IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 12, NO.

1, JANUARY 2008 7

3-D Segmentation Algorithm of Small Lung

Nodules in Spiral CT Images
Stefano Diciotti, Member, IEEE, Giulia Picozzi, Massimo Falchini, Mario Mascalchi, Natale Villari,
and Guido Valli

Abstract—Computed tomography (CT) is the most sensitive CT scans repeated after short-term follow-up are usually em-
imaging technique for detecting lung nodules, and is now being ployed to investigate nodule changes over time. In fact, nodule
evaluated as a screening tool for lung cancer in several large sam- growth is the most important clue to possible malignancy of
ples studies all over the world. In this report, we describe a semiau-
tomatic method for 3-D segmentation of lung nodules in CT images small solitary pulmonary nodules (i.e., having a mean diameter
for subsequent volume assessment. The distinguishing features of ≤10 mm) identified in CT screening trials for lung cancer [8],
our algorithm are the following. 1) The user interaction process. It [9]. The current measurement method of nodule sizes, the dig-
allows the introduction of the knowledge of the expert in a simple ital caliper, is a completely manual tool and enables only the
and reproducible manner. 2) The adoption of the geodesic distance estimation of linear measurements. According to the Early Lung
in a multithreshold image representation. It allows the definition
of a fusion–segregation process based on both gray-level similarity Cancer Action Project (ELCAP) protocol [10], nodule mean di-
and objects shape. The algorithm was validated on low-dose CT ameter is the average of the maximum diameter and the largest
scans of small nodule phantoms (mean diameter 5.3–11 mm) and diameter perpendicular to the maximum diameter, both mea-
in vivo lung nodules (mean diameter 5–9.8 mm) detected in the sured on the CT image in which the nodule shows the maximum
Italung-CT screening program for lung cancer. A further test on cross-sectional area. Such linear measurements do not take into
small lung nodules of Lung Image Database Consortium (LIDC)
first data set was also performed. We observed a RMS error less account the actual 3-D shape of the nodule and possible asym-
than 6.6% in phantoms, and the correct outlining of the nodule metric pattern of growth. The estimate of nodule volume offers
contour was obtained in 82/95 lung nodules of Italung-CT and in substantial advantages: 1) the volume characterizes the size of
10/12 lung nodules of LIDC first data set. The achieved results sup- the structure considering the complete 3-D data, and thus, is the
port the use of the proposed algorithm for volume measurements most appropriate parameter to evaluate size changes; 2) volume
of lung nodules examined with low-dose CT scanning technique.
measurements have the potential to increase the sensitivity of
Index Terms—Computer-aided diagnosis, computer vision, lung CT in demonstrating lesion growth, since an increase of only
cancer, lung nodules segmentation, medical imaging, multiscale 26% of diameter corresponds to a doubling of the volume for
processing, spiral computed tomography (CT).
a spherical lung nodule; and 3) 3-D measurements show bet-
ter intraoperator and interoperator reproducibility with respect
I. INTRODUCTION to diameter estimations [11]–[13], due to the reduced operator
dependence of the measuring process.
UNG CANCER is the leading cause of cancer death both
L in Europe [1] and in the United States [2]. The five-year
survival rate is about 15%, and it has not significantly increased
Lung nodule volume is usually determined after a segmen-
tation procedure by using one of several methods: a simple
voxel counting algorithm [14], the application of the diver-
over the last 20 years. The same survival rate for subjects who gence theorem [15], or methods that compensate for partial vol-
have localized cancer at diagnosis is about 49% [2]. These fig- ume effects [14], [16]. In recent years, several algorithms have
ures suggest that early diagnosis of lung cancer can improve been investigated for lung nodule segmentation. Thresholding
the effectiveness of the treatment. In earlier stages, lung can- methods have been widely exploited. Several criteria, based on
cer most commonly manifests itself radiologically as a solitary in vitro and in vivo studies, have been adopted to choose a proper
noncalcified pulmonary nodule. Computed tomography (CT) is threshold value. Fixed [14], [16]–[19] and variable thresholding
superior to chest radiography for detecting pulmonary nodules, methods have been described [14], [19], [20], and selection of
and is now being evaluated as a screening tool in several large the threshold as function of the image gradient has also been
samples studies all over the world [3]–[7]. considered [19], [21]. Simple thresholding methods are suitable
for discriminating well-circumscribed lung nodules from pul-
monary parenchyma, while for lung nodules near other anatom-
Manuscript received October 12, 2006; revised January 25, 2007 and March
27, 2007. This work was supported by the Italian Ministry of Education, Uni- ical structures with similar gray levels, like vessels or pleura,
versity, and Research under Grant 2003068017. other strategies must be applied to avoid the inclusion of such
S. Diciotti and G. Valli are with the Department of Electronics and structures in the segmentation object. The analysis of the shape
Telecommunications, University of Florence, 50100 Florence, Italy (e-mail:
[email protected]; [email protected]). of pulmonary structures is usually performed to separate lung
G. Picozzi, M. Falchini, M. Mascalchi, and N. Villari are with nodules from surrounding structures, since for instance, lung
the Department of Clinical Physiopathology, University of Florence, nodules are typically ellipsoidal objects, while vessels exhibit a
50100 Florence, Italy (e-mail: [email protected]; [email protected];
[email protected]; [email protected]). preferential direction. Several 3-D shape parameters have been
Digital Object Identifier 10.1109/TITB.2007.899504 utilized including compactness or sphericity factors, and surface

1089-7771/$25.00 © 2008 IEEE

8 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 12, NO. 1, JANUARY 2008

curvature analysis [19], [21]–[23]. Approaches based on mor-

phological algorithms were described by several authors [16],
[17], [24]. Kostis et al. [17] adopted a morphological opening
with a fixed-size structuring element to separate small nodules
from the attached vasculature, and an iterative dilation proce-
dure to reconstruct nodule borders. Kuhnigk et al. [16] described
a method for processing both small and large nodules taking
into account that a more pronounced fusion between large nod-
ules and vessels is usually present. Aoyama et al. [25] applied a
nodule segmentation algorithm for the determination of the like-
lihood measure of malignancy of lung lesions. The procedure
uses a multiple thresholding method for the creation of a set of
gray level contour lines, and a dynamic programming technique
was used to obtain a nodule outline. Okada et al. [26] strategy
was based on a fit of the lung nodule with an ellipsoid with a
Gaussian profile. The nodule volume was then calculated on the
fitted model. An active contour method was also proposed [27].
It is based on three novel energy components that take advan-
tages of 3-D information, such as 3-D gradient, 3-D curvatures,
and a mask term that reduces the possibility of contour growth
over the pleural wall.
In this report, we describe a semiautomatic algorithm for
3-D lung nodule segmentation in spiral CT scans for subse-
quent volume evaluation. The computational strategy is based Fig. 1. Typical lung nodule for each class is shown. (a) Well-circumscribed
nodule. (b) Juxta-vascular nodule. (c) Nodule with a pleural tail. (d) Juxta-
on two main processing modules: a focus of attention stage and pleural nodule.
a 3-D region growing algorithm. Two key features emerge by
these stages. 1) Pulmonary structures included in the volume of
interest are automatically detected by the focus of attention and 2) Juxta-vascular nodule: there is a connection between the
are displayed to the operator for visual inspection. It is up to the nodule and a vessel.
operator to confirm each pulmonary structure signaled by the 3) With pleural tail: there is a thin connection between the
algorithm as a correct detection or to discard it. The knowledge nodule and the pleural wall. The pleural tail is part of the
of the expert is thus introduced into the algorithm through a nodule, and hence, has to be included in nodular segmen-
controlled user interaction. 2) The 3-D segmentation algorithm tation. Therefore, differently from juxta-vascular nodules,
adopts geodesic influence zones in a multithreshold image repre- the nonnodular structure must not be cut out to avoid un-
sentation to allow the achievement of fusion–segregation criteria derestimating the real nodule volume.
based on both gray-level similarity and objects shape. 4) Juxta-pleural nodule: a high portion of the surface of the
The method was validated on both synthetic phantoms and nodule is abutting the pleural wall. Juxta-vascular nodules
small lung nodules identified in subjects scanned as a part of are different from juxta-pleural ones due to a) the connec-
Italung-CT lung cancer screening program [7]. A further test tion between the nodule and the nonnodular structure and
on small lung nodules of Lung Image Database Consortium b) the geometrical shape of the surrounding nonnodular
(LIDC) first data set [28] was also performed. structure.
In CT scans, vessels can be represented by blob-like structures
with a local tubular shape. CT density of lung nodules and
II. METHODS vessels are quite similar and their recognition requires prior
knowledge of lung anatomy, including location, size, shape,
A. Lung Nodules Classes
and connections between the pulmonary structures.
In a 3-D CT examination, a lung nodule is a blob-like object In our method, we exploit the expert’s knowledge by a con-
lighter than the background, with a typical spheroidal/ellipsoidal trolled procedure as described in the following section. In this
shape. Lung nodules are embodied in a complex anatomical paper, nodules belonging to classes 3 and 4, which require
structure. In the lungs, several objects of various sizes are a dedicated algorithm for pleural surface detection, were not
present, such as bronchi and blood vessels, and sometimes their considered.
proximity to lung nodules can hinder nodule detection and ham-
per nodule segmentation. According to location and connection
with surrounding pulmonary structures, lung nodules can be B. Computational Strategy
classified into four main classes (see Fig. 1) [17]. The architecture of the algorithm is shown in Fig. 2. The
1) Well-circumscribed nodules: without any connection with operator selects the nodule of interest in the spiral CT images by
other pulmonary structures. using a custom computer user interface. As we are interested in
DICIOTTI et al.: 3-D SEGMENTATION ALGORITHM OF SMALL LUNG NODULES IN SPIRAL CT IMAGES 9

spherical nodule and an elongated surrounding structure—both

with a Gaussian profile and equal maximum value.

C. Focus of Attention
This stage detects and locates blob structures, identified as
candidate markers for the lung nodules segmentation algorithm.
A method for detecting blob patterns in a multiscale setting
using Laplacian of Gaussian (LoG) kernels was proposed by
Blostein and Ahuja [29]. They demonstrated that blob structures
correspond to local maxima of LoG filtered images. LoG scale
space was previously adopted in 2-D to detect lung nodules in
chest radiograms with excellent results

[30].
In a 3-D space [31], assuming r = x2 + y 2 + z 2 , the LoG
kernel can be written as
LoGσ (r) = w(σ)∇2 Gσ (r)
   
w(σ) r2 r2
=− 3 − 2 exp − 2 (1)
(2π)3/2 σ 5 σ 2σ
where σ is the standard deviation (SD) of the Gaussian and
w(σ) is a normalization factor depending on σ. With the choice
−w(σ) = σ 2 , the volume of the central lobe of the LoG kernel
is independent of σ and has a positive sign. In this way, it is
possible to compare responses with a√different σ. We shall refer
to the width of the central lobe h = 2 3σ as the scale parameter.
In accordance with the linear scale–space theory [32], let us
Fig. 2. Algorithm architecture for 3-D lung nodules segmentation. Bold rect-
angles indicate the stages in which user interaction is required.
consider the representation L(x, y, z, h) of the volumetric image
f (x, y, z) as
L(x, y, z, h) = f (x, y, z) ⊗ LoGh (x, y, z) (2)
assessing volumes of small nodules, a cubic volume of interest
(VOI) of 25 × 25 × 25 mm3 , centered in the voxel indicated for the scales h = hm in , . . . , hm ax .
by the operator, is extracted for further processing, and thus, For each scale, local maxima of L(x, y, z, h) are detected,
we can assume that the nodule is completely included in the and for each maxima, a coarse-to-fine tracking is applied and a
VOI. The VOI is supersampled with trilinear interpolation to scale–space signature [L(x, y, z, h) versus h] is plotted [33]. It
obtain an isotropic voxel and to reduce partial volume effects can be proven that the scale–space signature presents a marked
[16], [17], [20]. The focus of attention stage identifies candidate peak for a certain scale h∗ , suggesting the existence of a blob
markers corresponding to blob structures in the scale space. Each at that scale value h∗ . We shall refer to this scale as the optimal
candidate marker is characterized by the location, scale, and scale, or the scale at which the normalized LoG response is
contrast of the associated blob. The candidate marker, located maximal. The coordinates of the local maxima at the optimal
within 2 mm from the center of the VOI and with the highest scale are the estimated location of the blob structure.
contrast, is automatically considered as a nodular marker. It can be shown that the optimal scale h∗ provides an es-
All the candidate markers are shown to the observer for in- timate of blob size and that the value of LoG response at h∗
spection. The observer has to confirm each candidate marker estimates the blob contrast [32]. For the scale–space analysis,
either as a nodular marker, a marker of a surrounding pulmonary we considered a LoG family with h = {1, 2, . . . , 10} mm, and
structure, or as an object to discard. A 3-D region growing seg- for scale–space tracking, a search region of 1 × 1 × 1 voxels
mentation algorithm, described in detail in Section II-D, takes around each maxima.
such markers into account in order to avoid a possible fusion Applying the focus of attention algorithm to the computer
between the nodule and different pulmonary structures. At the simulation of Fig. 3(a), six local maxima in the LoG scale space
end of the segmentation stage, if the obtained segmentation is are detected. The candidate marker with the highest contrast
not successful, the procedure of marker supervision and the seg- within 2 mm from the VOI’s center is automatically selected as
mentation process can be repeated. The volume of the success- a nodular marker. One of the other candidate markers identifies
ful nodule segmentation is thus computed by using a reasonable the center of the elongated structure while the remaining four
and simple procedure, as the voxel counting method [14]. In can be related to border effects in LoG filtering. Each candi-
any case, other methods for volume computation can also be date marker and the automatically determined nodular marker
employed. are represented by circles in the slice corresponding to the z
To explain our method in the following, we shall refer to a marker coordinate [Fig. 3(b)] with the estimated location as the
3-D computer simulation depicted in Fig. 3 including a synthetic center and the optimal scale as radius. Circle contour style and
10 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 12, NO. 1, JANUARY 2008

Fig. 3. Three-dimensional computer simulation composed of a spherical lung nodule on the left attached to an elongated structure on the right. The two objects
have Gaussian profiles and the same maximum value. (a) Central slice of the 3-D image. (b) Results of the focus of attention stage represented by circles. The
nodular marker is automatically selected and is represented by a solid thick line. (c) Marker of a non-nodule structure selected by the user and depicted with a
solid thin line. Superimposed to the image (in black) the segmentation by thresholding with a value above the fusion value between the two structures in (d) and
below the fusion value in (e). (f) Segmentation result of our algorithm.

thickness depend on the typology of the marker: dotted lines Let Tk (I) = {p ∈ DI | I(p) ≥ k} be the threshold decom-
for candidate or ignored markers, solid thick lines for nodule position of the 3-D image I, defined in the domain of the image
markers, and solid thin lines for nonnodule markers. The LoG voxels DI . We consider the threshold values k lower or equal
value at the optimal scale, that is the contrast of the considered than the maximum Hounsfield (HU) value of the voxels se-
structure, is represented by the circle’s gray level, ranging from lected as markers, both nodular and nonnodular, and greater
black (low values) to white (high values). With this graphic than the HU value of the air (∼ −1000 HU). The threshold
representation, black circles correspond to low contrast struc- value initially assumes the maximum possible value of the in-
tures or image noise, while white circles correspond to more terval and decreases at each iteration by ∆k, with a consequent
pronounced objects. Candidate markers, are thus, shown to the growth of the number of voxels equal or greater than the selected
user for supervision. At this step, the user indicates, for each threshold. The step decrement ∆k must take a value not too high
candidate marker, whether it is representative of the nodule, of so that the growing of the Tk (I) occurs with smoothness, but
the surrounding pulmonary structures, or of an object to discard. not too small to become too affected by image noise. A good
The candidate markers selected by the user become markers for compromise was experimentally reached fixing ∆k = 20 HU.
the 3-D segmentation algorithm. In Fig. 3(c), a representation For each threshold value k, an opening morphological operation
of the markers is shown after the user’s supervision, in which is applied to Tk , with a spherical structuring element of radius
one nonnodule marker was signaled and four candidate mark- equal to 1 voxel, to separate regions connected by a reduced
ers have been discarded by the user, and thus, ignored by the number of voxels, generally produced by image noise. Let Tk
segmentation algorithm. be the image resulting from the morphological processing. For
a generic marker j, a region Rkj composed of the voxel marker
and of voxels of Tk that are six-neighborhood connected to the
D. Three-Dimensional Segmentation Algorithm of Lung considered marker is produced. Let Rknod be the union of the
Nodules regions connected with nodular markers and Rksur be the union
We accomplish both the lung nodule and surrounding pul- of the regions connected to nonnodular markers. For a generic
monary structures segmentation by using a region growing ap- threshold value, the regions Rknod and Rksur could be disconnected
proach. Region-based segmentation algorithms adopt fusion– or fused into one region. Generally, for high threshold values,
segregation criteria based on gray-level similarity and proximity the regions Rknod and Rksur are composed of a reduced number of
of the objects in the image. In our method, we join a gray-level voxels and are disconnected. By decreasing the threshold value
representation with the geodesic distance concept, as described k, the regions Rknod and Rksur growth and probability of fusion
in the following (see Fig. 4). between them increases. In Fig. 3, the images Tk +∆ k and Tk
DICIOTTI et al.: 3-D SEGMENTATION ALGORITHM OF SMALL LUNG NODULES IN SPIRAL CT IMAGES 11

Fig. 4. Schema of our 3-D segmentation algorithm of lung nodules.

were shown superimposed on the VOI, for a threshold value

above the fusion value between the two binary structures [at
threshold value k + ∆k in (d)] and below the fusion value [at
threshold value k in (e)]. In the first case, two separated regions
are detected, one for a nodular marker and the other for the
nonnodular structure, while in the second case, a unique con-
nected region with both markers is identified. In the first case,
Rknod constitutes the nodular segmentation for the current itera-
tion, similarly to that described in [21]. In the second case, let
Fig. 5. (a) Example of a geodesic distance between voxels x and y in a region
N = Tk − Tk +∆ k be the set of “new” voxels appeared during A is shown. (b) Geodesic influence zones are displayed for the regions B 1 and
the current growing step with respect to the previous one. At B 2 in A.
this point of the current iteration, the nodular region Rknod and
the nonnodular region Rksur must be redefined in order to result
disconnected.
For this purpose, we recall the concepts of geodesic distance
and geodesic influence zones. According to Vincent et al. [34],
the geodesic distance between two voxels x and y in set A
is defined as the shortest length of the paths joining x and y
and totally included in A [Fig. 5(a)]. Let us consider a set A
and m connected regions B1 , B2 , . . . , Bm totally included in
A [Fig. 5(b)]. The geodesic influence zone izA (Bi ) of a con-
nected component Bi in A is the locus of the points of A whose Fig. 6. Assignment of the voxels to the nodular or nonnodular regions accord-
geodesic distance to Bi is smaller than their geodesic distance ing to the geodesic influence zones.
to any other component of B. Keeping these definitions in mind,
we assign each voxel of N according to the geodesic influence tance allows one to keep the shape of the structures into account
zones izN (Rknod sur
+∆ k ) and izN (Rk +∆ k ) defined in the Tk region when calculating the distance between a voxel and a region, con-
with respect to the connected regions Rknod sur
+∆ k and Rk +∆ k (see sidering only the path inside a spatial region defined by voxels
nod
Fig. 6). If the voxel belongs to izN (Rk +∆ k ), it is assigned to with gray level equal to or greater than a proper threshold. The
Rknod , otherwise to Rksur . In other words, each voxel is associated selection criteria of the optimal nodular segmentation are based
to the nearest (according to the geodesic distance) connected on the assumption that the nodule boundary is located where
region (nodular or nonnodular) obtained at the previous itera- the mean magnitude of the image gradient calculated on the
tion, in which two disconnected regions were defined. In this nodule contour is maximal [21] (see Fig. 7). For this purpose,
way, a criterion of spatial proximity is adopted, when gray-level the magnitude of the image gradient is initially determined, and
information only is not adequate to decide if a voxel belongs to for each growing step: 1) the mean magnitude of the gradient
a nodular or to a nonnodular region. The use of the geodesic dis- on the contour of the nodular segmentation is calculated and 2)
12 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 12, NO. 1, JANUARY 2008

Fig. 7. (a) Two 3-D lung nodules segmentation are shown with the gradient vector displayed at the nodule boundary. (b) Plots of the mean magnitude of the
gradient on the nodule boundary versus the threshold value are reported.

Fig. 8. CT slice of a phantom of type A, B, and C are shown in (a)–(c), respectively. The images are displayed with standard lung window settings (window level
= −500 HU, window width = 1500 HU).

the optimal nodular segmentation is defined as the segmentation and about 120 s for juxta-vascular nodules. The focus of atten-
that has obtained the maximum mean magnitude gradient up to tion computation requires about 8 s, while the remaining time
that iteration. The maximal gradient is searched for in the set is needed for the 3-D segmentation algorithm.
of segmentations obtained at each iteration of the threshold- Albeit very small nodules (<5 mm diameter) can be depicted
ing procedure. However, two practical rules can be adopted to in low-dose CT, screening trials focus on small nodules. Hence,
terminate the procedure in advance, thus improving time per- we performed experimental tests on both small synthetic nodule
formances without any change in segmentation results. The first phantoms and small lung nodules identified in a lung cancer
rule is applied in the case in which a nodular region contains screening program. An application of the proposed algorithm to
voxels belonging to the border of the VOI. That segmentation lung nodules of LIDC first data set was also carried out.
can be discarded because it is not in accordance with the hy-
pothesis that the nodule has to be completely contained in the
A. Nodule Phantoms
VOI, and the growing procedure can terminate. The second rule
is related to the search of the maximal gradient. The plot of We made three types of synthetic nodule phantoms (Fig. 8)
the gradient versus the threshold value for a high number of to simulate in vivo lung nodules in different situations.
nodules was analyzed. It was observed that by decreasing the
threshold value, if the maximal gradient on the contour of the Type A: Forty silicone synthetic nodules (mean CT density
nodule segmentation is not increased for six consecutive grow- ∼115 HU) with irregular shapes and known volume simu-
ing steps, that gradient value could be considered as the absolute lating well-circumscribed nodules [Fig. 8(a)]. The volume
maximum and the algorithm can be stopped. of the nodules was within the range 75.9–688.6 mm3 (mean
volume = 268.8 mm3 ) corresponding to equivalent diameter
(the diameter of a sphere with the same volume) within 5.3–
III. EXPERIMENTAL TESTS 11 mm (mean diameter = 8.0 mm). Each synthetic nodule
The proposed procedures were implemented using C++ lan- was weighed on an analytical balance (Gibertini, 0.1-mg res-
guage in a Linux environment by using the Insight Segmentation olution) and the volume was computed dividing the weight
and Registration Toolkit (ITK) [35] library for the image pro- by the density (0.98 g/cm3 ) provided by the manufacturer of
cessing algorithms. The development and test platform was a PC the material.
equipped with an AMD Athlon XP 2400+ processor clocked at Type B: Ten silicone synthetic phantoms simulating nodules
2 GHz and with 512 MB of RAM memory. The processing time of known volume adjacent to a different structure, such as a
was, on the average, about 40 s for well-circumscribed nodules blood vessel [Fig. 8(b)]. Ten type A phantoms were utilized to
DICIOTTI et al.: 3-D SEGMENTATION ALGORITHM OF SMALL LUNG NODULES IN SPIRAL CT IMAGES 13

simulate nodules. Each phantom was located in contact with TABLE I

NODULE PHANTOMS AND ITALUNG-CT ACQUISITION PROTOCOLS
an elongated structure made up of the same material adopted
for type A phantoms.
Type C: Ten deformable well-circumscribed phantoms of
silicone rubber mixed with polycarbonate microspheres
[Fig. 8(c)]. The microspheres, commonly adopted in hobby
modeling as fillers, contain air and decrease the CT density of
the rubber at a mean value of ∼50 HU. The volume of these
phantoms was not known. The phantoms were scanned before
and after a deformation made by a hand manipulation. Due to
the low values of pressure applied during the shape deforma-
tion, the phantoms, which are solids, cannot be compressed
and preserve their volume unchanged. and the remaining 42 (having eight well-circumscribed and 27
juxta-vascular nodules) with the four rows of detector scanner.
All nodules were embedded in a marjoram (mean CT den- In vivo, lung nodules were acquired by using only low-dose
sity ∼ −865 HU) background to reproduce the texture of the protocols, as part of the Italung-CT program. The low-dose pro-
lung parenchyma in the CT scans. Types A and B phantoms tocols were denoted as low-dose single detector (LDSD) and
allowed the evaluation of the accuracy of the algorithm in vol- low-dose multidetector (LDMD), respectively, for the single
ume measurements, while type C enabled the investigation of detector and the multidetector CT scanner. HRCT acquisition
the behavior of the algorithm in the processing of nodules with protocol was employed with the single detector CT scanner.
the same volume but with a different shape. The acquisition parameters of each CT protocol were detailed
in Table I.
B. Lung Nodules Data Sets
D. Data Analysis
1) Italung-CT: We collected CT scans from subjects en-
rolled in the multicenter randomized clinical trial on lung can- An expert radiologist examined the segmentation produced
cer screening with low-dose thin-section CT, called Italung-CT, for each phantom and lung nodule on a computer monitor.
promoted and funded by the government of Tuscany Region To evaluate the accuracy of the algorithm in volume mea-
(Italy) [7]. Ninety subjects showed 122 small nodules of inde- surements, the RMS error was calculated in phantom tests of
terminate nature with mean diameter within 5 and 10 mm (6.4 ± types A and B. Since different CT collimation width has been
1.2 mm; range 5.0–9.8 mm). Ninety-eight nodules out of 122 used, volume measurements were affected by different partial
belong to classes 1 and 2, while the remaining 24 nodules were volume effects. To compensate for this phenomena, all phan-
of class 3 and 4, which were discarded. Three nodule exami- tom measurements were calibrated by using a linear function
nations out of 98 were excluded due to technical artifacts. The (voltrue = a × volmeas + b). For each CT acquisition protocol, a
proposed algorithm was then applied to the remaining 95 lung subset of 20 type A phantoms were randomly chosen to consti-
nodules. tute a calibration set. The calibration parameters a and b were
2) LIDC First Data Set: As further independent test set, we obtained with the least-square method, and the R2 was calcu-
adopted the data set collected by the LIDC [28]. The LIDC first lated for each scan protocol. The linear calibration function was
data set is composed of CT examinations with 23 nodules from applied in all phantoms and Italung-CT lung nodule tests. No
screening and diagnostic studies. All CT images have a matrix calibration data were available for LIDC first data set. For types
of 512 × 512 pixels and spacing between 0.605 and 0.742 mm. A and B phantoms, absolute volume error was considered be-
The cases are annotated by a panel of six expert radiologists [36]. tween the calibrated value and the true value [14]. For type A
We tested our algorithm on all solid nodules of LIDC first tests, only the phantoms not considered in the calibration pro-
data set belonging to class 1 or 2 having mean diameter less cess were taken into account for evaluation. The percentage of
than 10 mm, that is on four well-circumscribed and eight juxta- successful segmentations and the RMS errors were evaluated
vascular nodules. for each CT acquisition protocol.
To quantify the variability of volume measurements in phan-
toms of the same volume but of a different shape (type C phan-
C. Nodule Phantoms and Italung-CT Images Acquisition toms), the Bland and Altman statistical method with 95% lim-
All nodule phantoms were scanned both with a single row of its of agreement was applied in HRCT, LDSD, and LDMD
detector CT scanner (Somatom Plus 4; Siemens Medical Sys- cases. Preliminarily, for each CT protocol in type C phan-
tem, Erlangen, Germany) and with four rows of detector CT tom tests the correlation between the SD and the magnitude
scanner (Siemens Somatom Volume Zoom). All phantom im- of volume measurements was evaluated using the Kendall τ
ages were acquired with low-dose thin-section CT acquisition test with a p < 0.05. In the case of a significant correlation,
protocols in use in the Italung-CT and with a high resolution CT log-transformed data were employed [37], [38].
(HRCT) standard dose protocol. Forty-eight subjects (having To assess the amount of interaction required to mark
28 well-circumscribed and 32 juxta-vascular nodules) undergo- the structures proposed by the focus of attention, we recorded
ing Italung-CT were examined with the single row of detector the number of nodular and nonnodular markers adopted for the
14 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 12, NO. 1, JANUARY 2008

B. Lung Nodules
All nodules and vessels were automatically detected by the
focus of attention stage. The number of nodular and nonnodular
markers and the false positive segmentations are reported in
Table IV.
1) Italung-CT: In 82 out of 95 nodules (86.3%), our method
provided a successful segmentation. As detailed in Table IV, the
percentage of successful segmentations were 97.2% and 79.7%
for well-circumscribed and juxta-vascular nodules, respectively.
In Figs. 12 and 13, the segmentation of two juxta-vascular nod-
ules is shown Fig. 12 refers to a nodule examined with the
LDSD CT protocol, while Fig. 13 refers to a nodule examined
with the LDMD CT protocol. In each figure, the segmentation
without any nonnodular marker and with a nonnodular marker
was compared. Two slices of the 3-D VOI with superimposed
the markers determined by the focus of attention stage are dis-
played in panels (a) and (b) for the segmentation without any
Fig. 9. Regression line for type A phantoms in LDSD scans.
nonnodular marker, and in panels (c) and (d) for the segmen-
tation with a nonnodular marker. The marker in solid line and
thick stroke was automatically selected by the algorithm and
the marker in solid line and thin stroke was confirmed by the
segmentation of types A and B phantoms and lung nodules. The radiologist as nonnodular. Other markers were ignored. From
number of false positive segmentations (i.e., runs of the algo- panels (e) to (h) of both the figures, the segmentations of the
rithm that do not carry out a correct segmentation but, with a corresponding slices showed from panels (a) to (d) are depicted.
different marker choice, become successful) of types A and B In particular, in panel (c) of Figs. 12 and 13, the nonnodular
phantoms and lung nodules were also annotated. marker confirmed by the user allowed to avoid the fusion be-
tween the segmentation of the nodule and the vessel, clearly
visible in panel (e) of each figure.
2) LIDC First Data Set: A correct segmentation was ob-
IV. RESULTS tained in 10 out of 12 (83.3%) lung nodules. Exactly, all well-
A. Phantoms circumscribed nodules were successfully segmented together
with six out of eight juxta-vascular nodules (see Table IV). The
According to radiologist’s judgment for each CT protocol, two nodules not correctly outlined were the #04 and #09. In
all type A phantoms were successfully segmented. Phantoms of Fig. 14, the segmentation of the nodule #03 is shown.
type B were correctly segmented in six out of ten cases in the
LDSD cases and in ten out of ten cases both in LDMD and HRCT
acquisitions. The algorithm showed an RMS error in 1.0%–
6.6% for volume measurements of types A and B phantoms. V. DISCUSSION AND CONCLUSION
The calibration coefficients, the R2 values, and the RMS errors We have proposed and tested a semiautomatic method for
are reported in detail in Table II. In Fig. 9, the regression line for 3-D segmentation of lung nodules. Pulmonary structures in the
type A phantoms in LDSD acquisitions was shown. In Figs. 10 neighborhood of the nodule, such as vessels, are automatically
and 11, the volume percentage error versus equivalent diameter detected and confirmed by the operator to guide nodule seg-
is shown for LDSD, LDMD, and HRCT protocols in type A mentation. With only the automatic nodular marker selection,
tests. It should be noted that for any CT protocol, the volume almost all well-circumscribed and some juxta-vascular nodules
percentage error decreases as the size of the nodule increases. All were correctly segmented. Juxta-vascular nodules require, on
nodule phantoms and surrounding structures were automatically average, the confirmation of 0.9, 1.8, and 1.0 nonnodular mark-
detected by the focus of attention stage. The number of the ers for Italung-CT LDSD, Italung-CT LDMD category, and
adopted nodular and nonnodular markers for phantoms of types LIDC first data set, respectively.
A and B are reported in Table II. Mean and SD of false positive A region growing algorithm keeps such markers into account
segmentations are also indicated in Table II. by using geodesic distance zones in a multithreshold image
Table III details the results of the Kendall τ test for volume representation. We believe that the adoption of the geodesic dis-
measurements in type C phantoms. No significant correlation tance is a reasonable choice since voxels were assigned to the
between SD and the magnitude of the volume measurements nearest region according to the shapes of the structures. On av-
for any CT protocol was observed. The Bland and Altman 95% erage, the false positive segmentations were at maximum 0.9 in
limits of agreement in type C phantom tests were −5.0% to the Italung-CT LDMD category for juxta-vascular nodules. Due
1.5% in HRCT acquisitions, −7.5% to 12.2% for LDSD scans, to the geodesic approach applied after the confirmation of nonn-
−7.4% to 5.6% in the LDMD case. odular markers, computational time is larger, but still limited,
DICIOTTI et al.: 3-D SEGMENTATION ALGORITHM OF SMALL LUNG NODULES IN SPIRAL CT IMAGES 15

Fig. 10. Volume percentage error versus true volume for type A phantoms. (a) LDSD acquisitions. (b) LDMD acquisitions.

TABLE II
CALIBRATION COEFFICIENTS, R 2 , AND RMS ERRORS FOR EACH CT PROTOCOL

TABLE III
KENDALL τ CORRELATION COEFFICIENTS BETWEEN SD AND MAGNITUDE OF
THE VOLUME MEASUREMENTS

Fig. 11. Volume percentage error versus true volume for type A phantoms in
HRCT acquisitions.

for juxta-vascular nodules as compared to well-circumscribed fixed and variable thresholding. Partial volume methods were
nodules. also proposed. The authors reported high errors (1.2–10.0 mm3
Since an user interaction is required, the effect of this man- for phantoms of size 8–60 mm3 ) as compared to other studies,
ual intervention should be verified on the performance of the probably for the employment of large FOV and a realistic lung
algorithm. For this reason, in another study, described in [39], background. Recently, Kuhnigk et al. [16] reported a median
both intraoperator and interoperator reproducibility of volume error of −3.1% on juxta-vascular phantoms in standard and
measurements were assessed on low-dose thin-section CT scans. low-dose CT scans.
Ideally, a comparison between segmentation algorithms on
It is well known that volume measurements are heavily af- lung nodules should be performed on a large public data set
fected by CT image quality. As reported by Ko et al. [14], the with a defined ground truth [16]. In the future, important re-
best results are obtained with standard dose and sharp recon- sources still under construction, like the LIDC data set, will
struction kernel protocols. Moreover, collimation width plays an probably become a reference standard, but nowadays such
important role since it has a relevant influence on partial volume archives are at an initial stage. As a matter of fact, most
effects. Some previous studies on nodule phantoms were per- studies were performed on systems developed with private
formed by other researchers. Yankelevitz et al. [40] obtained an databases and, since segmentation performance is heavily af-
RMS error of 3% on phantoms with size ≥3 mm examined with fected by CT protocols and nodule characteristics, the compari-
a standard dose and small field of view (FOV). Ko et al. [14] son between different research groups remains, currently, quite
applied several algorithms for nodule segmentation based on limited.
16 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 12, NO. 1, JANUARY 2008

TABLE IV
RESULTS OBTAINED FOR LUNG NODULES IDENTIFIED IN THE ITALUNG-CT AND FOR LUNG NODULES OF LIDC FIRST DATA SET

Fig. 12. Segmentation without and with the confirmation of a nonnodular marker of a juxta-vascular lung nodule of Italung-CT trial acquired with the LDSD
CT protocol: two slices of the 3-D VOI with superimposed the markers are displayed in panels (a) and (b) for the segmentation without any nonnodular markers,
and in panels (c) and (d) for the segmentation with a nonnodular marker; on the bottom [panels from (e) to (h)], the corresponding segmentations are displayed.

Fig. 13. Segmentation without and with the confirmation of a nonnodular marker of a juxta-vascular lung nodule of Italung-CT trial acquired with the LDMD
CT protocol. See caption of Fig. 12 for panels explanation.
DICIOTTI et al.: 3-D SEGMENTATION ALGORITHM OF SMALL LUNG NODULES IN SPIRAL CT IMAGES 17

Fig. 14. Segmentation without and with the confirmation of a nonnodular marker of nodule #03 of LIDC first data set. See caption of Fig. 12 for panels
explanation.

Kostis et al. [17] reported 80% of successful segmentations on protocols, good performances were obtained since the RMS
a data set of 21 juxta-vascular nodules examined with standard error was lower than 6%. In type B cases, the segmentation
dose acquisitions. With a free manual correction of the kernel approach, based on the geodesic influence zones, permitted the
size of a morphological operation, such a percentage increased detachment of nodules from surrounding pulmonary structures:
to 95%. Kuhnigk et al. [16] achieved a high percentage of cor- the segmentations were all successful except in four phantoms
rect segmentation (91.4%) on a large data set of low-dose scans in LDSD scans. With LDSD protocol, partial volume effects
including small and large nodules. Lastly, Okada et al. [26] in- were more pronounced compared to LDMD and HRCT proto-
dicated about 81% of successful segmentations on two data sets cols because of the wider collimation width; therefore, a seg-
including juxta-pleural and ground-glass opacities, all examined mentation failure can more likely occur in the case of a nod-
with HRCT. ule near other surrounding structures. It should be pointed out
In this paper, both a standard dose HRCT protocol (which that a few studies have addressed the role of low-dose tech-
ensures better image quality) and low-dose CT protocols, as nique with respect to the accuracy (and reproducibility) of the
used in lung cancer screening programs, were employed for nodule volume measurements [14], [16], [41], [42]. Like other
in vitro validation. To study phantoms in similar conditions to researchers [14], we observed that volume measurements were
those employed with in vivo lung nodules, we paid attention to significantly influenced by the scan dose, at least in in vitro
the realistic model of synthetic phantoms and also on the use studies in which a gold standard is available. Type C phantoms
of a large FOV during image acquisitions. Nodule phantoms enabled us to explore the influence of the shape of the nodule
with irregular shape were made, both well circumscribed and in in volumetric analysis. Bland and Altman’s analysis revealed
proximity to other structures. As regarding CT image density that volume measurements performed in phantoms examined
and texture, they were embodied in a material that simulates with HRCT have reduced variability compared to that employed
the lung parenchyma. All these factors are likely to affect the on low-dose scans. For example, if a nodule measurement of
segmentation performances, and therefore, the volume estima- 300 mm3 was performed, by varying only the shape of the nod-
tion [14]. Moreover, we used large FOV, similar to that used ule we should obtain, in another measurement process with a
in patient examinations, with the drawback of a reduced im- confidence level of 95%, a measurement between 277.8 and
age quality. As a fact, CT images acquired with a large FOV 316.8 mm3 .
show a worse in-plane spatial resolution and increased partial In regard to lung nodules identified in the Italung-CT,
volume effects compared to those generated with a targeted we observed successful segmentations in 97.2% of well-
(small) FOV [17], [40]. Hence, the use of a targeted FOV can circumscribed nodules and in 79.7% of juxta-vascular nodules.
produce more accurate volume estimations [14], but it is not Overall, correct segmentations were obtained in 86.3% on
recommended in current protocols for lung cancer screening the entire data set. No statistical comparison among the
[10]. segmentation performances of nodules examined with different
For type A phantoms, in HRCT images, the RMS error was CT scanners was attempted. It will be performed as soon
lower than in low-dose protocols due to both higher SNR and re- as results from extended samples are available. A further
duced partial volume effects. However, with LDSD and LDMD test on the LIDC first data set revealed a 83.3% of correct
18 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 12, NO. 1, JANUARY 2008

segmentations (100% for well-circumscribed nodules and 75% [16] J. Kuhnigk, V. Dicken, L. Bornemann, A. Bakai, D. Wormanns, S. Krass,
for juxta-vascular nodules). and H. Peitgen, “Morphological segmentation and partial volume analysis
for volumetry of solid pulmonary lesions in thoracic CT scans,” IEEE
In conclusion, the results reached by the proposed method Trans. Med. Imag., vol. 25, no. 4, pp. 417–434, Apr. 2006.
on phantoms and on in vivo nodules support the validity of [17] W. Kostis, A. Reeves, D. Yankelevitz, and C. Henschke, “Three-
the algorithm to analyze well-circumscribed and juxta-vascular dimensional segmentation and growth-rate estimation of small pulmonary
nodules in helical CT images,” IEEE Trans. Med. Imag., vol. 22, no. 10,
nodules identified in a screening trial with low-dose thin-section pp. 1259–1274, Oct. 2003.
CT. [18] J. Goo, T. Tongdee, R. Tongdee, K. Yeo, C. Hildebolt, and K. Bae, “Vol-
umetric measurement of synthetic lung nodules with multi-detector row
CT: Effect of various image reconstruction parameters and segmentation
ACKNOWLEDGMENT thresholds on measurement accuracy,” Radiology, vol. 235, pp. 850–856,
The authors would like to thank the anonymous reviewers 2005.
[19] W. Mullally, M. Betke, and J. Wang, “Segmentation of nodules on chest
for their helpful comments about the preliminary version of this computed tomography for growth assessment,” Med. Phys., vol. 31, no. 4,
manuscript. pp. 839–848, 2004.
[20] A. Reeves, A. Chan, D. Yankelevitz, C. Henschke, B. Kressler, and
W. Kostis, “On measuring the change in size of pulmonary nod-
REFERENCES ules,” IEEE Trans. Med. Imag., vol. 25, no. 4, pp. 435–450, Apr.
2006.
[1] International Agency for Research on Cancer. (2005). [Online]. Available: [21] B. Zhao, A. Reeves, D. Yankelevitz, and C. Henschke, “Three-dimensional
http://www-dep.iarc.fr. multicriterion automatic segmentation of pulmonary nodules of helical
[2] American Cancer Society. (2005). Cancer fact and figures [Online]. computed tomography images,” Opt. Eng., vol. 38, no. 8, pp. 1340–1347,
Available: http://www.cancer.org/downloads/STT/CAFF2005f4 PWSe- 1999.
cured.pdf. [22] Y. Kawata, N. Niki, H. Oshmatsu, K. Eguchi, and N. Moriyama, “Shape
[3] S. Diederich, M. Thomas, M. Semik, H. Lenzen, N. Roos, A. Weber, analysis of pulmonary nodules based on thin section CT images,” in Proc.
W. Heindel, and D. Wormanns, “Screening for early lung cancer with low- SPIE, 1997, vol. 3034, pp. 964–974.
dose spiral computed tomography: Results of annual follow-up examina- [23] J. Kostis, A. Reeves, D. Yankelevitz, and C. Henschke, “Three-
tions in asymptomatic smokers,” Eur. Radiol., vol. 14, no. 4, pp. 691–702, dimensional curvature analysis of small pulmonary nodules in helical
Apr. 2004. CT scans,” in Proc. RSNA Scientific Program 2000, Suppl. Radiol., Nov.,
[4] S. Sone, L. Li, Z. G. Yang, T. Honda, Y. Maruyama, S. Takashima, vol. 217P, p. 549.
M. Hasegawa, S. Kawakami, K. Kubo, M. Haniuda, and T. Yamanda, [24] C. Fetita, F. Preteaux, C. Beigelman-Aubry, and P. Grenier, “3D automatic
“Results of three-year mass screening programme for lung cancer using lung nodule segmentation in HRCT,” in Proc. MICCAI, (Series Lecture
mobile low-dose spiral computed tomography scanner,” Brit. J. Cancer, Notes in Computer Science 2878) R. E. Ellis and T. M. Peters, Eds.
vol. 84, pp. 25–32, 2001. Berlin, Germany: Springer-Verlag, pp. 626–634, 2003.
[5] C. I. Henschke, D. F. Yankelevitz, J. P. Smith, O. M. Miettinen, ELCAP [25] M. Aojama, Q. Li, S. Katsuragawa, F. Li, S. Sone, and K. Doi, “Comput-
Group, “Screening for lung cancer: The early lung cancer action ap- erized scheme for determination of the likelihood measure of malignancy
proach,” Lung Cancer, vol. 35, pp. 143–148, 2002. for pulmonary nodules on low-dose CT images,” Med. Phys., vol. 30,
[6] S. J. Swensen, J. R. Jett, J. A. Sloan, D. E. Midthun, T. E. Hartman, no. 3, pp. 387–394, Mar. 2003.
A. M. Sykes, G. L. Aughenbaugh, F. E. Zink, S. L. Hillman, G. R. Noetzel, [26] K. Okada, D. Comaniciue, and A. Krishnan, “Robust anisotropic gaussian
R. S. Marks, A. C. Clayton, and P. C. Pairolero, “Screening for lung cancer fitting for volumetric characterization of pulmonary nodules in multi-
with low-dose spiral computed tomography,” Amer. J. Respir. Crit. Care slice CT,” IEEE Trans. Med. Imag., vol. 24, no. 3, pp. 409–423, Mar.
Med., vol. 165, pp. 508–513, 2002. 2005.
[7] G. Picozzi, E. Paci, A. L. Pegna, M. Bartolucci, G. Roselli, A. D. Francisci, [27] T. Way, L. Hadjilski, B. Sahiner, H. Chan, P. Cascade, E. Kazerooni,
S. Gabbrielli, A. Masi, F. Perri, M. L. Rottoli, F. Carozzi, N. Villari, and N. Bogot, and C. Zhou, “Computer-aided diagnosis of pulmonary nodules
M. Mascalchi, “Screening of lung cancer with low dose spiral CT. Results on CT scans: Segmentation and classification using 3D active contours,”
of a three year pilot study and design of the randomized clinical trial Med. Phys., vol. 33, no. 7, pp. 2323–2337, 2006.
“Italung-CT”,” La Radiol. Med., vol. 109, pp. 17–26, Jan./Feb. 2005. [28] K. W. Clark, D. S. Gierada, S. M. Moore, D. R. Maffitt, P. Koppel, S.
[8] D. M. Libby, J. P. Smith, N. K. Altorki, M. W. Pasmantier, D. Yankelevitz, R. Phillips, and F. Prior, “Creation of a CT image library for the lung
and C. I. Henschke, “Managing the small pulmonary nodule discovered screening study of the national lung screening trial,” J. Digit Imag.,
by CT,” Chest, vol. 125, pp. 1522–1529, 2004. vol. 20, no. 1, pp. 23–31, Mar. 2007.
[9] D. Yankelevitz, R. Gupta, B. Zhao, and C. Henschke, “Small pulmonary [29] D. Blostein and N. Ahuja, “Shape from texture: Integrating texture-
nodules: Evaluation with repeat CT-preliminary experience,” Radiology, element extraction and surface estimation,” IEEE Trans. Pattern Anal.
vol. 212, pp. 561–566, 1999. Mach. Intell., vol. 11, no. 12, pp. 1233–1251, Dec. 1989.
[10] C. Henschke. (2006, Feb.) International Early Lung Cancer Action Pro- [30] G. Coppini, S. Diciotti, M. Falchini, N. Villari, and G. Valli, “Neural
gram: Protocol [Online]. http://www.ielcap.org/ielcap.pdf, accessed May networks for computer aided diagnosis: Detection of lung nodules in
31, 2006. chest radiograms,” IEEE Trans. Inf. Technol. Biomed., vol. 7, no. 4,
[11] M. Revel, C. Lefort, A. Bissery, M. Bienvenu, L. Aycard, G. Chatellier, pp. 344–357, Dec. 2003.
and G. Frija, “Pulmonary nodules: Preliminary experience with three- [31] S. Diciotti, L. Petrolo, M. Falchini, G. Picozzi, M. Mascalchi, N. Villari,
dimensional evaluation,” Radiology, vol. 231, pp. 459–466, 2004. and G. Valli, “Multi-scale neural network system for lung nodules detec-
[12] L. Goodman, M. Gulsun, L. Washington, P. Nagy, and K. Piacsek, “In- tion in spiral CT images: Preliminary results,” presented at the Mediterr.
herent variability of CT lung nodule measurements in vivo using semi- Conf. Med. Biol. Eng., IFMBE, Ischia, Italy, 2004.
automated volumetric measuements,” Amer. J. Roentgenol., vol. 186, [32] T. Lindeberg, Scale-Space Theory in Computer Vision. Dordrecht, The
pp. 989–994, 2006. Netherlands: Kluwer, 1994.
[13] K. Marten, F. Auer, S. Schmidt, G. Kohl, E. Rummeny, and K. Engelke, [33] T. Netsch and H. Peitgen, “Scale-space signatures for the detection of
“Inadequacy of manual measurements compared to automated CT vol- clustered microcalcifications in digital mammograms,” IEEE Trans. Med.
umetry in assessment of treatment response of pulmonary metastases using Imag., vol. 18, no. 9, pp. 774–786, Sep. 1999.
RECIST criteria,” Eur. Radiol., vol. 16, pp. 781–790, 2006. [34] L. Vincent and P. Soille, “Watershed in digital spaces: An efficient algo-
[14] J. P. Ko, H. Rusinek, E. L. Jacobs, J. S. Babb, M. Betke, G. McGuinness, rithm based on immersion simulations,” IEEE Trans. Pattern Anal. Mach.
and D. P. Naidich, “Small pulmonary nodules: Volume measurement at Intell., vol. 13, no. 6, pp. 583–598, Jun. 1991.
chest CT-phantom study,” Radiology, vol. 228, no. 8, pp. 864–780, [35] L. Ibanez, W. Schroeder, L. Ng, and J. Cates, The ITK Software Guide.
2003. New York: Kitware, 2003.
[15] A. M. Alyassin, J. L. Lancaster, J. H. Downs, III, and P. T. Fox, “Evaluation [36] National Cancer Institute. LIDC: Datasets as a public resource. (2005).
of the new algorithms for the interactive measurement of surface area and [Online]. Available: http://imaging.cancer.gov/reportsandpublications/ re-
volume,” Med. Phys., vol. 21, no. 6, pp. 741–752, Jun. 1994. portsandpresentations/firstdataset.
DICIOTTI et al.: 3-D SEGMENTATION ALGORITHM OF SMALL LUNG NODULES IN SPIRAL CT IMAGES 19

[37] J. M. Bland and D. G. Altman, “Statistical methods for assessing agree- Mario Mascalchi was born in Florence, Italy, in
ment between two methods of clinical measurement,” Lancet, vol. 1, 1959. He received the Laurea degree (with honors)
pp. 307–310, 1986. in medicine from the Medical School of the Univer-
[38] J. Bland and D. Altman, “Measuring agreement in method comparison sity of Florence, Florence, in 1984, two Postgraduate
studies,” Stat. Meth. Med. Res., vol. 8, pp. 135–160, 1999. degrees in neurology and radiology from the Univer-
[39] G. Picozzi, S. Diciotti, M. Falchini, S. Foresti, F. Gallesi, E. Cavigli, sity of Florence, Florence, Italy, in 1988 and 1996,
L. Livi, N. Villari, and M. Mascalchi, “Operator-dependent reproducibility respectively, and the Ph.D. degree in applied phys-
of size measurements of small phantoms and lung nodules examined with iopathology from the University of Florence, in 1992.
low dose thin-section CT,” Investig. Radiol., vol. 41, no. 11, pp. 831–839, In 1992, he became an Assistant Professor of the
Nov. 2006. Department of Radiology, University of Pisa, Pisa,
[40] D. Yankelevitz, A. Reeves, W. Kostis, B. Zhao, and C. Henschke, “Small Italy. In 1998, he became a Reseacher at the Univer-
pulmonary nodules: Volumetrically determined growth rates based on CT sity of Florence, where, in 2001, he became an Associate Processor of radiology.
evaluation,” Radiology, vol. 217, pp. 251–256, 2000. His current research interests include computed tomography of the lung and neu-
[41] D. Wormanns, G. Kohl, E. Klotz, A. Marheine, F. Beyer, W. Heindel, and romagnetic resonance. He is the author of 150 papers in refereed journals and
S. Diederich, “Volumetric measurements of pulmonary nodules at multi- his impact factor (IF) is 320.
row detector CT: In vivo reproducibility,” Eur. Radiol., vol. 14, pp. 86–92,
2004.
[42] J. Ko, R. Marcus, E. Bomsztyk, J. Babb, C. Stefanescu, M. Kaur,
D. Naidich, and H. Rusinek, “Effect of blood vessels on measurement
of nodule volume in a chest phantom,” Radiology, vol. 239, pp. 79–85,
2006.
Natale Villari was born in Messina, Italy, in 1940. He
received the Laurea degree (with honors) in medicine
Stefano Diciotti (S’03–M’04) was born in Florence, from the Medical School of Messina University,
Italy, in 1975. He received the Laurea degree (with Messina, in 1964. He specialized in radiology in 1966
honors) in electronic engineering from the University and nuclear medicine in 1970, all from the University
of Florence, Florence, in 2001, and the Ph.D. degree of Florence, Florence, Italy, and in oncology in 1975
in bioengineering from the University of Bologna, from Modena University, Modena, Italy.
Bologna, Italy, in 2005. In 1970, he became an Assistant of the Department
Since 2005, he has been a Contract Professor of of Radiology, University of Florence. In 1976, he was
biomedical technologies, University of Florence. His a Professor of radiology with the National University
current research interests include medical imaging, of Mogadiscio, Mogadiscio, Somalia, where in 1981,
soft computing, and CAD systems. he became an Associate Processor of radiology. Since 1987, he has been a
Dr. Diciotti is a member of the IEEE Engineering Professor of general and special odontostomatologic radiology and the Director
in Medicine and Biology Society. of Radiology Unit at the University of Florence, where, since 1992, he has also
been a Professor of radiology, and, in 1996, the Director of the Radiology Unit
at the Department of Clinical Physiophatology. His current research interests
Giulia Picozzi was born in Prato, Italy, in 1973. include magnetic resonance and computed tomography. He is the author of
She received the Laurea degree in medicine and the more than 200 publications and the coauthor of many books covering subjects
Postgraduate degree in radiology from the Univer- as general radiology and new imaging techniques.
sity of Florence, Florence, Italy, in 1999 and 2003, Prof. Villari is a member of the Italian Society of Medical Radiology (SIRM),
respectively. and from 1996 to1998, he was the Vice President.
Since 2006, she has been a Researcher at the De-
partment of Clinical Phisiopathology, University of
Florence, where she is involved in research programs
for lung cancer screening with computed tomogra-
phy (CT), and is a Radiologist with the Italung-CT
Screening Group.
Guido Valli received the Laurea degree in physics
from the University of Padua, Padua, Italy, in 1963.
Since 1982, he has been a Researcher with the
Massimo Falchini was born in Florence, Italy, in National Council of Researches (CNR). In 1982, he
1959. He received the Laurea degree in medicine and became an Associate Professor of biomedical data
two Postgraduate degrees in gastroenterology and in and signal processing at the Faculty of Engineer-
radiology from the University of Florence, Florence, ing, University of Florence, Florence, Italy, where
in 1985, 1989, and 1989, respectively. since 2000, he has been a full Professor of biomedi-
Since 1995, he has been a Researcher at the cal technology. His current research interests include
Department of Clinical Phisiopatology, University the study of innovative methods for the processing
of Florence. His current research interests include of biomedical data and signals, and the investigation
biomedical imaging and interventional radiology. of computer vision systems for medical diagnosis. He is the author of a number
of publications and reports in the related fields.