multiplying one of the dose quantities appropriate
for CT wit the product pf the number of slices n and the slice thickness h:
DLPx,y,z = CT DIx,y,z · n · h [mGy · cm] (5.3)
The suffix ”w” or ”air” tells us which kind of CTDI was used. This differ-
ence is important if DLP is used to assess effective dose. In this definition of
DLP, pitch-related effects have already been taken into account, since instead
of the scan length L, the product of the number of slices and their thickness
is used.
Since the pitch factor is already implicitly contained in (5.3), some care is
necessary in order not to correct for pitch for a second time when calculating
DLP from the effective CTDI displayed at the operator’s console. In these
cases, the calculation of DLPw must be made using a different formula such
as
DLPw = CT DIw,ef f · p · n · h (5.4)
so as to convert from effective CTDI into weighted CTDI. It would appear
that the DLP discussed here is the same quantity as that already mentioned
in the context of the definition of CTDI. It is important to recognize that here
the DLP stands for the entire scan series, with ”length” meaning the dimen-
sion of the irradiated part of the body. In the case of the dose length product
for a single slice, ”length” means the dimension over which the contributions
from the dose profile are summed.
112 Dose Assessment

Figure 5.2: Illustration of the term ”Dose Length Product”.

In some examinations, more than one series of scans, is subsequently

made. Dose-relevant parameters such as mAs product, slice thickness and
scan length may differ from series to series. Therefore the DLP should be
calculated separately for each scan series. The total radiation exposure for
the complete examination is obtained by simply adding the contribution from
each series.

5.3 CTDI and DLP Measurements

The CT Head/Body/Pediatric CTDI (Computed Tomography Dose Index)
Phantom, in combination with a specialized CT-ion chamber, provides a
means of determining the approximate dose to the patient for a given series
of scans. The CT head and body phantoms are designed in accordance with
the FDA standard for diagnostic x-ray units, specifically as applied to CT
systems.

Figure 5.3: Phantom kit to evaluate CTDI (LEFT) and internal view with pencil chamber (RIGHT).

The phantoms can be used with any computed tomography system de-
signed to image both adult and pediatric head and body. They can separately
dose information for each. When performing dose profile measurements, the
5.3 CTDI and DLP Measurements 113

dose phantoms allow the user to collect information for the maximum, min-
imum, and mid-range value of the nominal tomographic sections thickness.
This essential phantom kit consists of three parts: an adult body phan-
tom, an adult head phantom that doubles as a pediatric body phantom, and
a pediatric head phantom. All are made of solid acrylic and 15 cm thick
with diameters of 32 cm, 16 cm, and 10 cm respectively. Each part contains
five probe holes-one in the center and four around the perimeter-that are 90
degrees apart and 1 cm from the edge. The inside diameter of the holes is
1.3 cm. Each part includes five acrylic rods for plugging all the holes in the
phantom.
In this case, since it uses a long pencil-like detectors with active lengths
of 100 mm, Eq.(5.1) becomes
1 +50
Z
CT DI100 = D(z)dz [mGy] (5.5)
h −50
CTDI100 has been generally accepted as a standard CT dose descriptor.
Finally, CTDI obtained by Dosewatch software can be written as
CT DIw
CT DIvol = [mGy] (5.6)
pitch
Whereas CTDIw represents the average absorbed radiation dose over the
x and y directions at the center of the scan from a series of axial scans where
the scatter tails are negligible beyond the 100 mm integration limit, CTDIvol
represents the average absorbed radiation dose over the x, y and z directions.

5.3.1 Protocols and Method

The CTDIvol is an objective technical dose parameter based on a measured
quantity. It takes into account protocol-specific parameters and is useful
to compare different scan protocols across various CT scanners. Thus, IEC
standards require the prospective display of the CTDIvol on the console of
the CT scanner.
To represent the overall dose of a given scan protocol (in this case ab-
domen, chest and head), the CTDIvol is multiplied with the examination
range which then yields the DLP.
We utilized Dosewatch GE software, that retrieves, tracks and reports
the radiation dose administrated to patients during medical exams and au-
tomatically organizes the data for hospital leaders. The system collects data
may be analyzed in different way, from imaging device. It can than compare
and contrast the CTDI and DLP being administered in one exam to another
in the past.
114 Dose Assessment

We consider the main protocols on which is applied the iterative recon-

struction algorithm and evaluate CTDI and DLP to compare with filtered
back projection.
Table 5.1 shows the sample exams taken into account to calculate CTDI
and DLP, which we than compared with the national DLR values (dose refer-
ence values for CT examinations, D.L. 187/00). For iterative algorithm, each
of these protocols uses a particular convolution kernel that has reconstructed
them on various levels.
Sample exams
Protocols Standard Iterative DLR CTDI DLR DLP
CT Head 2896 112 60 1050
CT Abdomen 6285 1163 35 780
CT Chest 9364 1968 30 650

Table 5.1: Number of exams analyzed to calculate Computed Tomography Dose Index and Dose
Length Product.

Standard Algorithm
Protocols Kernel CTDI DLP
[mGy] [mGy · cm]
CT Head SOFT 64± 5 1140± 200
CT Abdomen SOFT 20± 7 840± 390
CT Chest DETAIL 18± 6 563± 220

Table 5.2: Comparison of calculated CTDIvol and DLP mean values for different protocols, by using
standard reconstruction algorithm.

Iterative Algorithm
Protocols Kernel iDose level CTDI DLP
[mGy] [mGy · cm]
CT Head SOFT 2 60± 1 1014± 58
CT Abdomen SOFT 3 12± 5 520± 230
CT Chest DETAIL 4 10± 4 390± 130

Table 5.3: Comparison of calculated CTDIvol and DLP mean values for different protocols, by using
iterative reconstruction algorithm.

Tabs 5.2 and 5.3 show the examined protocols and the results obtained
with standard reconstruction and iterative reconstruction.
5.3 CTDI and DLP Measurements 115

• The standard algorithm show a good agreement for CT chest and CT

abdomen with DLR ; while the value of CT head is to the limit with
reference values.

• The iterative algorithm show a good agreement for all three protocols,
but the most important thing that we may notice is a reduction in the
values of CTDI and DLP compared to the standard algorithm.

CTDI and DLP values were extrapolated on a large numbers of patients,

subjected to the same type of examinations.
The variations on abdomen and chest regions is due to the patient size
variability.
Errors in DLP values take into account both patient size and variability
in multiphase length, in particular for abdomen.
Chapter 6

Conclusions

In this thesis we studied the approaches to the objective assessment of the

quality of iteratively reconstructed CT images, compared to filtered back
projection (FBP). The work presented here provides an objective and robust
way to quantitatively assess the image quality impact of newly introduced CT
iterative reconstruction (IR) analyzing the noise power spectrum, the mod-
ulation transfer function and the contrast-to-noise ratio. The main results
obtained have been:
• According to our measures, the performance of the IR shows that the
image quality is better than the traditional FBP algorithm. Our results
indicate that iDose4 IR can be another tool allowing for a reduction of
noise. Scans reconstructed with iDose4 have a higher CNR compared to
filtered back projection. The ability of iDose4 to improve low-contrast
object detection depends on the levels of reconstruction.
• Our MTF measurements showed that the high-contrast resolution not
differs for the FBP and iDose4 reconstructions. Spatial resolution was
largely unaffected by the IR levels.
• iDose4 allows us to tailor our CT scanning protocols to each clinical
indication, resulting in optimal image quality with the lowest radiation
dose. This algorithm reduces the radiation dose compared to FBP and
national DLR values.
The potential for patient dose reduction is generally considered the biggest
advantage of IR methods, and applications such as CT imaging, where dose
figures are of great concern, are a prime target. For this reason, an optimiza-
tion process with a Radiologist and a Physicist should be made to determine
the appropriate dose to obtain diagnostic image quality for the particular
clinical task.

117
118 Conclusions

This suggests that further study are necessary to analyze image quality
and dose patient. In addition, further verification using patient data and
human observer study are needed to fully develop low dose patient scan
protocols using IR algorithm. Real time reconstruction speed, the availability
to all CT imaging modes, and the large dose saving make this IR method a
good candidate to replace FBP as the routine reconstruction method used
in CT.
Appendix A

Images, tables and surface plot

A.1 Body Phantom images

Figure A.1: Filtered Back Projection reconstruction with kernel A (LEFT). Iterative reconstruction
with kernel A (RIGHT).

Figure A.2: Filtered Back Projection reconstruction with kernel B (LEFT). Iterative reconstruction
with kernel B (RIGHT).

119
120 Images, tables and surface plot

Figure A.3: Filtered Back Projection reconstruction with kernel C (LEFT). Iterative reconstruction
with kernel C (RIGHT).

Figure A.4: Filtered Back Projection reconstruction with kernel D (LEFT). Iterative reconstruction
with kernel D (RIGHT).

Figure A.5: Filtered Back Projection reconstruction with kernel DH (LEFT). Iterative reconstruction
with kernel DH (RIGHT).
A.2 Head Phantom images 121

A.2 Head Phantom images

Figure A.6: Filtered Back Projection reconstruction with kernel A (LEFT). Iterative reconstruction
with kernel A (RIGHT).

Figure A.7: Filtered Back Projection reconstruction with kernel UB (LEFT). Iterative reconstruction
with kernel UB (RIGHT).

Figure A.8: Filtered Back Projection reconstruction with kernel EB (LEFT). Iterative reconstruction
with kernel EB (RIGHT).
122 Images, tables and surface plot

Figure A.9: Filtered Back Projection reconstruction with kernel C (LEFT). Iterative reconstruction
with kernel C (RIGHT).

Figure A.10: Filtered Back Projection reconstruction with kernel D (LEFT). Iterative reconstruction
with kernel D (RIGHT).

Figure A.11: Filtered Back Projection reconstruction with kernel DH (LEFT). Iterative reconstruc-
tion with kernel DH (RIGHT).
A.3 NNPS tables both FBP and iterative algorithm, head and
body phantoms. 3D surface plot. 123
A.3 NNPS tables both FBP and iterative al-
gorithm, head and body phantoms. 3D
surface plot.

Figure A.12: Average NNPS both Filtered Back Projection and Iterative reconstruction algorithm,
head phantom. Spatial Frequency [mm−1 ], NNPS [mm2 ].

Figure A.13: Average NNPS both Filtered Back Projection and Iterative reconstruction algorithm;
body phantom. Spatial Frequency [mm−1 ], NNPS [mm2 ].
124 Images, tables and surface plot

Figure A.14: Noise texture fluctuations of Filtered Back Projection algorithm with convolution kernel
B (LEFT). Noise texture fluctuations of Iterative reconstruction algorithm (iDose, level 4) with same
convolution kernel of FBP (RIGHT). The teflon insert is not affected by noise texture and reconstruction
algorithm. Body phantom.

Figure A.15: Noise texture fluctuations of Filtered Back Projection algorithm with convolution kernel
C (LEFT). Noise texture fluctuations of Iterative reconstruction algorithm (iDose, level 4) with same
convolution kernel of FBP (RIGHT). The teflon insert is not affected by noise texture and reconstruction
algorithm. Body phantom.
A.3 NNPS tables both FBP and iterative algorithm, head and
body phantoms. 3D surface plot. 125

Figure A.16: Noise texture fluctuations of Filtered Back Projection algorithm with convolution kernel
D (LEFT). Noise texture fluctuations of Iterative reconstruction algorithm (iDose, level 4) with same
convolution kernel of FBP (RIGHT). The teflon insert is not affected by noise texture and reconstruction
algorithm. Body phantom.

Figure A.17: Noise texture fluctuations of Filtered Back Projection algorithm with convolution kernel
DH (LEFT). Noise texture fluctuations of Iterative reconstruction algorithm (iDose, level 4) with same
convolution kernel of FBP (RIGHT). The teflon insert is not affected by noise texture and reconstruction
algorithm. Body phantom.
126 Images, tables and surface plot

Figure A.18: Noise texture fluctuations of Filtered Back Projection algorithm with convolution kernel
UB (LEFT). Noise texture fluctuations of Iterative reconstruction algorithm (iDose, level 4) with same
convolution kernel of FBP (RIGHT). Head phantom.

Figure A.19: Noise texture fluctuations of Filtered Back Projection algorithm with convolution kernel
EB (LEFT). Noise texture fluctuations of Iterative reconstruction algorithm (iDose, level 4) with same
convolution kernel of FBP (RIGHT). Head phantom.
Appendix B

CNR Plot → Catphan 600

Figure B.1: CNR values for kernel UB; nominal contrast 1%.

127
128 CNR Plot → Catphan 600

Figure B.2: CNR values for kernel EB; nominal contrast 1%.

Figure B.3: CNR values for kernel A; nominal contrast 1%.

 129

Figure B.4: CNR values for kernel C; nominal contrast 1%.

Figure B.5: CNR values for kernel UB; nominal contrast 0.5%.
130 CNR Plot → Catphan 600

Figure B.6: CNR values for kernel EB; nominal contrast 0.5%.

Figure B.7: CNR values for kernel A; nominal contrast 0.5%.

 131

Figure B.8: CNR values for kernel C; nominal contrast 0.5%.

Figure B.9: CNR values for kernel UB; nominal contrast 0.3%.
132 CNR Plot → Catphan 600

Figure B.10: CNR values for kernel EB; nominal contrast 0.3%.

Figure B.11: CNR values for kernel A; nominal contrast 0.3%.

Appendix C

CNR plot → CIRS 061

Figure C.1: CNR values for kernel A; nominal contrast 2%.

133
134 CNR plot → CIRS 061

Figure C.2: CNR values for kernel B; nominal contrast 2%.

Figure C.3: CNR values for kernel C; nominal contrast 2%.

 135

Figure C.4: CNR values for kernel A; nominal contrast 1%.

Figure C.5: CNR values for kernel B; nominal contrast 1%.

136 CNR plot → CIRS 061

Figure C.6: CNR values for kernel C; nominal contrast 1%.

Figure C.7: CNR values for kernel A; nominal contrast 0.5%.

 137

Figure C.8: CNR values for kernel B; nominal contrast 0.5%.

Figure C.9: CNR values for kernel C; nominal contrast 0.5%.

Bibliography

[1] J. Beutel, H.L. Kundel, R.L. Van Metter “Handbook of Medical Imaging
”, SPIE Press, (2000).

[2] International Commission “Radiation Dose and Image Quality Assess-

ment in Computed Tomography ”, Oxford University Press, (2012).

[3] G.F. Knoll,“Radiation Detection and Measurements”, Wiley, (2000).

[4] N.A. Dyson, “X-Rays in Atomic and Nuclear Physics”, Cambridge Uni-
versity Press, (1990).

[5] P.R. Griffiths, “Handbook of Vibrational Spectroscopy”, Wiley, (2006).

[6] L.W. Goldman, “Principles of CT and CT Technology”, Journal of Nu-

clear Medicine Technology, (2007).

[7] L. Pierotti, “Evoluzione dei sistemi TC ”, Impact Scan.org, (2013).

[8] L.W. Goldman, “Principles of CT: Multislice CT ”, Journal of Nuclear

Medicine Technology, (2008).

[9] J. Hsieh, B. Nett, Z. Yu, K. Sauer, J.B. Thibault, C.A. Bouman, “Recent
Advances in CT Image Reconstruction”, Springer-Science, (2013).

[10] A.H. Hara, R.G. Paden, A.C. Silva, “Iterative Reconstruction Technique
for Reducing Body Radiation Dose at CT: Feasibility Study”, American
Journal of Roengtenology, 2009.

[11] B.P. Fahimian, Y. Zhao, Z. Huang, . Fung, Y. Mao, C. Zhu, M. Kha-

tonabadi, J.J. DeMarco, S.J. Osher, M.F. McNitt-Gray and J. Miao,
“Radiation dose reduction in medical x-ray CT via Fourier-based itera-
tive reconstruction”, Medical Physics, 2013.

[12] H. Shi, S. Luo, “A novel scheme to design the filter for CT reconstruction
using FBP algorithm”, BioMedical Engineering OnLine, 2013.

139
140 BIBLIOGRAPHY

[13] M. Beister, D. Kolditz, W.A. Kalender, “Iterative Reconstruction

method in X-ray CT ”, European Journal of Medical Physics, 2012.

[14] S. Kaczmarz, “Angenäherte Auflösung von Systemen Linearer Gleichun-

gen”, Bull Internat Acad Polonsci Lettres A, 1937.

[15] K.E. Jang, J. Lee, Y. Sung, S. Lee, “Information-theoretic discrepancy

based iterative reconstructions (IDIR) for polychromatic x-ray tomogra-
phy”, Medical Physics, 2013.

[16] L. Liu, “Model-based Iterative Reconstruction: A Promising Algorithm

for today’s Computed Tomography Imaging”, Journal of Medical Imag-
ing and Radiation Sciences, 2014.

[17] H.W. Tseng, J. Fan, M.A. Kupinski, P. Sainath, J. Hsieh “Assessing im-
age quality and dose reduction of a new x-ray computed tomography iter-
ative reconstruction algorithm using model observers”, Medical Physics,
2014.

[18] A. Hara, “ASiR reconstruction sharpens images, slices abdominal CT

dose”, Mayo Clinic Arizona, 2009.

[19] D.A. Kaplan, “Iterative reconstruction in CT evolves for lower dose,

increased clarity”, DiagnosticImaging.com, 2011.

[20] H.S. Park, H.M. Cho, J. Jung, C.L. Lee, H.J. Kim “Comparison of the
image Noise Power Spectra for Computed Tomography”, Journal of the
Korean Physical Society, 2008.

[21] M.F. Kijewski, P.F. Judy, “The noise power spectrum of CT images”,
Physics in Medicine and Biology, 1986.

[22] O. Rampado, “Interconfronto metodi di valutazione NPS ”, Review pre-

sentata alla città della scienza e della salute di Torino, 2014.

[23] O. Rampado, “Algoritmi Iterativi: Review Letteratura”, Review presen-

tata alla città della scienza e della salute di Torino, 2014.

[24] O. Rampado, “Interconfronto metodi NPS: fantocci ”, Review presentata

alla città della scienza e della salute di Torino, 2014.

[25] K.M. Hanson, “Noise and contrast discrimination in computed tomog-

raphy”, Radiology of the skull and Brain, 1981.
BIBLIOGRAPHY 141

[26] L. Berta, L. Mascaro, P. Feroldi, R. Maroldi “Optimisation of an MDCT

abdominal protocol: Image quality assessment of standard vs. iterative
reconstruction”, Physica Medica, 2013.
[27] K.L. Boedeker, V.N. Cooper, M.F. McNitt-Gray “Application of the
noise power spectrum in modern diagnostic MDCT: part 1. Measurement
of noise power spectra and noise equivalent quanta”, Physics in Medicine
and Biology, 2007.
[28] K.L. Boedeker, M.F. McNitt-Gray “Application of the noise power spec-
trum in modern diagnostic MDCT: part 2. Noise power spectra and sig-
nal to noise”, Physics in Medicine and Biology, 2007.
[29] S.N. Friedman, G.S.K. Fung, J.H. Siewerdsen, B.N.W. Tsui, “A simple
approach to measure computed tomography (CT) modulation transfer
function (MTF) and noise power spectrum (NPS) using the American
College of Radiology (ACR) accreditation phantom”, Medical Physics,
2013.
[30] J.H. Siewerdsen, I.A. Cunningham, D.A. Jaffray, “A framework for noise
power spectrum analysis of multidimensional image”, Medical Physics,
2002.
[31] K. Yang, A.L.C. Kwan, S. Ying Huang, N.J. Packard, J.M. Boone,
“Noise power properties of a cone-beam CT system for breast cancer
detection”, Medical Physics, 2008.
[32] J.T. Dobbins, E. Samei, N.T. Ranger, Y. Chen, “Intercomparison of
methods for image quality characterization: Noise power spectrum”,
Medical Physics, 2006.
[33] C. Ghetti, O. Ortenzia, G. Serreli, “CT Iterative reconstruction in image
space: A phantom study”, Physica Medica, 2011.
[34] J.T. Dobbins, E. Samei, N.T. Ranger, Y. Chen, “Intercomparison of
methods for image quality characterization: Modulation transfer func-
tion”, Medical Physics, 2006.
[35] J.Y. Vaishnav, W.C. Jung, L.M. Popescu, R. Zeng, K.J. Myers, “Ob-
jective assessment of image quality and dose reduction in CT iterative
reconstruction”, Medical Physics, 2014.
[36] E. Buhr, S.G. Kohfahl, U. Neitzel, “Accuracy of a simple method for
deriving the presampled modulation transfer function of a digital radio-
graphic system from an edge image”, Medical Physics, 2003.
142 BIBLIOGRAPHY

[37] I.A. Cunningham, B.K. Reid, “Signal and noise in modulation trans-
fer function determinations using the slit, wire, and edge techniques”,
Medical Physics, 1992.

[38] K. Li, J. Zambelli, N. Bevins, Y. Ge and G.H. Chen, “Spatial resolution

characterization of differential phase contrast CT systems via modula-
tion transfer function (MTF) measurements”, Physics in medicine and
biology, 2013.

[39] S.M. Akbari, M.R. Ay, A.R. Kamali, H. Ghadiri, H. Zaidi “Experimental
measurements of modulation transfer function (MTF) in five commercial
CT scanners”, XII Mediterranean Conference on Medical and Biological
Engineering and Computing, 2010.

[40] J.M. Wilson, O.I. Christianson, S. Richard, E. Samei “A methodology

for image quality evaluation of advanced CT systems”, Medical Physics,
2013.

[41] A. Love, M.L. Olsson, R. Siemund, F. Stalhammar, I.M. Bjorkman-

Burtscher, M. Soderberg, “Six iterative reconstruction algorithms in
brain CT:a phantom study on image quality at different radiation dose
levels”, Journal of Radiology, 2013.

[42] J. Hausleiter, S. Martinoff, M. Hadamitzky, “Image quality and radiation

exposure with a low tube voltage protocol for coronary CT angiography:
result of the PROTECTION Trial ”, Cardiovascular imaging, 2010.

[43] D. Marin, R.C. Nelson,H. Barnhart, “Detection of pancreatic tumors,

image quality, and radiation dose during the pancreatic parenchymal
phase: effect of a low-tube-voltage, high-tube-current CT technique-
preliminary results”, Radiology, 2010.

[44] A. Love, M.L. Olsson, R. Siemund, F. Stalhammar, I.M. Bjorkman-

Burtscher, M. Soderberg, “Impact of a reduced tube voltage on CT an-
giography and radiation dose. Results of the PROTECTION study”,
Cardiovascular imaging, 2009.

[45] K. Tang, L. Wang, R. Li, X. Zheng, G. Cao “Effect of Low Tube Voltage
on Image Quality, Radiation Dose, and Low-Constrast Detectability at
Abdominal Multidetector CT: Phantom Study”, Journal of Biomedicine
and Biotechnology, 2012.
BIBLIOGRAPHY 143

[46] G. F. Acquah, B. Schiestl, A. Y. Cofie, J. O. Nkansah, M. Gustavsson

“Radiation dose reduction without degrading image quality during com-
puted tomography examinations: Dosimetry and quality control study”,
International Journal of Cancer and Oncology, 2014.

[47] K. L. Dobeli, S. J. Lewis, S. R. Meikle, D. L. Thiele, P. C. Brennan “Op-

timization of Computed Tomography Protocols: Limitations of a Method-
ology Employing a Phantom with Location-Known Opacities”, Journal
of Digital Imaging, 2013.

[48] E. Samei, N. T. Ranger, J. T. Dobbins, C. E. Ravin “Effective dose

efficiency: an application-specific metric of quality and dose for digital
radiography”, Physics in Medicine and Biology, 2011.

[49] S. Kim, H. Song, E. Samei, F. Yin, T. Yoshizumi “Computed Tomogra-

phy dose index and dose length product for cone-beam CT: Monte Carlo
simulations of a commercial system”, Journal of Applied Clinical Med-
ical Physics, 2010.

[50] C. Descamps, M. Gonzalez, E. Garrigo, A. Germanier, D. Venencia

“Measurements of the dose delivered during CT exams using AAPM
Task Group Report No. 111”, Journal of Applied Clinical Medical
Physics, 2012.