Magnetic Resonance Imaging 20 (2002) 437– 446

MRI inter-slice reconstruction using super-resolution

H. Greenspana,*, G. Oza, N. Kiryatia, S. Peledb
a
Faculty of Engineering, Tel Aviv University, Tel Aviv 69978, Israel
b
Tel Aviv Sourasky Medical Center, Tel Aviv, Israel

Received 14 February 2002; accepted 19 May 2002

Abstract
MRI reconstruction using super-resolution is presented and shown to improve spatial resolution in cases when spatially-selective RF
pulses are used for localization. In 2-D multislice MRI, the resolution in the slice direction is often lower than the in-plane resolution. For
certain diagnostic imaging applications, isotropic resolution is necessary but true 3-D acquisition methods are not practical. In this case, if
the imaging volume is acquired two or more times, with small spatial shifts between acquisitions, combination of the data sets using an
iterative super-resolution algorithm gives improved resolution and better edge definition in the slice-select direction. Resolution augmen-
tation in MRI is important for visualization and early diagnosis. The method also improves the signal-to-noise efficiency of the data
acquisition. © 2002 Elsevier Science Inc. All rights reserved.

Keywords: Echo-planar imaging (EPI); Magnetic resonance angiography (MRA); Fast spin echo (FSE); Signal-to-noise ratio (SNR); Superresolution; Thin
slices

1. Introduction A technique based on super-resolution is proposed in which

several sets of 2-D slices are merged, to provide the high-
resolution 3-D image. The method we present consists of
In cases in which true 3-D image acquisition is not
effective or possible in MRI, it is common practice to two stages. First, the acquisition of a small number of
acquire a set of 2-D slices. Such is sometimes the case, for multislice data sets, each volume shifted by a sub-pixel
example, in T2-weighted imaging, diffusion-weighted im- amount in the slice-select direction with respect to the other
aging, and MR angiography. These are all imaging tech- volumes. Second, the use of super-resolution post-process-
niques that are important for early medical diagnosis and ing in the inter-slice ( z) dimension. Experimental validation
visualization purposes, and usually require coverage of ex- of the technique shows that the resolution in the z direction
tensive 3-D volumes in the imaged subject. The problem, as is significantly improved. We show also that the SNR effi-
illustrated in Fig. 1, is that a set of 2-D slices does not give ciency (SNR per unit acquisition time) of image data sets
a good isotropic 3-D image. MRI slice thickness is deter- reconstructed using super-resolution is better than that of
mined by hardware limitations coupled with pulse sequence images acquired directly with higher spatial resolution in
timing considerations. This results in resolution that is high the slice-select direction. Preliminary results appeared in a
in-plane and is lower in the slice-select (or “through-plane”) short abstract [1].
direction. Even data acquisition in all three orientations, The structure of the paper is as follows: In section 2 a
would not provide self consistent 3-D data sets viewable in brief overview of MRI is given, with a specific focus on 2-D
any oblique orientation. vs. 3-D MRI spatial encoding techniques. Section 3 dis-
In this work we address the challenge of achieving high- cusses super-resolution algorithms. In section 4 the appli-
resolution, isotropic 3-D MRI images from 2-D MRI slices. cation of super-resolution algorithms to MRI data sets is
investigated. Experimental results using echo-planar and
fast spin-echo imaging methods are given in section 5.
Computational aspects and quantitative performance evalu-
* Corresponding author. Tel.: ⫹972-3-6407398; fax: ⫹972-3-
6407939. ation are discussed in section 6 and conclusions are drawn
E-mail address: [email protected] (H. Greenspan). in section 7.

0730-725X/02/$ – see front matter © 2002 Elsevier Science Inc. All rights reserved.
PII: S 0 7 3 0 - 7 2 5 X ( 0 2 ) 0 0 5 1 1 - 8
438 H. Greenspan et al. / Magnetic Resonance Imaging 20 (2002) 437– 446

(FSE) methods [2]. The problem stems from the need for
relatively long signal recovery between excitations to en-
able operation of the spin-echo mechanism that provides T2
contrast. Since all the spins are excited by every pulse, the
recovery time cannot be utilized and the sequence takes a
long time.
Multi-slab 3-D Fourier imaging in which a number of
slabs are excited and each one is encoded by 3-D techniques
attempts to address the problem of wasted time in 3-D T2
imaging, but creates artifacts of its own. The artifacts in-
clude signal deterioration at slab edges, and increased point-
spread function and truncation artifacts due to the small
number of Fourier phase encodes in the slab-select dimen-
sion. A proposed solution to the problems of 3-D multi-slab
encoding involves the use of non-Fourier methods such as
Hadamard wavelets for encoding the third dimension [3].
Fig. 1. MRI slice acquisition. The resolution in the slice-select direction is
usually much lower than in the in-plane directions.
Time-of-Flight (TOF) MR angiography (MRA) is an-
other popular application that sometimes performs better in
the 2-D rather than the 3-D version, e.g., in peripheral
2. Background non-contrast angiography.
For the important MRI application of diffusion imaging,
Conventional two-dimensional multi-slice MR imaging no 3-D technique for humans currently exists. Sequences
usually relies on the selection of 2-D planes (or slices) using that acquire raw data pertaining to the same slice or volume
radio frequency (RF) pulses, and on in-plane encoding us- over many excitations cannot be modified to provide diffu-
ing frequency encoding of the signal in one direction and sion-weighted contrast because of motion artifacts resulting
phase encoding in the other in-plane direction. Usually each from phase inconsistencies in the data. Today, the most
slice requires multiple RF excitations to be applied to it in popular solution is to base diffusion imaging on 2-D single-
order to be fully encoded. For reasons of pulse sequence shot techniques—mostly echo-planar imaging (EPI). Even
timing, and signal-to-noise ratio (SNR), there are disadvan- the in-plane resolution here is sub-optimal, but usually bet-
tages to thin-slice excitation. RF pulses are finite in dura- ter than the resolution in the third dimension. Isotropic
tion, thus a perfectly rectangular slice excitation is not resolution is particularly important in diffusion tensor im-
feasible. The thinner the slice that should be excited by the aging—the technique used for delineating white matter fiber
RF pulse (i.e., the narrower the frequency band), the sharper tracts in the human brain in vivo [4].
the edges of the slice excitation profile are required to be. In cases in which 3-D acquisition is not possible, it is
This translates to longer pulse durations, given the same common practice to use a set, or sets, of 2-D slices. In this
magnetic field gradient strength. Thin slice excitation is work we propose a novel approach for achieving high-
particularly problematic when spectral-spatial pulses (used resolution isotropic 3-D images by merging sets of 2-D
for fat suppression) are applied. Long slice-selective pulses slices.
increase echo times which increases the signal loss caused Newly developed parallel imaging techniques (see, for
by dephasing of the magnetization. example, Ref. [5] and references therein) will allow faster
Three-dimensional acquisition techniques provide thin acquisition, or higher in-plane resolution, at the expense of
sections and the ability to view interpolated data at high SNR. Increasingly high commercial gradient strengths and
resolution from any angle. In true 3-D image acquisition, slew rates generated by local gradient coils will do the same.
the third dimension is also phase encoded, and the RF High resolution in plane will call for thin-slice acquisition
excitation pulses excite a thick slab defining the whole which could bring the SNR down to values requiring signal
imaging volume. Use of a short echo time is possible be- averaging in order to provide clinically useful images. The
cause of the short duration of the minimally-selective RF ability to use super-resolution post-processing of thick
excitation pulses. T1-weighted imaging is compatible with slices, as shown in this work, might in some cases provide
3-D acquisitions, and, in this case, a short repetition time a much needed extra boost to the SNR.
improves the image contrast. True 3-D acquisition may not
be advisable in other imaging cases, such as in T2-weighted
imaging, diffusion-weighted imaging, and occasionally in 3. Super-resolution algorithms
MR angiographic (MRA) imaging.
True T2-weighting is difficult to obtain in reasonable Super-resolution algorithms are a family of techniques
imaging times by 3-D acquisition methods, although for creating a high resolution image from several lower
progress is being made in this direction using fast-spin-echo resolution images of the same scene, taken from slightly
 H. Greenspan et al. / Magnetic Resonance Imaging 20 (2002) 437– 446 439

Fig. 2. Linear imaging system.

different viewpoints. The first to address the problem have the given data set of low resolution images. This is accom-
been Huang and Tsai [6]. They were followed by Ur and plished by minimizing the differences between the given
Gross [7], Irani and Peleg [8], Kim et al. [9], Tekalp et al. low resolution images, and low resolution versions of the
[10] and others. Super-resolution is a very active research high resolution image. The latter are obtained by simulating
area, motivated by emerging video technologies (e.g., [11]). the imaging process, taking geometric transformations and
A general model of an imaging system is given in Fig. 2. blur into account. Formally, if f is the high resolution image
K
A high-resolution scene undergoes a geometrical transfor- we wish to reconstruct, and { g k } k⫽1 are the given set of
mation and optical blur. The low-resolution image is the low resolution images, the imaging process can be modeled
outcome of sampling and additive noise. The principle of by:
super-resolution algorithms is to acquire several low-reso-
g k ⫽ 共T k共 f 兲ⴱh兲2s ⫹ ␩ k (1)
lution images { g k }, with slightly different imaging condi-
tions, and then estimate the high-resolution source, f, that where T k describes the geometric transformation between
best explains the low-resolution data. A variety of ap- the k-th image and the reference frame, h corresponds to the
proaches can be found in the super-resolution literature. imaging blur kernel, 2 s is a down-sampling operator that
decreases the resolution by the factor s, and ␩ k is an addi-
3.1. The Irani-Peleg algorithm tive noise term. A flowchart describing the Irani-Peleg al-
gorithm is shown in Fig. 3.
The algorithm we use is based on the iterative back- In the Irani-Peleg algorithm, p is a “back-projection” kernel
projection (IBP) method of Irani and Peleg [8]. A high (see Fig. 3). In order for the algorithm to converge, p should be
resolution image is constructed iteratively to best explain chosen so that the following condition holds: 储␦ ⫺ h ⴱ p储2 ⱕ 1 [8].

Fig. 3. Block diagram of the Irani-Peleg superresolution algorithm.

440 H. Greenspan et al. / Magnetic Resonance Imaging 20 (2002) 437– 446

Fig. 4. Acquisition of voxel-shifted data sets. The geometric relationship between low and high resolution frames in 3-D sampling of MRI data. The large
transparent boxes represent voxels in the low-resolution acquired data. The shaded box shows the position of a single high-resolution voxel within each of
the low-resolution data sets. The Field of View (FOV) defined by the acquisition parameters on the MR scanner is shifted in this case by half a voxel in each
direction. Thus, a voxel in the high resolution image volume is “covered” by 8 low resolution voxels.

Stability considerations may suggest that p be other than the sequence of 2-D video frames, or even from the incremental
inverse of h, at the expense of a slower convergence rate. generation of high resolution video frames from a low
resolution input sequence, as suggested by Elad and Feuer
[12]. The difference is the pooling, in true 3-D super-
4. Super-resolution in MRI resolution, of several 3-D data sets to simultaneously obtain
an entire 3-D higher resolution output set. An illustration of
Applying the super-resolution principle to MRI raises the acquisition of voxel-shifted data sets is shown in Fig. 4.
several fundamental as well as technical issues. It turns out The geometric relationship between low and high resolution
that the differences between an MRI system and a common frames in 3-D sampling is shown. The large transparent
camera hinder the straightforward extension of the super- boxes represent voxels in the low-resolution acquired data.
resolution algorithm to 3-D for use in MRI. These and other The shaded box shows the position of a single high-resolu-
issues specific to the application of super-resolution in MRI tion voxel within each of the low-resolution data sets. In the
are presented in this section. presented example, a voxel in the high resolution image
volume is “covered” by 8 low resolution voxels.
4.1. Super-resolution dimensionality The 3-D methodology is applicable to cases where:
1. Data acquisition covers a 3-D volume;
Super-resolution, for images and image sequences, has 2. Successive data acquisitions add new information in
customarily been treated as a 2-D problem. Note that super- each dimension.
resolution can also be applied to 1-D signals; the feasibility
of applying a 1-D super-resolution procedure to 2-D images, The extension of the Irani-Peleg algorithm from 2-D to 3-D
improving resolution in one dimension at a time, depends is conceptually straightforward, in principle requiring only
primarily on the separability of the physical 2-D blur kernel. 3-D models of the blur kernel h, the back-projection oper-
In our investigation of 3-D MRI data-sets we explored the ator p, the individual geometric transformations T k and their
possibility of extending super-resolution methodologies to inverse. Note that 2-D images (or 1-D “needles”) can be
3-D. The 3-D super-resolution methodology is a novel ap- extracted from the 3-D image.
proach for treating volumetric data. We introduce the gen-
eral 3-D methodology next, and pursue an investigation of 4.2. In-plane vs. inter-slice investigation
its applicability to MRI data-sets in the following sub-
section. An important issue for consideration when applying su-
The concept of obtaining a high resolution 3-D image per-resolution to MRI data, is the dimensionality of the
volume from a set of low resolution 3-D image volumes is problem. An examination of Fourier-encoded MRI data sets
different from the extraction of a 2-D still image from a reveals distinct characteristics of the in-plane vs. through-
 H. Greenspan et al. / Magnetic Resonance Imaging 20 (2002) 437– 446 441

Fig. 5. Spectrum analysis, ( y, z) plane. The horizontal axis is the slice-select direction. Top row: Low-resolution input (left), spectrum of low-resolution input
(right). Bottom row: High-resolution output (left), spectrum of high-resolution output (right).

plane encoding. These differences impose constraints on the z plane, where y is one of the in-plane coordinates, and z is
dimensions in which super-resolution can lead to useful the slice-select direction. The output of the super-resolution
results. process (double size in each dimension) is shown (bottom
Fourier-encoded in-plane MRI data is inherently band- left), with its power spectrum (bottom right). The sharp
limited. This is due to the time limit of the acquisition frequency cut-off in the y direction is evident. A comparison
mechanism and the fact that the information is gathered in between the low-resolution spectrum and high-resolution
the frequency domain (k-space). The spatial frequencies in spectrum shows that the attempt to augment the in-plane
the z (inter-slice) direction exhibit a less sharp cut-off. image resolution is equivalent to zero-padding extension of
There is sufficient information in the z dimension such that the frequency domain. A spreading-out of the power-spec-
sampling of the data in that direction results in aliasing, and trum is evident in the z direction.
thus provides the basis for using a super-resolution algo- The investigation leads to the following conclusions:
rithm in enhancing the resolution. First, without prior information about the data, the predic-
In order to verify that the above statements do in fact tion of higher spatial frequencies is not possible for Fourier-
reflect the reality of MR imaging, we acquired multislice encoded MR data. The best that can be done in the in-plane
2-D image data sets shifted in all three spatial directions by ( x and y dimensions) is to interpolate, via zero-padding, the
half a voxel (as shown in Fig. 4). A 3-D iterative super- given data to the desired resolution. Previous results that
resolution algorithm was applied. In Fig. 5, we see an show improved images after application of super-resolution
original low-resolution image (top left) with its original to the in-plane dimensions [13] may be, in most part, rep-
power spectrum (top right). This image is parallel to the y ⫺ licated by zero-padding. An important second observation is
that in the inter-slice direction, sub-voxel spatial shifts can
in fact be utilized to increase the resolution. It is in this
dimension that we apply the super-resolution algorithm.1
Fig. 6 shows the sub-voxel spatial shifts in the slice-
select direction, as used in this work. The input low-reso-
lution voxels are shown left, with the high-resolution voxels
via sub-voxel shifts, shown right. Experiments are con-
ducted with two, three and four shifts per voxel, enabling a
substantial augmentation in the slice-select resolution. The
number of shifts is determined so as to achieve (as much as
possible) isotropic resolution in all dimensions.

1
In principle, with non-Fourier encoding, other dimensions should also
Fig. 6. Acquisition with 1-D voxel shifts in the slice-select direction. be amenable to enhancement through super-resolution, although we have
Low-resolution voxels (left). High-resolution voxels (right). not investigated this route.
442 H. Greenspan et al. / Magnetic Resonance Imaging 20 (2002) 437– 446

4.3. Signal-to-noise ratio The difficulty here depends on the number of degrees of
freedom in the inter-image transformations that have to be
When considering a method for resolution improvement accommodated. In the Irani-Peleg algorithm, the geometric
in MRI, one should make sure that the signal-to-noise ratio transformations T k need to be known to sub-pixel accuracy,
(SNR) is not compromised. The signal-to-noise ratio (SNR) preferably to 1/10th of a pixel. Then the iterative process
of an MR image is often measured by taking the mean of a converges rapidly and provides good results. The threshold
high-intensity region of interest and dividing by the stan- used as a stopping condition in the Irani-Peleg algorithm
dard deviation of a region of noise outside the imaged was 2% (see Fig. 3).
object. For the purpose of our comparison between recon- In our case, the accuracy and stability in the field of view
struction methods, this definition is satisfactory. of MRI machines allows us to acquire a set of images of a
RF excitation of a volume or slice leaves the affected subject, each translated by a predefined sub-voxel vector
nuclear spins in a state of partial saturation. Multi-slice 2-D with respect to the 3-D reference frame. Depending on the
imaging takes advantage of the time required for signal physical orientation of the slice-select direction, the shift
recovery after RF excitation in order to excite and encode will be implemented either by shifting the patient bed, or by
many slices in an interleaved fashion. The time constant moving the RF transmitter center frequency. In the super-
associated with this signal recovery is referred to as T 1 . resolution algorithm, this eliminates the need for registra-
Thus increasing volume coverage by adding more slices tion, since all the geometric transformations T k (and their
increases the acquisition time approximately in proportion inverse) are known in advance. We are provided with the
to the number of slices, but may also improve the SNR due keys necessary for successful super-resolution processing:
to a longer repetition time (TR) between individual excita- all image volumes are brought to the same 3-D reference
tions of a particular slice. However, when many slices need frame reliably, accurately and without any computational
to be acquired, as is often the case in clinical settings, the
effort. It should be noted that we are assuming minimal
incremental improvement in SNR of adding more slices
movement of the subject. For the more general case of a
(thus adding time between consecutive excitations of the
moving imaged subject, the method proposed herein would
same slice) is insignificant. In such cases, serial acquisition
also benefit from image registration.
of 2 or 3 sets of a smaller number of thicker slices with
correspondingly shorter repetition times, as described in this
work, does not negatively affect the spin recovery. In such
5. Experimental results
cases we find a clear SNR advantage to using our method of
data acquisition with super-resolution post-processing.
In order to evaluate the inter-slice super-resolution meth-
odology in MRI, we carried out a number of experiments:
4.4. Parameterization
Y Phantom experiments, in which acquired MRI data
Successful implementation of the super-resolution algo- sets of inanimate objects, thus assuring no subject
rithm requires an estimate for the transformation and blur motion.
parameters that most closely correspond to the true imaging Y MR brain imaging experiments, in which we use
system parameters. human brain data as input to the super-resolution
The point-spread function (PSF), or blur, h, in the slice- algorithm.
select direction can be inferred from the slice excitation
profile. We measured typical slice profiles and found them All imaging was performed with an RF head coil on either
to be well approximated by Gaussian functions, where the a 1.5 Tesla General Electric Signa Echospeed MRI system
full width at half maximum (FWHM) was the originally or a 3 Tesla General Electric MRI system.
selected slice width. In our work we experiment with two We present a sample result from each of the experimen-
PSFs. The first is a rectangular pulse PSF (a crude estima- tal paradigms. The presented results include visual compar-
tion for the slice profile), hereon termed Box-PSF. The box isons between the input low-resolution image (zoomed to
width is taken as the selected slice width, in the desired the desired size of the high-resolution result), the low-
high-resolution pixel units. The second is a Gaussian PSF resolution image following Sinc interpolation (zero-pad-
(Gaussian-PSF), with FWHM set to the selected slice ding), the result of interleaving the sets of low-resolution
width. In the implementation of the Irani-Peleg algorithm, images and the result of using the super-resolution algo-
the h filter is taken as one of the above two PSFs. In both rithm on the low-resolution image set.
cases, the p filter is taken as an impulse function. The choice Interleaving is a method to achieve a high-resolution
for p satisfies the Irani-Peleg requirement for convergence. image from a set of shifted low-resolution images by com-
A key requirement for successful implementation of su- bining the pixels, one by one from alternating low-resolu-
per-resolution algorithms in general is precise image regis- tion image inputs, to generate a single large image. This
tration accurate to a small fraction of a pixel and capable of method will give best results when the PSF of the sensor is
bringing all the input images to a common reference frame. the same size as the high-resolution pixel. In our case, the
 H. Greenspan et al. / Magnetic Resonance Imaging 20 (2002) 437– 446 443

Fig. 7. Super-resolution applied to comb-phantom MRI data. The horizontal axis is the slice-select direction. Shown are: original low resolution data (a);
zero-padded data (b); interleaved slices (c); inter-slice direction improved by super-resolution (Box-PSF) (d); inter-slice direction improved by super-
resolution (Gaussian-PSF) (e).

PSF is much larger and its width is approximated by the size the comb-phantom MRI data. The horizontal axis is the
of the low-resolution pixel in the z direction. slice-select direction. In (a), the original low resolution
data is shown, followed by zero-padding interpolation
5.1. MRI resolution phantom (b), interleaving (c) and super-resolution results in the
inter-slice direction (d, e). The super-resolution results
The homemade phantom consists of long thin plastic include a box-PSF (d) and a Gaussian-PSF (e).
partitions (“teeth”), lodged in a plastic block, placed 4 mm Several observations may be made from the presented
apart, surrounded by Gd-DTPA-doped water. The imaging results. The visibility of the comb teeth has greatly im-
sequence consists of multislice fast spin-echo (FSE) with 16 proved by using super-resolution rather than zero-padding
slices, 3 mm thick, approximately parallel to the plastic interpolation. Moreover, super-resolution brings out more
partitions. Three sets of multislice data were acquired, with information than interleaving. The super-resolution brings
1 mm shifts in the slice-select direction. The low-resolution out features that are inseparable in the source images. The
input voxel size is 1 ⫻ 1 ⫻ 3 mm. Following the super- implementation of the super-resolution algorithm with a
resolution procedure, an output voxel will be a 1 mm iso- Gaussian-PSF (e) gives slightly better results than when
tropic cube. using a Box-PSF (d). The comparison between the two PSFs
Fig. 7 shows the results of super-resolution applied to is interesting: better estimation of the high resolution image
444 H. Greenspan et al. / Magnetic Resonance Imaging 20 (2002) 437– 446

Fig. 8. Human brain MRI. Shown are the low-resolution image (a), zero-padding interpolation (b), super-resolution result, Box-PSF (c), super-resolution
result, Gaussian-PSF (d).

is achieved by using a blurring filter, h, that more closely 1

matches with the MRI imaging system and the MRI image y共 x兲 ⫽ . (2)
1 ⫹ exp 共⫺a共 x ⫺ c兲兲
characteristics in the slice-select direction.
The parameter a is inversely proportional to the width, and
5.2. Human brain c corresponds to the center location. Following the fitting
step, a measure of “rise length” is computed, defined as the
In the following example, we show super-resolution on width (in high-resolution pixels) from 10% to 90% of the
human brain data. The imaging sequence was fast-spin-echo edge height. It is easy to show that:
(FSE) with 3 shifts in the slice direction. The slice thickness
4.4
was 4.5 mm, the in-plane resolution was 1.5 mm and the width [pixels] ⫽ . (3)
number of slices was 22. a
Fig. 8 shows the low-resolution image (a), zero-padding
For the quantitative analysis of super-resolution, we used
interpolation (b), super-resolution result with Box-PSF (c),
an apple as the input source (due to the well defined outer
and the super-resolution result with Gaussian-PSF (d).
edges and relative lack of structure close to those edges).
There is a clear improvement in the resolution of the images
Imaging was performed using the FSE sequence with an
as we progress from the low-resolution input to the Gauss-
echo train length of 16, a slice width of 4.5 mm with no gap
ian-PSF super-resolution result.
between them, and 3 shifts differing in spatial location by
1/3 of a pixel. The number of slices in each low resolution
data set was 22 and the repetition time between RF excita-
6. Image analysis—resolution and SNR efficiency tions of the same slice was 8 sec. A high resolution data set
of 66 contiguous slices covering the same volume was also
In addition to visual evaluation of the output images, acquired. We compared the width of 20 outer edges of the
quantitative measures of resolution and signal-to-noise ratio apple for 5 different data sets: (a) the input low resolution
(SNR) are computed and used to evaluate the performance images following sinc interpolation (zero padding), (b) in-
of the super-resolution algorithm. terleaved result, (c) super-resolution result, Box-PSF, (d)
super-resolution result, Gaussian-PSF and (e) high-resolu-
6.1. Spatial resolution tion source. A sample edge from the apple image is shown
in Fig. 9, along with the corresponding sigmoid function
Quantitative comparison of resolution is not a trivial that was fitted to it, for each of the cases compared. Results
process. Currently we are basing our resolution estimate on are summarized in Table 1.
the measurement of edge widths. A sample of edges are Several points may be learned from Table 1. Looking
selected in one image, and then consistently compared across the columns, we note that the resolution in edge
across all the images. The width of each edge is measured width improves as we shift from the zero-padded input to
by least-squares fitting it to a sigmoid function of the form: the high-resolution source. A clear improvement is present
 H. Greenspan et al. / Magnetic Resonance Imaging 20 (2002) 437– 446 445

Fig. 9. A sample edge from the apple image and the corresponding sigmoid function fitted to it, for each of three cases compared. The squares, circles and
triangles show points on the sample edge, for zero-padding interpolated low-resolution image, super resolution image and high resolution image, respectively.
The dashed, solid and dotted lines show the sigmoid corresponding to the low-resolution, super resolution and high resolution data, respectively.

in the edge-width of the super resolution result, in compar- sition is possible). The mean edge width of the Gaussian-
ison to the zero-padded low-resolution source and the in- PSF super-resolution result is (almost exactly) identical to
terleaved results. The super-resolution process with Box- the mean width of the edges in the high-resolution source.
PSF gives high-resolution results but not as high as a true This statistic indicates a successful augmentation of the
high-resolution acquisition (in cases where such an acqui- image resolution via the super-resolution procedure.

Table 1
Quantitative measures of SNR and resolution

Super-resolution Super-resolution
Zero-padded Interleaved Box-PSF Gaussian-PSF High resolution

Acquisition time (min:sec) 1:28 4:24 4:24 4:24 4:00

SNR 287 276 170 124 95

Edge widths (pixels) 2.9 3.0 2.5 1.9 1.9

(7 of 20) 3.9 3.7 3.2 2.6 2.3
3.7 3.5 2.6 2.0 2.3
3.5 3.8 3.3 2.8 2.7
3.3 2.8 2.1 1.8 1.0
4.0 4.2 3.1 2.4 3.1
3.5 3.8 3.1 2.3 1.8
: : : : :

Mean (edge width pixels) 3.7 3.7 2.9 2.2 2.3

Comparison to ⫹58% ⫹57% ⫹25% ⫺3% 0%

high resolution

SNR values are given as absolute ratio values. Resolution is measured as edge widths in the z direction. Results are given for an apple input. The edge
widths in each row correspond to the same edge. In the two last rows the mean edge width in each data set is listed, as well as the percentage difference
from the high-resolution data.
446 H. Greenspan et al. / Magnetic Resonance Imaging 20 (2002) 437– 446

6.2. SNR efficiency resolution methodology is suggested for general volumetric

data-sets.
The SNR values for the corresponding images decrease In conclusion, we are proposing a novel framework of
as expected with the increased resolution. As one collects using super-resolution algorithms in MRI resolution aug-
more data in the MRI process (a long sequence), one obtains mentation. The method is easily implementable because it is
a result with higher SNR for the same resolution. The ratio based on a simple reconstruction algorithm and on existing
between the SNR of the result and the square root of the pulse sequences. In the MRI application domain super-
time length of the data acquisition sequence is called SNR resolution may be of significant potential importance.
Efficiency. In an MRI process the goal is to obtain high
resolution images with a high SNR efficiency. Table 1
shows the SNR obtained from each of the methods under Acknowledgments
comparison.
The sequence time of the high-resolution source was This research was supported in part by the Ela Kodesz
4:00 minutes (⫽240 sec) and the sequence time for each of Institute for Medical Engineering and Physical Sciences,
the low resolution data sets was 1:28 min which gives 4:24 and by the Adams Super-Center for Brain Studies, Tel-Aviv
minutes (⫽264 sec) for complete collection of three shifts. University.
Division of the SNR by the square root of the image acqui-
sition times gives a SNR efficiency of 7.63 sec⫺1/2 for the
super-resolution result, and 6.13 sec⫺1/2 for the high-reso- References
lution acquisition. It is possible to state from these results
that the super-resolution SNR efficiency is better than a [1] Greenspan H, Peled S, Oz G, Kiryati N. MRI Inter-slice reconstruc-
high-resolution MRI collection (at the same effective reso- tion using super-resolution. In: Proceedings of MICCAI 2001, Fourth
lution as measured by the edge widths). International Conference on Medical Image Computing and Comput-
er-Assisted Intervention (short abstract), Lecture Notes in Computer
Science, Springer, October 2001.
[2] Mugler JP, Bao S, Mulkern RV, Guttmann CRG, Robertson RL,
7. Conclusion Jolesz FA, Brookeman JR. Optimized single-slab three-dimensional
spin-echo MR imaging of the brain. Radiology 2000;216:891–9.
This paper presents and demonstrates MRI inter-slice [3] Goelman G. Fast 3D T2-weighted MRI with Hadamard encoding in
reconstruction using super-resolution. With inter-slice su- the slice select direction. Magn Reson Imag 2000;18:939 – 45.
[4] Basser PJ, Pajevic S, Pierpaoli C, Duda J, Aldroubi A. In vivo fiber
per-resolution we may be able to effectively break limits on
tractography using DT-MRI data. Mag Reson Med 2000;44:625–32.
slice thickness posed by the physical properties of existing [5] Madore B, Pelc NJ. SMASH and SENSE: Experimental and numer-
MR imaging hardware. ical comparisons. Magn Reson Med 2001;45:1103–11.
For large volume coverage using multislice 2-D imaging [6] Huang TS, Tsai RY. Multi-frame image restoration and registration.
techniques, the acquisition of multiple data sets does not add In: Advances in computer vision and image processing, Vol. 1. JAI
Press Inc., 1984. pp. 317–39.
significant time to the data acquisition when compared to
[7] Ur H, Gross D. Improved resolution from subpixel shifted pictures.
the acquisition of thin slices. In such cases, the SNR effi- CVGIP: Graphical Models and Image Processing 1992;54:181– 6.
ciency of this method appears to be better than 2-D thin [8] Irani M, Peleg S. Motion analysis for image enhancement: resolution,
slice acquisitions. From a purely theoretical point of view, occlusion, and transparency. J Vis Comm Image Rep 1993;4(4):324 –
the above finding can be explained by the following argu- 35.
[9] Kim SP, Bose NK, Valenzuela HM. Recursive reconstruction of high
ment: acquisition of n sets of 2-D slices that are n times
resolution image from noisy undersampled multiframes. IEEE Trans-
thicker results in a SNR per unit volume that is 公n greater actions on Acoustics Speech and Signal Processing 1990;38(6):1013–
than that of a similarly timed total acquisition of one set of 27.
thin slices. In any case, super-resolution provides the flex- [10] Tekalp AM, Ozkan MK, Sezan MI. High resolution image recon-
ibility to determine the extent of the trade-off between the struction for lower resolution image sequences and space-varying
image restoration. In: IEEE International Conference on Acoustics,
high SNR of the thick slices, and the high anatomical
Speech, and Signal Processing (San Francisco, CA), March 23–36,
accuracy of thin slices. 1992, Vol. III, pp. 169 –72.
The overall spatial accuracy and stability in the field of [11] Eren PE, Sezan MI, Tekalp AM. Robust object based high resolution
view of MRI machines, in all three axes, is far better than image reconstruction from low resolution video. IEEE Trans on
their voxel size. This means that the imaged subject can be Image Processing 1998;6:1446 –51.
[12] Elad M, Feuer A. Super-resolution reconstruction of image se-
reliably and accurately shifted by predetermined sub-voxel
quences. IEEE Trans Patt Anal Machine Intell 1999;21:1646 –58.
steps between scans. Although for Fourier-encoded MRI [13] Peled S, Yeshurun Y. Superresolution in MRI: Application to human
data-sets we showed that super-resolution may only be white matter fiber tract visualization by diffusion tensor imaging.
usefully applied in the slice-select direction, a 3-D super- Magn Reson Med 2001;45(1):29 –35.