264 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 61, NO.

2, FEBRUARY 2014

Amorphous Regions-of-Interest Projection Method

for Simplified Longitudinal Comparison of Dynamic
Regions in Cancer Imaging
Yaniv Gal∗ , Nicholas Dowson, Pierrick Bourgeat, Olivier Salvado, Paul Thomas, Michael Fay, Stephen Rose,
Rosalind L. Jeffery, Amir Fazlollahi, and Stuart Crozier

Abstract—Tumors are typically analyzed as a single unit, despite obtained and that tumors are biologically heterogeneous, the
their biologically heterogeneous nature. This limits correlations treatment outcomes are generally measured in terms of pa-
that can be drawn between regional variation and treatment out- tient survival and whole-tumor growth [1]–[3]. Such analysis
come. Furthermore, despite the availability of high resolution 3-D
medical imaging techniques, local outcomes, (e.g., tumor growth), has limitations in revealing the local factors involved in the
are not easily measured. This paper proposes a method that uses tumor progression. Similarly, histological and genetic studies
streamlines to divide a 3-D region of interest (e.g., tumor) into units of cancer use biopsy samples often taken from a few (often
where local properties can be measured over the paths of growth. one) locations [4] with the assumption that the whole tumor is
The parameters such as directional length and mean intensity can homogeneous. Regional analysis, however, may provide better
be measured locally at sequential time points and then compared.
The method is evaluated on synthetic objects, simulated tumors, data for establishing prognostics and for treatment planning and
and medical images of brain tumors. The evaluations suggest that monitoring.
the method is suitable for mapping amorphous dynamic objects. Although resection allows detailed characterization of tu-
Index Terms—Brain cancer, image analysis, image processing, mors, methods for analyzing residual tumor in vivo are needed.
medical imaging, region of interest (ROI). Imaging can characterize factors such as metabolism and blood
brain barrier breakdown at every location within a tumor vol-
ume. The analysis of such data is challenging due to the over-
I. INTRODUCTION whelming amount of information in the 3-D images and the
difficulty in associating corresponding regions from different
HREE-dimensional medical imaging is widely used for
T the diagnosis and treatment planning for cancer. Despite
the fact that high-resolution 3-D images of tumors can be
images.
The outcome can be measured in different ways, including
changes in metabolism or changes in shape and volume. Such
measurements consider changes in local tissue properties as a
function of time. Hence, there is a requirement for a one-to-one
Manuscript received December 12, 2012; revised July 23, 2013; May 30, mapping from a location at one time point to its new posi-
2013; accepted September 12, 2013. Date of publication September 17, 2013; tion at another time point. Moreover, as tumor shape usually
date of current version January 16, 2014. This work was supported by the Aus- changes in time, the comparison of the images of the tumor ac-
tralian National Health and Medical Research Council funding scheme under
Project Grant 631567. First two authors contributed equally to this work. Aster- quired at different time points (e.g., before and after the course
isk indicates corresponding author. of radiotherapy) is often limited in showing in which regions
∗ Y. Gal is with the School of Information Technology and Electrical
of the tumor growth or shrinkage has occurred. Spatial map-
Engineering, University of Queensland, Brisbane, Qld. 4072, Australia (e-mail:
ygal@itee.uq.edu.au). pings can be obtained using registration, either rigid [5], [6] or
N. Dowson, P. Bourgeat, O. Salvado, S. Rose, and A. Fazlollahi are with the nonrigid [7], [8]. However, registration methods suffer from an
Australian E-Health Research Center, Royal Brisbane and Women’s Hospital, implicit ambiguity as to whether a change is due to local changes
Herston, Qld. 4006, Australia (e-mail: Nicholas.Dowson@csiro.au; Pierrick.
Bourgeat@csiro.au; olivier.salvado@csiro.au; stephen.rose@cai.uq.edu.au; in intensity or changes in the object morphology. Sensible con-
Amir.Fazlollahi@csiro.au). straints on registration algorithms [9] can be made that limit
P. Thomas is with the Queensland Nuclear Medicine, Royal Bris- the ambiguity, but a second problem arises, namely the division
bane and Women’s Hospital, Herston, Qld. 4006, Australia (e-mail: Paul_
Thomas@health.qld.gov.au). of the tumor into units for which outcome can be meaning-
M. Fay is with the Department of Radiation Oncology, Royal Brisbane and fully evaluated. In simulations, meaningful units such as cells
Women’s Hospital, Herston, Qld. 4006, Australia (e-mail: mikefay@me.com). or contiguous biological environments can be selected [10], but
R. L. Jeffery is with the Department of Neurosurgery, Royal Brisbane and
Women’s Hospital, Herston, Qld. 4006, Australia (e-mail: r.jeffree@uq.edu.au). they may not always be inferred from images without bias. Fur-
S. Crozier is with the School of Information Technology and Electrical thermore, registration algorithms are intrinsically challenged to
Engineering, University of Queensland, Brisbane, Qld. 4072, Australia (e-mail: give results that are biologically feasible. This problem is the
stuart@itee.uq.edu.au).
This paper has supplementary downloadable material available at main application of interest in this paper: finding spatial corre-
http://ieeexplre.ieee.org (file size: 554 kB). spondences between regions within a tumor in images taken at
Color versions of one or more of the figures in this paper are available online different stages of the tumor development (e.g., before and after
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TBME.2013.2282402 radiotherapy).

0018-9294 © 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
GAL et al.: AMORPHOUS REGIONS-OF-INTEREST PROJECTION METHOD FOR SIMPLIFIED LONGITUDINAL COMPARISON 265

The previous work has looked at localizing tumors using A. Defining the ROI Surface
expectation maximization and using normal brain atlases as
An image can be described as a mapping of D-dimensional
a prior [11], using nonparametric generative approaches [12],
spatial coordinates into intensities. Given an image, I(x) :
and Bayesian models [13], with some success. Likewise, the
RD → R, with quantitative intensities, I(xi ), and an ROI,
methods to detect subtle changes from images have been pro-
R ⊆ RD , to be processed, it is assumed that R is a single con-
posed [14]. However, all the aforementioned methods consider nected component. It is preferred (but not essential) that the
tumors as a whole. Nonrigid image registration fields have never,
surface of R is homeomorphic to a sphere (i.e., an orientable
to our knowledge, been successfully used to locally link biolog-
surface of genus zero). This makes the projection map easier to
ical factors with local outcomes in longitudinal studies. interpret in terms of correspondence between the map and the
This paper proposes a new method for achieving local anal-
original ROI topology. For practical reasons, we assume that I
ysis of outcomes in tumors. The method is based on projecting and R are defined on a discrete set of voxel coordinates, i.e.,
(2-D or 3-D) a region of interest (ROI) from an internal refer- xi ∈ ZD . In such case, the surface of R is defined to be the set of
ence (core) point to the exterior surface, using streamlines as an
voxels R̄ = {xi | xi ∈/ R ∧ {N (xi ) ∩ R} = ∅}, where N (x) is
indicator for tumor local direction of growth. This allows the the set of (26) immediate neighbors of the voxel x. In the case
placement of tumors into a standard coordinate system despite of multiple connected components or ROIs from the same tu-
their amorphous shape. Further, with the modest requirements
mor (e.g., metastases), it is possible to either apply the method
of rigid alignment and moderate movement, spatial correspon- to each individual ROI, if independent analysis is required or
dences over time are obtained allowing local outcomes to be to define an encapsulating (bounding) ROI that contains all the
referred to. The method is agnostic to dimensionality and im-
other ROIs. In such case, the voxels in the encapsulating ROI
age resolution as demonstrated by its application to synthetic that are outside of all the original ROIs will be ignored during
images. The correspondence of the streamlines to the local tu-
the projection process.
mor growth was tested on simulated tumor images using the
simulation scheme proposed in [15].

B. Obtaining a Projection Reference Point

II. METHOD Defining a reference (i.e., core) point is the first step of the
The method, proposed in this paper, called amorphous ROI ARP method. It is required to be inside the ROI, regardless of
projection (ARP), is designed to project a 3-D object, described its morphological properties (e.g., convexity) in order for the
by an ROI, to a 2-D projection map using a set of streamlines to method to be able to capture internal properties of the ROI un-
allow effective, yet efficient, local analysis of tissue properties. der inspection (e.g., intensity mean and variance in different
The projection of a 3-D object onto a 2-D map imposes some regions of the ROI). Also, the core points of similar ROIs (e.g.,
loss of information. However, the underlying assumption of this a follow-up study of a tumor) are required to be in close prox-
paper is that the ROI analysis is not performed on an individual imity in order to allow analysis of change. A morphological
voxel basis, and thus, the information from sets of voxels can skeletonization algorithm was adapted for this purpose. For the
be augmented into a single value with no impact on the analysis 2-D ROI case, an algorithm from [16] was adapted, while for
results. Furthermore, the geometrical correspondence between a 3-D ROI, the authors adapted the skeletonization algorithm
the projection map and the original image is preserved to allow from [17]. Once a skeleton, K ⊂ R,  has been formed, the ROI
x i ∈R
tracking the source voxels in the projection map. center of mass is calculated, m = |R | , where |R| is the
In this paper, tumors are represented as ROIs in the image number of elements in R. The voxel ROI core point is then
where a reference (core) point is defined. Finally, a set of stream- chosen to be xc = arg minx i ∈K xi − m . In some special, yet
lines is constructed from the ROI surface (i.e., boundary) to the rare, cases (e.g., when the ROI shape is a torus), the core point
core point and from the surface to a bounding sphere that is then may become unstable under small changes in the ROI topology.
projected to an easy-to-interpret 2-D map. None of the images that were tested for this paper exhibited
A desirable feature of such projections is that local changes in such ROI topology. When comparing two or more ROIs, the
the original ROI (e.g., local growth of tumor over time) will yield core points may not correspond due to such sensitivity. This
local changes in the projection map. This means that changes can usually be solved by selecting a reference ROI, calculating
in the ROI morphology or intensity levels in a specific region its core point xc and using (the same) xc for all other ROIs
will be translated to changes in a corresponding region in the under comparison. In case xc falls outside the ROI, the closest
projection map and thus makes it possible to locate the region point inside the ROI can be selected as a core. This solution,
in the original ROI that produced these changes. however, assumes that all ROIs are spatially registered which is
The method is designed for longitudinal studies, where ROIs sometimes hard to achieve. Thus, when spatial alignment can-
from different time points are compared (e.g., tumor post treat- not be achieved, the aforementioned method should be used for
ment and a follow-up scan for detecting regrowth). However, the each ROI independently. Further, stability of the core location
proposed method is also useful for within time point analysis is obtained by iteratively computing the core point, while down-
as the spatial correspondence allows interpretation of measure- weighting the influence of points distant from the center of mass
ments, despite the object’s amorphous shape. (i.e., outliers).
266 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 61, NO. 2, FEBRUARY 2014

measured as an indication of local changes in metabolic activity.

The aforementioned properties are calculated and stored for
each surface voxel, resulting in a pair of values for each voxel
xj ∈ R̄. These measures were chosen to demonstrate the utility
of the approach. Other useful measures can be extracted from
the streamlines based on the application. The total variance,
for instance, can provide information on irregular changes in
intensity which is a meaningful feature in the breast cancer
analysis [20], [21].
The characterization of the region’s properties along each
streamline is essentially a summarization of the object onto a
Fig. 1. Synthetic object formed from a uniform sphere with irregularities in
boundary and intensity level. (a) shows the original object outline (dark grey)
manifold with one less dimension. Hence, the scalar properties
and the region of high uptake (white) and (b) shows the object streamlines, R̄, of a 3-D region can be projected onto a 2-D scalar field (and
to the core, xc , and from R̄ to the sphere S. similarly a 2-D object onto a 1-D profile), which is not only
convenient for visualization, but is useful for certain applica-
C. Extracting ROI Streamlines tions. In particular, when reporting on tumors after resection,
individual regions of the residual tumor within the periphery of
In order to project the ROI onto its surface, Laplace’s equation
the resection cavity may progress at different rates.
is utilized in a similar way to [18] and [19]. Laplace’s equation
is a second-order partial differential equation defined on a scalar
field, f , and is given by D. Projecting the ROI Onto a 2-D Map
∇2 f (x) = 0. (1) An additional advantage of the ARP method is its ability to
project the internal content of the ROI on a 2-D rectangular map
As boundary conditions, we set f (xc ) = 0 and f (xi ∈ R̄) =
that can be easily interpreted by humans. The projection process
1. We use LR to denote the solution of the Laplace equation
is based on the constructed streamlines and is performed by
inside R, i.e., LR : (R\xc ) → R. Next, the gradient field, ∇LR ,
projecting the outer surface onto a sphere and then transforming
is calculated and a streamline, ψj , is extracted for each voxel,
the sphere into a 2-D map (plane).
xj ∈ R̄, by following the gradient directions from xj to xc .
1) Projecting the ROI Surface Onto a Sphere: Once the de-
Each streamline is treated as a parameterized curve ψj (t) : R →
sired properties have been measured inside the ROI, based on
RD , such that ψj (0) = xj and ψj (1) = xc . An example of the
the internal streamlines, the calculated values are stored in cor-
streamlines created in this manner, using the synthetic data,
respondence to the boundary voxels which are the starting points
is shown in Fig. 1(b), in blue. The properties of the Laplace
of the streamlines. The ROI surface topology is then simplified
solution ensure that the streamlines never cross as they converge
by projecting it onto a sphere, S. S is chosen to be a sphere
on the reference point xc . Also, all the voxels in R are intersected
with a center at xc and a radius large enough to entirely en-
by at least one streamline, because no local minima apart from
compass R̄, i.e., R ⊂ B, where B is the ball bounded by S.
xc exist within LR .
Although the ARP method is relatively insensitive to the choice
The reference point is chosen in a manner that is insensitive
of radius, too large a radius may “diffuse” local changes to
to changes in region shape, and the Laplacian method ensures
surrounding streamlines thus diminishing the visibility of the
that the streamlines are primarily influenced by the position of
change. It is hence recommended that S be as small as possible
the proximal region boundary. Hence, the variation of stream-
under the above constraints. The transformation between R̄ and
lines with small changes in region shape is low. This limited
S is performed by solving the Laplace equation where S and R̄
variation allows correspondences between streamlines from a
are the initial conditions (i.e., f (xj ∈ R̄) = 1, f (xk ∈ S) = 0).
region defining a tumor at two points in time to be inferred.
Once a solution is established, gradient-based stream lines,
The defined streamlines, ψj , allow multiple useful properties
ηj are formed for each xj ∈ R̄, such that ηj (0) = xj and
to be measured, including:
ηj (1) = sj ∈ S. For each point sj ∈ S, we can now assign the
1) The total length of a streamline
corresponding values PL j and PI j . The points {sj } are then
 1
   transformed into a pair (one each for growth and change in in-
PL j = PL (ψj ) = ψj (t) dt. (2)
0
tensity) of 2-D projection maps, ML and MI , holding the values
{PL j } and {PI j } of the corresponding voxels sj on the sphere.
2) The mean intensity along the streamline
 1 An example of the streamlines created in this manner, using the
1 synthetic data is shown in Fig. 1(b), in red. A simplified ver-
PI j = PI (ψj ) = I (ψj (t)) dt. (3)
L(ψj ) 0 sion of the proposed method can skip the sphere construction
stage by projecting the measured values PL j and PI j directly to
The length of the streamline is measured in real world units the corresponding maps ML and MI , based on the correspond-
and when measured at two time points and subtracted is an ing streamline direction of entry toward the core point. This
approximate measure of growth. Similarly, the regional mean method, however, has decreased sensitivity to small changes in
intensity in brain positron emission tomography (PET) images is R̄ as streamlines originating at small curvatures on the surface
GAL et al.: AMORPHOUS REGIONS-OF-INTEREST PROJECTION METHOD FOR SIMPLIFIED LONGITUDINAL COMPARISON 267

(i.e., “outwards” tumor growth) tend to gather toward the core

point and thus represent the change on the surface in a very small
angular region near the core. This property can cause such small
changes to be averaged away when calculating ML and MI .
2) Transforming the Sphere S Into a 2-D Map: The sphere
S can be transformed into a 2-D rectangular map using a Mer-
cator projection [22] (i.e., projecting the sphere onto a bounding
cylinder, aligned with the z-axis, and then unfolding the cylin-
der). For consistency, azimuth 0 is defined as the direction of
y − (i.e., inferior) and is increasing clockwise. This definition
is especially useful for simplifying the analysis of the maps
in brain imaging, for instance, as the center of the map corre-
sponds to the direction of the patient’s nose (i.e., anterior). Once
a correspondence between voxels on the sphere and pixels in
the map has been established, each pair of values PL j , PI j is
assigned to the row and column in ML and MI that correspond
to the azimuth and slice of sj , respectively. The set sj ⊆ S will
usually not occupy all the voxels in S. Hence, the maps ML
and MI may also not be fully occupied and will contain gaps
of undefined values. The gaps in the maps can be covered by
(linear or nearest-neighbor) interpolation of the existing values.
It is important to note that such interpolation must be cyclic on
the azimuth (column) axis of the map (i.e., the first column is
subsequent to the last).

Fig. 2. Results of two 2-D objects (ROIs) with irregularity in shape and
III. EXPERIMENTS AND RESULTS intensity. Top row: original ROIs, Middle row: corresponding M L , Bottom
row: corresponding M I . In the graphs, the vertical axis represents the measured
A. Synthetic Data quantity while the horizontal axis represents azimuth (in degrees).

In the first experiment, the ARP method was qualitatively

applied to the synthetic data in order to test its ability to identify B. Simulated Growth
small inhomogeneities in shape and intensity. The method was The goal of the second experiment was to test the ARP method
first applied to selected slices in a 3-D image and then to the on small changes in the tumor shape with as realistic data as
image as a whole. possible. The Konukoglu [15] tumor growth simulation was
For an ROI with a shape of a perfect sphere and uniform used for this purpose. The Konukoglu simulation is based on
intensity, ML and MI are, respectively, expected to be uniform. the reaction–diffusion method [23] and accounts for directional
When irregularities in the boundary or intensity occur, a cor- diffusion based on tract direction. The Konukoglu method was
responding change is observed in the projections of ML (for chosen because it accounts for the diffusion properties of white
shape) or MI (for intensity). Fig. 2 demonstrates this for two matter and grey matter in the brain, providing a good approxi-
2-D regions with irregularities in boundary and intensity. In this mation to the actual brain tumor growth process. Nine different
experiment, each object was constructed from a (discrete) circle tumors were simulated in different locations in the brain. The
and a narrow rectangle attached to it. A high intensity location extent of each tumor was recorded and analyzed with the pro-
(lesion) was embedded inside the spherical part of the region. posed model in several time points during the growth of the
In order to examine the results of the model with 3-D data, tumor (approximately every five days, based on the simulation
various synthetic 3-D objects were constructed, including: 1) time scale). The extent of sample simulated tumors in the differ-
a sphere with uniform intensity, 2) a sphere with a “hot spot” ent time points and the resulting ML maps in the corresponding
(i.e., small region with high intensity), 3) a sphere with a hot time points are presented in Fig. 4, note that all the simulations
spot and a narrow cuboid attached to it, pointing toward y − , and and resulting maps demonstrated similar behavior. A core loca-
4) a sphere with a hot spot and a narrow cuboid attached to it, tion sensitivity test has also been carried out using the simulated
pointing toward z + . data. The results showed that movement of the core within a
The calculated ML and MI for item 3 is presented in Fig. 3. range of 5 mm can cause insignificant changes in the difference
The results demonstrate how an irregularity in either intensity between the ML maps (less than 5% of maximum intensity).
levels or boundary shape can be easily identified in the projected
maps by looking for local changes. The changes can then be
C. Analyzing the Biological Meaning of the Model Streamlines
tracked back to the location of the corresponding streamline by
constructing a hash table that relates each pixel in the maps to The simulations were also used to test the hypothesis that
the corresponding voxel in R̄ during the projection process. the model streamlines ψj provide a reasonable prediction of
268 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 61, NO. 2, FEBRUARY 2014

Fig. 5. Mean results of the streamline quasi-distance. The horizontal axis

represents simulation days prior to the final tumor state while the vertical axis
represents mean distance between the model streamlines and simulated line of
Fig. 3. M L and M I resulting from a 3-D sphere with an irregularity in shape growth. The dashed red lines represent the maximum and minimum values, the
(horizontal cuboid) and intensity. blue dashed lines represent one standard deviation and the green line represents
the mean over all simulations.

distance between the streamlines ψi and γi for a selected time

T is given by
 
ds = max min ( γi (t1 ) − ψi (t2 ) ) . (4)
t 1 ∈[0,T ] t 2 ∈[0,1]

The directional correlation at location xi is given by the inner

product
  + 
ψj (0 ), γi (0+ )
Ci = . (5)
ψj (0+ ) · γi (0+ )

The quasi-metric (4) measures the spatial difference between

the line of growth and the model corresponding streamline. In-
tuitively, this asymmetric variation of the Hausdorff distance
indicates how far the model streamline can get from the real
line of growth in the defined past time period. The streamline
quasi-distance was averaged over all streamlines of each simu-
lation and the results are presented in Fig. 5. The results show
Fig. 4. Simulated tumor growth, using the Konukoglu method. Top row: the that the model streamlines follow the actual line of growth, with
extent of the tumor in different time points; ten days after tumor seeding, and only a few millimeters maximum deviance for a period of al-
every five days afterwards. Bottom row: the corresponding M L maps (in mm).
most a (simulation) month prior to tumor imaging. Thus, the
results indicate that the model streamlines correspond well with
short-term local direction of growth. In order to test this hypoth- the actual tissue line of growth for the past few weeks. The di-
esis, two properties were measured: First, the mean streamline rectional correlation on the surface of the tumor tests the ability
quasi-distance (defined below) between the model streamlines of the model streamlines to predict the direction of growth on
and the simulation line of growth for a given time, and; second, the surface in the short term. The directional correlation was
the directional correlation between each model streamline and measured and averaged over all streamlines in each simulated
the actual direction of growth (extracted from the simulation), tumor. The results of the comparison for one of the simulated
on the surface of the tumor. tumors are given in Fig. 6. The mean directional correlation over
Let γi be the “line of growth” of the tumor from the tumor all tumors was 0.98, with minimum correlation of 0.95, maxi-
seed xs (i.e., where it started from) to the voxel xi , such that mum 0.99, and median value of 0.99, providing evidence that
γi (0) = xi and γi (1) = xs . Given a streamline ψi between the the direction of the streamlines ψj (t) (t ≈ 0) closely concurs
ROI core xc and the ROI surface voxels xi ∈ R̄, the quasi- with the direction of simulated growth.
GAL et al.: AMORPHOUS REGIONS-OF-INTEREST PROJECTION METHOD FOR SIMPLIFIED LONGITUDINAL COMPARISON 269

Fig. 6. Directional correlation between model streamlines ψ j and the actual

direction of the simulated growth, using the Konukoglu method. The blue lines
represent the model streamlines while the red segments represent the direction
of growth near the boundary. The mean directional correlation values between
the model streamlines and the tumor growth for the above tumor was 0.99.

D. Clinical Data
To qualitatively evaluate the utility of the method on clinical
data, longitudinal brain and breast tumor images were acquired.
Datasets from three brain cancer patients (patient 1, 2, and 3)
and one breast cancer patient (patient 4) were used. Fig. 7. Patient 1 brain tumor images: FDOPA-PET in “hot iron” on top,
1) Brain Tumor Data: All brain patients (aged 58, 71, CEMR on bottom, with manual delineation overlaid (green line): (a,c) three
and 70, respectively) were diagnosed with a (Glioblastoma) months postresection, (b,d) six months postresection.
high grade primary brain tumor. 4-dihydroxy-6-[1 8F]-fluoro-
L-phenylalanine (FDOPA) PET and contrast enhanced MR 2) Breast Cancer Data: The breast cancer data (patient 4)
(CEMR) images were acquired at two time points: four weeks were acquired in two time points: initial image (t1), at the age
after resection of the tumor (t1), and four months after resection of 42 and a follow-up scan (t2) three years later. Each dataset
(t2). The two time points were, respectively, after a tumor resec- contains a T1 /T2 anatomical and a dynamic contrast-enhanced
tion and after a course of (temozolamide) chemo-radiotherapy. magnetic resonance imaging. The initial scan detected a sus-
In each of the studies, the tumor was delineated manually picious lesion in the right breast. The cytological examination
by an experienced nuclear-medicine physician (PT) and the un- confirmed that the suspicious lesion was benign. The follow-
modified delineations were used to define the ROI. The sample up scan showed changes in the lesion shape. The cytological
slices from the images with the manual delineation overlaid are examination confirmed the lesion to be a malignant carcinoma.
presented in Figs. 7 and 9. Note the significant change in size and In each of the images, the lesion was delineated manually
intensity between the first (t1) and second image (t2) of patient by an experienced clinician and the unmodified delineations
1. The ML and MI generated from patient 1 ROI images are were used to define the ROI. The sample slices from the images
presented in Fig. 8. The core location points were selected in the with the manual delineation overlaid are presented in Fig. 13.
t1 image using the method described at Section II-B and were The ML generated for the two time points are presented in
reused in the t2 images after rigid registration (i.e., translation Fig. 14. Due to deformation of the organ under examination,
and orientation). The ML graphs indicate that the tumor shrank no unambiguous spatial correspondence between the images
due to therapy. However, the MI graphs show an increase of could be established. Thus, the location of the core point was
tracer uptake levels within the tumor, which suggest that treat- estimated using the method in Section II-B for each time point
ment did not eliminate all the active tumor and that recurrence independently. Also, as breast MR images are acquired in a
is likely. The resulting MI and ML from a 3-D ROI projection prone position (i.e., face down), it is assumed that the organ
of patient 2 data, for both time points, are presented in Figs. 10 orientation is approximately consistent between scans and no
and 11, respectively. In these images, ML was measured in further correction is required.
millimeters, MI of the PET was measured in standardized up- The resulting ML maps indicate that the tumor has generally
take value ratio [24] which is a method for interpatient PET shrunk but the shape of the tumor has changed (see Fig. 15).
uptake normalization. MI of the CEMR was measured in abso-
lute contrast enhancement between the precontrast anatomical
and postcontrast, taken 4 min after the injection of the contrast IV. DISCUSSION
agent. The region of largest change in ML between t1 and t2 is A new method was proposed to describe dynamic amorphous
marked by a black arrow in Fig. 11, with a corresponding arrow ROIs. The method allows spatial correspondences between the
in the 3-D diagram of the tumor in Fig. 12. objects to be obtained despite structural changes, by defining a
270 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 61, NO. 2, FEBRUARY 2014

Fig. 8. Intensity maps M I of a 2-D ROI from patient 1: Left: PET, center: CEMR, and; Right: length maps M L . Green: Three months postresection (t1). Blue:
Six months postresection (t2).

Fig. 11. 3-D distance maps M L of t1 (Left) and t2 (Right) time points of
patient 2. The black arrows point to a region with indicated shrinkage of the
tumor.

Fig. 12. 3-D model of patient 2 tumor: left image shows the tumor at t1, while
the right image shows the tumor at t2. The colors represent the M L values at the
corresponding position. The arrow indicates the location of a significant change
Fig. 9. Patient 2 brain tumor images: FDOPA-PET in “hot iron”on top, CEMR
in the M L value between t1 and t2 and corresponds to the arrow location in
on bottom, with manual delineation overlaid (green line): (a,c) Three months
Fig. 11.
postresection. (b,d) Six months postresection.

Fig. 10. 3-D intensity maps M I of t1 (Left) and t2 (Right) time points of
patient 2. The black arrows points to an area of indicated increase in uptake Fig. 13. Patient 4 breast CEMR images with manual delineation overlaid
(potential tumor regrowth) between t1 and t2. (green line): (a) Initial image. (b) Three years follow-up.
GAL et al.: AMORPHOUS REGIONS-OF-INTEREST PROJECTION METHOD FOR SIMPLIFIED LONGITUDINAL COMPARISON 271

The selection of the core point for the model is a crucial step
in producing a meaningful representation of changes in the ob-
ject under inspection. The task of selecting the core point in
longitudinal studies should be carefully examined when spa-
tial correspondence is not achievable. Also, the orientation of
the object under inspection should be approximately consistent
between scans.

V. CONCLUSION
This paper has proposed a simple method to subdivide tumors
Fig. 14. 3-D distance maps M L of t1 (Left) and t2 (Right) time points of
patient 4. Note the overall shrinkage of the tumor. using streamlines. The streamlines are shown to be biologically
meaningful models of paths of previous growth with predictable
error. Furthermore, the streamlines demonstrated to be useful
for longitudinal analysis of dynamic amorphous 3-D shapes
such as tumors and can be used to measure multiple useful
parameters such as metabolism or geometry. The experiments
demonstrated that the method can be an effective tool for the
analysis of amorphous tumors in 3-D medical images.

ACKNOWLEDGMENT
Fig. 15. 3-D model of patient 4 breast tumor: left image shows the tumor at The authors would like to thank D. McClymont for providing
t1, while the right image shows the tumor at t2. The colors represent the M L the breast images.
values at the corresponding position. Note the overall shrinkage of the tumor
and the change in shape on the left end (arrow).
REFERENCES
[1] W. Chen. (2007, Sep.). Clinical applications of PET in brain tumors. J. Nu-
reference point, a boundary, and a set of streamlines between clear Med.: Offic. Publication, Soc. Nucl. Med. [Online]. 48(9), pp. 1468–
the boundary and reference point. 1481, PMID: 17704239. Available:http://www.ncbi.nlm.nih.gov/pubmed/
17704239
The method also allows the description of complex 3-D ob- [2] B. J. Fueger, J. Czernin, T. Cloughesy, D. H. Silverman, C. L. Geist, M.
jects via a 2-D projection of the data where each streamline is A. Walter, C. Schiepers, P. Nghiemphu, A. Lai, M. E. Phelps, and W.
presented by a single point. This property is useful for examin- Chen. (2010, Oct.). Correlation of 6-18F-fluoro-L-dopa PET uptake with
proliferation and tumor grade in newly diagnosed and recurrent gliomas.
ing biologically relevant measures such as tissue heterogeneity J. Nuclear Med.: Offic. Publ., Soc. Nucl. Med. [Online]. 51(10), pp. 1532–
and regional mean uptake in PET and MR images. 1538, PMID: 20847166. Available:http://www.ncbi.nlm.nih.gov/pubmed/
The method was evaluated on the synthetic data, images of 20847166
[3] C. Schiepers, W. Chen, T. Cloughesy, M. Dahlbom, and S. Huang.
simulated tumor growth, and on brain tumor PET and MR im- (2007, Oct.,). 18f-fdopa kinetics in brain tumors. J. Nucl. Med. [Online].
ages from clinical practice. The results on the synthetic data 48(10), pp. 1651–1661. Available:http://jnm.snmjournals.org/cgi/content/
suggest that there is intuitive correspondence between the pro- abstract/48/10/1651
[4] M. Gerlinger, A. J. Rowan, S. Horswell, J. Larkin, D. Endesfelder, E.
jection maps and the original images for simple shapes. The Gronroos, P. Martinez, N. Matthews, A. Stewart, P. Tarpey, I. Varela, B.
tumor growth simulation provides evidence that such correspon- Phillimore, S. Begum, N. Q. McDonald, A. Butler, D. Jones, K. Raine,
dence also exists when the shape is amorphous. The analysis of C. Latimer, C. R. Santos, M. Nohadani, A. C. Eklund, B. Spencer-Dene,
G. Clark, L. Pickering, G. Stamp, M. Gore, Z. Szallasi, J. Downward,
directional correlation between the model streamlines and the P. A. Futreal, and C. Swanton. (2012, Mar.). Intratumor heterogeneity
simulated tumor suggests that the model streamlines provide and branched evolution revealed by multiregion sequencing. New Engl. J.
good indication of expected direction local tumor growth. A Med. [Online]. 366(10), pp. 883–892, PMID: 22397650. Available:http:
//www.ncbi.nlm.nih.gov/pubmed/22397650
possible explanation for that is that tumor cells have higher [5] J. West, J. M. Fitzpatrick, M. Y. Wang, B. M. Dawant, J. Maurer, C R,
probability to grow toward tissue with low cell density which R. M. Kessler, R. J. Maciunas, C. Barillot, D. Lemoine, A. Collignon,
sits with the tendency of the streamlines to remain “as far” as F. Maes, P. Suetens, D. Vandermeulen, P. A. van den Elsen, S. Napel, T.
S. Sumanaweera, B. Harkness, P. F. Hemler, D. L. Hill, D. J. Hawkes, C.
possible from each other. Studholme, J. B. Maintz, M. A. Viergever, G. Malandain, and R. P. Woods.
The evaluation on clinical data shows that the method is (1997, Aug.). Comparison and evaluation of retrospective intermodality
stable enough to deal with large and complex 3-D volumes brain image registration techniques. J. Comput. Assist. Tomogr. [Online].
21(4), pp. 554–566, PMID: 9216759. Available:http://www.ncbi.nlm.nih.
that change significantly over time. Furthermore, the projection gov/pubmed/9216759
maps provide simplified visualization of the data, which reveal [6] M. Jenkinson, P. Bannister, M. Brady, and S. Smith, “Improved optimiza-
subtle changes in shape and intensity in the ROI over time. tion for the robust and accurate linear registration and motion correction
of brain images,” NeuroImage, vol. 17, no. 2, pp. 825–841, Oct. 2002.
The model can be generated in a relatively short-time (usu- [7] B. B. Avants, C. L. Epstein, M. Grossman, and J. C. Gee, “Symmetric
ally less than a minute) using a single-core workstation, for an diffeomorphic image registration with cross-correlation: evaluating au-
average size ROI. Furthermore, parallelization of the method on tomated labeling of elderly and neurodegenerative brain,” Med. Image
Anal., vol. 12, no. 1, pp. 26–41, Feb. 2008.
a multicore CPU or a general-purpose graphical processing unit [8] D. Rueckert, C. Hayes, C. Studholme, P. Summers, M. Leach, and D.
is relatively straight-forward. J. Hawkes. (1998). “Non-rigid registration of breast MR images using
272 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 61, NO. 2, FEBRUARY 2014

mutual information,” in Medical Image Computing and Computer- [16] L. Lam, S.-W. Lee, and C. Y. Suen, “Thinning methodologies-a compre-
Assisted Interventation? MICCAI98, vol. 1496, W. M. Wells, A. Colch- hensive survey,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 14, no. 9,
ester, and S. Delp, Eds. Berlin/Heidelberg, Germany: Springer-Verlag, pp. 869–885, Sep. 1992.
1998, pp. 1144–1152. [Online]. Available:http://www.springerlink.com/ [17] Y. Ge, D. Stelts, and D. Vining, “3D skeleton for virtual colonoscopy,”
content/f272v611l25g27q7/ in Visualization in Biomedical Computing.. New York, NY, USA:
[9] J. P. Pluim, J. B. Maintz, and M. A. Viergever, “Mutual-information-based Springer-Verlag, 1996, pp. 449–454.
registration of medical images: A survey,” IEEE Trans. Med. Imag., [18] S. Jones, B. Buchbinder, and I. Aharon, “Three-dimensional mapping
vol. 22, no. 8, pp. 986–1004, Aug. 2003. of cortical thickness using Laplace’s equation,” Human Brain Mapping,
[10] Z. Wang and T. S. Deisboeck, “Computational modeling of brain tumors: vol. 11, no. 1, pp. 12–32, 2000.
discrete, continuum or hybrid?,” in Scientific Modeling and Simulations, [19] O. Acosta, P. Bourgeat, M. Zuluaga, J. Fripp, O. Salvado, and S. Ourselin,
S. Yip, T. D. Rubia, T. J. Barth, M. Griebel, D. E. Keyes, R. M. Nieminen, “Automated voxel-based 3d cortical thickness measurement in a combined
D. Roose, and T. Schlick, Eds., (Lecture Notes in Computational Science Lagrangian–Eulerian PDE approach using partial volume maps,” Med.
and Engineering Series). vol. 68Amsterdam, The Netherlands: Springer- Image Anal., vol. 13, no. 5, pp. 730–743, 2009.
Verlag, 2009, pp. 381–393. [20] American College of Radiology, “Breast imaging reporting and data sys-
[11] A. Gooya, K. M. Pohl, M. Bilello, G. Biros, and C. Davatzikos. (2011). tem (BIRADS),” American College of Radiology (ACR), 2006. Available:
“Joint segmentation and deformable registration of brain scans guided www.arc.org
by a tumor growth model,” Med. Image Comput. Comput.-Assist. In- [21] Y. Gal, A. Mehnert, A. Bradley, D. Kennedy, and C. Stuart, “New spa-
tervent.: MICCAI . . . Int. Conf. Med. Image Comput. Comput.-Assist. tiotemporal features for improved discrimination of benign and malignant
Intervent., vol. 14, no. Pt 2, pp. 532–540, PMID: 21995070 PM- lesions in dynamic contrast-enhanced-magnetic resonance imaging of the
CID: PMC3246749, [Online]. Available:http://www.ncbi.nlm.nih.gov/ breast,” J. Comput. Assist. Tomography, vol. 35, no. 5, pp. 645–654, 2010.
pmc/articles/PMC3246749/ [22] O. M. Miller, “Notes on cylindrical world map projections,” Geograph.
[12] M. R. Sabuncu, B. T. T. Yeo, K. Van Leemput, B. Fischl, and P. Gol- Rev., vol. 32, no. 3, pp. 424–430, 1942.
land. (2010, Oct.). A generative model for image segmentation based on [23] O. Clatz, M. Sermesant, P. Bondiau, H. Delingette, S. Warfield,
label fusion. IEEE Trans. Med. Imag. [Online]. 29(10), pp. 1714–1729. G. Malandain, and N. Ayache, “Realistic simulation of the 3-D growth of
Available:http://dspace.mit.edu/handle/1721.1/64791 brain tumors in MR images coupling diffusion with biomechanical defor-
[13] J. Corso, E. Sharon, S. Dube, S. El-Saden, U. Sinha, and A. Yuille, “Effi- mation,” IEEE Trans. Med. Imag., vol. 24, no. 10, pp. 1334–1346, Oct.
cient multilevel brain tumor segmentation with integrated Bayesian model 2005.
classification,” IEEE Trans. Med. Imag., vol. 27, no. 5, pp. 629–640, May [24] B. Lopresti, W. Klunk, C. Mathis, J. Hoge, S. Ziolko, X. Lu, C. Meltzer,
2008. K. Schimmel, N. Tsopelas, S. DeKosky, and J. Price, “Simplified quan-
[14] K. Pohl, E. Konukoglu, S. Novellas, N. Ayache, A. Fedorov, I. Talos, tification of Pittsburgh compound b amyloid imaging pet studies: A com-
A. Golby, W. Wells, R. Kikinis, and P. Black, “A new metric for detect- parative analysis,” J. Nucl. Med., vol. 46, no. 12, pp. 1959–1972, 2005.
ing change in slowly evolving brain tumors: Validation in meningioma
patients,” Neurosurgery, vol. 68, pp. 225–233, 2011.
[15] E. Konukoglu, O. Clatz, B. Menze, B. Stieltjes, M. Weber, E. Mandonnet,
H. Delingette, and N. Ayache, “Image guided personalization of reaction-
diffusion type tumor growth models using modified anisotropic Eikonal
equations,” IEEE Trans. Med. Imag., vol. 29, no. 1, pp. 77–95, Jan. 2010. Authors’ photographs and biographies not available at the time of publication.