3D Variational Brain Tumor Segmentation using a High Dimensional Feature Set
Dana Cobzas, Neil Birkbeck
Computer Science
University of Alberta
Mark Schmidt
Computer Science
Univ. of British Columbia
Martin Jagersand
Computer Science
University of Alberta
Albert Murtha
Department of Oncology
University of Alberta
Abstract
Tumor segmentation from MRI data is an important but
time consuming task performed manually by medical experts. Automating this process is challenging due to the
high diversity in appearance of tumor tissue, among different patients and, in many cases, similarity between tumor
and normal tissue. One other challenge is how to make use
of prior information about the appearance of normal brain.
In this paper we propose a variational brain tumor segmentation algorithm that extends current approaches from
texture segmentation by using a high dimensional feature
set calculated from MRI data and registered atlases. Using manually segmented data we learn a statistical model
for tumor and normal tissue. We show that using a conditional model to discriminate between normal and abnormal regions significantly improves the segmentation results
compared to traditional generative models. Validation is
performed by testing the method on several cancer patient
MRI scans.
1. Introduction
Radiation oncologists, radiologists, and other medical
experts spend a substantial portion of their time segmenting medical images. Accurately labeling brain tumors and
associated edema in MRI (Magnetic Resonance Images) is
a particularly time-consuming task, and considerable variation is observed between labelers. Furthermore, in most
settings the task is performed on a 3D data set by labeling
the tumor slice-by-slice in 2D, limiting the global perspective and potentially generating sub-optimal segmentations.
Subsequently, over the last 15 years, a large amount of research has been focused on semi-automatic and fully automatic methods for detecting and/or segmenting brain tumors from MRI scans.
The process of segmenting tumors in MRI images as op-
posed to natural scenes is particularly challenging [22]. The
tumors vary greatly in size and position, have a variety of
shape and appearance properties, have intensities overlapping with normal brain tissue, and often an expanding tumor can deflect and deform nearby structures in the brain
giving an abnormal geometry also for healthy tissue. Therefore, in general it is difficult to segment a tumor by simple
unsupervised thresholding [12]. Other more successful approaches consider fuzzy clustering [7, 19] or texture information [1, 6].
Often the information extracted from the MRI images is
incorporated into a supervised approach that uses labeled
data and automatically learns a model for segmentation.
Different machine learning (ML) classification techniques
have been investigated: Neural Networks [6], SVMs (Support Vector Machines) [8, 28], MRFs (Markov Random
Fields) [11, 2], and most recently CRFs (Conditional Random Fields) [16].
But, as previously mentioned, statistical classification
may not allow differentiation between non-enhancing tumor
and normal tissue due to overlapping intensity distributions
of healthy tissue with tumor and surrounding edema. One
major advantage when segmenting medical images as opposed to natural scenes is that structural and intensity characteristics are well known up to a natural biological variability or the presence of pathology. Therefore, a geometric
prior can be used by atlas-based segmentation in which a
fully labeled template MR volume is registered to an unknown data set [15, 33, 17, 23].
Having extracted different types of information from the
MRI data (e.g., texture, symmetry, atlas-based priors), one
challenge is to formulate a segmentation process that accounts for it. Although classification techniques have been
widely explored in the field of medical image segmentation
and for this problem, variational and level set techniques are
a powerful alternative segmentation strategy that is now of
substantial interest to this field [24, 19]. It has been shown
that variational techniques can integrate different types of
information in a principled way (e.g., boundary information [3], region information [4, 21], shape priors [5, 27],
texture for vector valued images [26, 25]). Some advantages of using a level set representation is that since the
curve/surface is implicitly represented, topological changes
are naturally possible. There also exist efficient numerical
implementations [29].
In this paper, we propose a variational MRI tumor segmentation method that incorporates both atlas-based priors
and learned statistical models for tumor and healthy tissue.
The formulation extends the Chan-Vese region-based segmentation model [4] in a similar way to texture-based approaches [26, 25]. But instead of using an unsupervised
approach we learn a statistical model from a set of features
specifically engineered for the MRI brain tumor segmentation task. The set of features that we use are calculated
from the original MRI data, in addition to registered brain
templates and atlases similar to Schmidt [28]. We show the
advantage of using a conditional model based on logistic regression compared to the generative model (e.g. Gaussian)
usually used in variational segmentation.
Our multimodal feature set uses specific anatomical priors fully integrated into a region-based variational segmentation, as opposed to Prastawa et al. [24] who uses the level
set only to smooth the segmentation results at the final stage.
It also differs from Ho et al. [13], who proposed a region
competition method implemented on a level set ’snake’ but
without considering template and atlas priors.
In summary the main contributions of this paper are:
• We extract a high dimensional multiscale feature set
from brain MRI images using brain atlases and templates registered with the data. The multi-modal,
multi-scale feature set incorporates both anatomical
priors for the brain tissue and texture information.
• We incorporate this set of features into a 3D variational
region based segmentation method that uses a learned
statistical model defined on the same set of features
to differentiate between normal and pathological tissue
(tumor).
The remaining of this paper is organized as follows. The
next section briefly presents the problem formulation. In
Section 3 we introduce the multidimensional feature set
and mention the methods used for registration and preprocessing. Section 4 describes the choice of statistical
model used for characterizing the regions and Section 5
gives an overview of the entire system. Finally, Section 6
presents the experimental results.
2. Problem formulation
This section presents the general formulation for the
variational 3D segmentation problem without committing
to a general set of features or statistical model for the data.
The next two sections will give detail on the specific choices
for our particular MRI segmentation problem.
Assume we have a multivariate M -dimensional feature
volume
V = {Vi |i = 1 . . . M }
where Vi : Ω ⊂ ℜ3 → ℜ+ and the domain Ω is assumed
open and bounded . The segmentation task consists of finding a surface S (assumed regular) that splits the domain Ω in
two disjoint regions Ω1 , Ω2 . S represents the interface between the regions denoted ∂Ω. Following [21] we can segment by maximizing the a posteriori partitioning probability p(P(Ω)|V(x)) with P(Ω) = {Ω1 , Ω2 } and x ∈ Ω. This
optimization is equivalent to an energy minimization [21].
Two assumptions are necessary: (i) all partitions are equally
possible and (ii) the pixels within each region are independent. Denote the two probability density functions for the
value V(x) to be in region Ω1 and Ω2 with p1 (V(x)) and
p2 (V(x)), respectively. The optimal segmentation is then
found by minimizing the energy:
Z
Z
log p2 (V(x))dx
log p1 (V(x))dx −
E(Ω1 , Ω2 ) = −
Ω2
Ω1
(1)
+α
Z
dS
(2)
S
The first two terms are referred to as data terms and the last
term represents the regularization on the area of S.
One further challenge is defining a family of probability
density functions (PDF) p1 , p2 that approximate the information of each region and are able to discriminate between
the two regions. Section 4 gives detail on the choice of statistical models used for the MRI segmentation.
We now introduce the level set representation by extending the integrals in Equation 1 to the whole domain using
the level set function:
Φ(x) = D(x, S),
if x ∈ Ω1
(3)
Φ(x) = −D(x, S), if x ∈ Ω2
where D(x, S) represents the distance from point x to the
surface (interface) S. Further, set Hǫ (z) and δǫ (z) the regularized Heaviside and Dirac function, respectively. The
energy function from Equation 1 can then be written as:
E(Ω1 , Ω2 ) =
Z
( − Hǫ (Φ) log p1 (V(x))
(4)
Ω
− (1 − Hǫ (Φ)) log p2 (V(x))
(5)
+ α|∇H(Φ)|) dx
(6)
The Euler Lagrange evolution equation for Φ is
3. Intra-volume bias field correction (Nonuniform intensity normalization N3 [30]).
(see [26]):
Φt (x) = δǫ (Φ(x))( log p1 (V(x)) − log p2 (V(x))
∇Φ
)
+ αdiv
|∇Φ|
(7)
(8)
This region segmentation strategy was proposed in [25, 26],
using an unsupervised approach where the parameters for
the region probability distributions are updated at each step
using the corresponding Euler Lagrange equations derived
from (4). We instead adopt a supervised approach where the
parameters are learned a priori from labeled training data.
3. Feature Extraction
Our experimental MRI data consists of T1, T1c (T1 after
injection with contrast agent - gadolinium), and T2 images.
We used two types of features. The first type of features
- image-based features - are extracted from image intensities alone. The second type of features - alignment-based
features - use templates and tissue prior probability atlases
registered with the data. The motivation in using the two
types of features is that although carefully designed imagebased features (such a textures) can improve the segmentation results [1, 6], they do not take advantage of anatomical prior information that is known about the brain, and
hence require a lot of training data. Therefore, as recently
shown [28, 15, 24, 11] spatially aligned templates overcome
many image-based problems such as intensity overlap, intensity heterogeneity and the lack of contrast at structure
boundary. They also allow accurate results to be obtained
with a relatively small number of training images.
4. Alignment of different modalities (normalized mutual
information).
5. Linear and non-linear alignment of the template with
the data (We use the T1 template from [14]).
6. Resampling (β-splines) and intensity normalization of
the template with the data (weighted regression).
Figure 1 illustrates the pre-processing step for two cases.
The left image is the original T1, the middle image shows
the effect of step 1-3 and the right image shown the registered T1 template.
Figure 1. The effect of pre-processing : (left) original T1 (middle) noise-free and intensity corrected (step 1-3) (right) registered
template for T1.
3.1. Data pre-processing
3.2. Features
To define alignment-based features, the data and templates have to be registered both spatially and in intensity.
We perform a pre-processing step that reduces the effect of
noise and intensity variations within and between images,
in addition to achieving spatial registration. This is done in
a pre-processing pipeline as in [28, 34]. For segmentation
purposes, we chose to always align the templates with the
original data in order to minimize the amount of distortion
applied to the original data (e.g. to preserve texture properties). This is also the reason why we used the noise free
and intensity corrected images (step 1,2,3) only for registration purposes and use the original ones in the segmentation.
For the registration and resampling stages we used Statistical Parametric Mapping implementations [32]. The pipeline
consist of:
The image-based features include the original data
modalities and a multi-scale texture characterization. These
types of features are well studied in the context of texture
segmentation. In the framework of variational segmentation, the most common texture features are Gabor (wavelet)
filters (e.g. [21]) and structure tensors [25]. Wavelet-based
texture features have been previously used also for medical
image segmentation [1]. Another approach that has been
discussed in this field is to include intensities of neighboring pixels as additional features and let a classifier learn the
method to combine these intensities [6, 8]. In the present
approach we use a multi-scale Gabor-type feature set [18].
Figure 2 shows 6 examples of texture features.
Alignment-based features were previously used by Kaus
et al. [15] where they define a ’distance transform’ based
on a labeled template. The abnormality metric from [10]
could also be used as a feature of this type. Schmidt [28] extended the use of templates by defining features that use the
template data directly. This approach is valid when using
1. Noise reduction (non-linear filtering).
2. Inter-slice intensity variation correction (weighted regression).
Figure 2. Example of texture-based features (MR8 [18]) - from
top/left t bottom/right: Gaussian, Laplacian of Gaussian, symmetric (3) and antisymmetric (3) features at 3 scales.
machine learning on a pixel-by-pixel basis for segmentation but is not be appropriate for a variational energy based
segmentation.
We use three types of alignment-based features. The first
type is the spatial likelihoods of 3 normal tissues (WMwhite matter, GM-gray matter and CSF) obtained from [14].
The actual features are calculated by taking the differences
between the registered priors with T1 (for GM) and T2 (for
WM,CSF). Figure 3 shows the GM and WM priors (first
row, last two images) and the corresponding features (second row, last two images). A second type is the average intensity maps from a set of individuals aligned with the template coordinate system (also obtained from [14]). Again,
the features were calculated by taking the difference between the registered templates and the data (comparing images of corresponding modalities).
Finally a third type of alignment-based feature is a characterization of left-to-right symmetry. Tumors are typically
asymmetric while normal areas are typically symmetric.
This type of feature was originally proposed by Gering [11].
We characterize the symmetry by subtracting the intensity
value of the pixel on the opposite side of the line of symmetry. Figure 3 shows the symmetry features corresponding to
T1 and T2 (first two images on the bottom row). The line
of symmetry is extracted from the template and registered
with the data (along with the template).
In summary, the original data is transformed in a high dimensional feature set, where each voxel in the brain volume
is characterized by a M dimensional vector x.
4. Region Statistics
The variational segmentation framework presented in
Section 2 requires a statistical model (in the form of a probability density function) to describe the consistency of the
region. The most common choice is the Gaussian distribution [25, 26]. Alternatively, the PDF can be estimated based
on the histogram(s). Rousson et al. [25] proposed a continuous version of the Parzen density while Malik et al. [20]
Figure 3. Example of alignment-based features: (top:) T1, T2,
prior on gray matter GM and prior on white matter WM; (bottom)
symmetry features for T1, T2, features based on GM and WM
priors
used a more complex measure based on texon histograms.
We implemented and evaluated two types of PDFs. The
first one is based on a Gaussian distribution and the other
one is based on a discriminatively-trained Generalized Linear Model that accounts for the discrete nature of the labels
(Logistic Regression).
4.1. Gaussian approximation
First, a general Gaussian approximation is used to model
the vector valued intensity information for ’tumor’ (Ω1 ) and
’normal brain’ (Ω2 ) regions. Since we are working with
multivariate data volumes, the parameters of the Gaussian
model are the M dimensional vector mean µi , and a covariance matrix Σi of dimension M × M (i = 1, 2 one set for
each region: Ω1 - ’tumor’, Ω2 ’normal brain’). The probability of a voxel V(x) to be in Ωi is:
pi (V(x)) = g(V(x)|µi , Σi )
(9)
T −1
1
1
=
e− 2 (V(x)−µi ) Σi (V(x)−µi )
(2π)2 |Σi |1/2
(10)
The parameters {µi , Σi } are estimated from the N labeled
data volumes:
µi =
N
1 XX
V(x)
ni j=1
(11)
N
1 XX
(V(x) − µi )(V(x) − µi )T
ni − 1 j=1
(12)
x∈Ωi
Σi =
x∈Ωi
Under the hypothesis that the channels are not correlated,
the class-conditional PDF can be estimated using the joint
density probabilities from each component. This is equivalent to having a joint diagonal covariance matrix on class
variables.
For modeling the ’normal brain’ area we also tried using
a mixture of two Gaussians as there are two major histogram
Figure 4. Three stages in the surface evolution. (top) 3D surface
and T2 brain intensities; (bottom) horizontal section. the color
code shows how far is the surface from the manual label.
peaks (corresponding to white and gray matter). However,
we did not get better results with this strategy.
4.2. Logistic regression approximation
As a second choice for computing the PDF we used the
Logistic Regression, a discriminative (rather than generative) training strategy. The PDF for ’tumor’ and ’normal
brain’ pixels are given by:
p1 (V(x)) = lr(V(x)|α, β) =
1
1 + exp(−α − β T V(x))
(13)
p2 (V(x)) = 1 − p1 (V(x));
(14)
The Maximum (Log-)Likelihood parameters α (scalar)
and β (vector of dimension M - the number of features),
are estimated from the labeled data (x) using a 2nd-order
nonlinear optimization strategy:
{α, β} = arg max(lr(V(x)|α, β, x))
α,β
(15)
5. Segmentation System and Implementation
Details
We now present an overview of the segmentation system
and some implementation related details (see Figure 5 for
an overview).
In the training phase we used data manually labeled by
radiation oncologists and radiologists. The training data is
pre-processed as described in Section 3.1. Next, the imagebased and alignment-based features presented in Section 3.2
are extracted. We also perform a skull removal using FSL
tools [31] and mask all the features such that we further
process only brain area (note that this remains fully automatic). The PDF (probability density functions) for ’tumor’
and ’normal brain’ regions are computed from all data using
voxels inside and outside the labels, respectively. We used
both models (Gaussian and Logistic Regression) outlined in
Equations 11 and 15.
For the actual segmentation the data set is run through
the same pipeline as the training data including feature extraction and skull removal. The level set is initialized with
the mask corresponding to the extracted brain area. The
evolution uses Equation 7. The voxel probabilities for ’tumor’/’normal brain’ are calculated using Equations 9 and 13
(using the pre-computed PDF parameters). The evolution
PDE is solved using an explicit finite differences scheme on
an irregular grid (the data has 1mm resolution on the slice
plane while about 5mm resolution on the vertical dimension
due to inter-slice distances used during acquisition). The
discretization of the data term is straightforward while for
the parabolic term representing the mean curvature motion
we used central differences.
The only parameter that controls the evolution is α, the
ratio between the regularization and data terms. The parameter was fixed during evolution and set so that the data
and regularization terms were balanced in magnitude (e.g.
about 1/8 for Logistic Regression and 1/300 for the Gaussian). The evolution is stopped when the surface reaches a
stable position. The level set can change topology during
evolution. Although this is an advantage in other applications, it might not be desirable in the case of tumor segmentation (that is in most cases localized in one spot). In
a post-processing step, we remove small surface pieces that
are not part of the tumor area (we remove pieces that are
smaller than half size of the biggest piece). Figure 4 shows
3 stages in the surface evolution.
For visualization, we have created a user interface (see
Figure 6) where the user can visualize the surface evolution
in 3D and on 2D slices (one horizontal and two vertical Training:
(1) preprocess labeled training data (Section 3.1)
(2) extract features (Section 3.2)
(3) compute PDF:’tumor’ p1 and ’brain’ p2
(Section 4)
Segmentation:
Initialization:
(1) preprocess (Section 3.1)
(2) extract features (Section 3.2)
(3) skull removal and init level set
Evolution:
(4) evolve level set until convergence
(Equation 7)
Postprocess:
(5) remove small surface pieces
Figure 5. Overview of segmentation method
Figure 6. Visualization interface (top) and two other data modalities (bottom: left - slices; right - labeled data).
parallel to the tree main plains). The three planes corresponding to the 2D slices can be moved along their axis.
There are two modes of displaying the 3D brain information (as transparent voxels or using the slices). If a manual
segmentation is provided a color code shows the distance
error between the level set and manual segmented surface.
Also, different 3D brain modalities can be displayed (e.g.,
T1, T2, labeled data).
6. Experiments
We validated the proposed method using data from 9 patients having either a grade 2 astrocytoma, an anaplastic astrocytoma or a glioblastoma multiforme. The tumor area
was manually segmented slice-by-slice is each data set by
an expert radiologist. We performed inter-patient training
(training on 8 patients and testing on 1).
For qualitative evaluation we present in Figure 7 the results of using our technique with the two types of statistics (Gaussian and Logistic Regression as described in Section 4) for three patients (each row corresponds to a patient).
We present both the segmented volumes as 3D surfaces as
well as one slice corresponding to the same segmentation.
The first two columns show one slice of the original T1, T2
data, the third column shown the manual segmentation, the
next two column present results for the automatic segmentation and the last two columns illustrate the final segmented
3D surfaces. The color code on the volume shows the distance from the manual segmentation.
For quantitative evaluation we used the VALMET validation framework [9]. The results for the 9 tumor data
sets are shown in Table 1. The volume overlap measures the normalized intersection in voxel space of the
two segmentations (manual X and automatic Y ), given
by Overlap(X, Y ) = (X ∩ Y )/(X ∪ Y ) (equivalent
to Jaccard similarity measure). The Hausdorff distance
from X to Y is Hausdorff(X, Y ) = maxx∈X dist(x, Y ).
To make the Hausdorff distance symmetric the greater of
Hausdorff(X, Y ) and Hausdorff(Y, X) is taken. The last
measure MeanDist(X, Y ) represents the mean absolute
surface distance.
We first compared the two types of statistics - Gaussian
(G) and Logistic Regression (LG). As expected the Logistic
Regression always performs better due to the fact that the
generative distribution for both classes is complex, while
the Logistic forms a better model of binary data. For one
volume the Gaussian model completely failed. For showing
the importance of using the feature set we compared the
method with the conventional variational segmentation on
T1,T2,T1c alone. Results show that when using the features
the method performs about 20% better. Over the data sets
for the Logistic Regression, the highest mean distance was
about 1 voxels (4 mm) and the highest Hausdorff distance
3.5 voxels (14 mm) which is good compared to the 20 mm
margin that is commonly used in brain tumor surgery.
All techniques tend to over-segment (results are smaller
than the manual labels) probably due to the influence of regularization term.
7. Conclusions
We have presented a variational method for brain tumor
segmentation. Existing region-based variational segmentation methods based on texture features are not suited for
tumor segmentation as they are not discriminative enough
when the appearance of tumor and normal tissue overlap.
Using priors of the appearance of anatomical structures in
the normal brain in the form of templates and atlases, we
define a set of multidimensional features and use them to
calculate statistics for ’tumor’ and ’normal brain’ area from
labeled MRI data. We show that a discriminatively-trained
conditional model based on Logistic Regression gives better
results than traditional generative models.
To further improve the results we are going to investigate more sophisticated probability models, including Regularized (MAP) Logistic Regression to reduce the effects
of noise, using kernels to expand the feature representation,
and Bayesian parameter estimation. We are also interested
in exploring multi-class scenarios, where anatomical prior
9
6
3
0
mm
T1
T2
Original data
Manual
labeling
2D view LogReg 2D view Gauss
Segmentation
3D view LogReg
3D view Gauss
3D Surface with distance error
Figure 7. Results for the automatic segmentation compared to the manual segmentation. Each row represents a patient data set. The color
code on the volume shows the distance error from the manual segmentation (see the bar color code).
Case
1
2
3
4
5
6
7
8
9
LRNF
16%
76%
61%
64%
46%
53%
38%
68%
35%
Overlap
G
30%
72%
71%
37%
47%
61%
08%
51%
39%
LR
56%
82%
74%
47%
48%
68%
48%
72%
47%
Hausdorff
LRNF
G
LR
4.83
3.36 2.91
1.32
1.79 1.71
2.90
2.41 1.75
2.13
5.04 2.99
2.69
8.37 2.67
3.63
9.94 3.32
3.84
8.00 3.49
2.89
3.12 2.74
3.09
2.29 2.61
Mean Dist
LRNF
G
LR
1.83
1.09 1.02
0.29
0.35 0.22
0.72
0.51 0.43
0.47
1.10 0.57
0.73
0.76 0.57
0.86
0.76 0.56
0.90
1.54 0.67
0.57
0.81 0.46
0.98
0.78 0.71
Table 1. VALMET scores for the 9 patient data sets with two types of statistics G(Gaussian) LR(Logistic Regression) with full set of
features and without features - only Log. Reg. (LFNF). The overlap score represents the percentage of overlap between the automatic and
manual segmentation (with respect to tumor size). The Hassdorf distance and mean distance are in voxel units (about 3mm).
information could be used to help initialization (as in [23]).
An advantage of variational methods compared to discrete
ones (e.g., MRFs) is that any type of regularization can be
easily incorporated into the energy function. We plan to investigate anisotropic regularization that would preserve discontinuities at boundaries, and encode the expected shape
information of tumor volumes.
References
[1] C. Busch. Wavelet based texture segmentation of multimodal tomographic images.
Computer and Graphics,
21(3):347–358, 1997.
[2] A. Capelle, O. Colot, and C. Fernandez-Maloigne. Evidential segmentation scheme of multi-echo MR images for the
detection of brain tumors using neighborhood information.
Information Fusion, 5(3):203–216, September 2004.
[3] V. Caselles, R. Kimmel, and G. Sapiro. Geodesic active contours. Int. J. Comput. Vision, 22(1):61–79, 1997.
[4] T. Chan and L. Vese. Active contours without edges. IEEE
Trans. Image Processing, 10(2):266–277, 2001.
[5] D. Cremers, T. Kohlberger, and C. Schnorr. Nonlinear shape
statistics in mumford-shah functional. In ECCV, pages 93–
108, 2002.
[6] S. Dickson and B. Thomas. Using neural networks to automatically detect brain tumours in MR images. International
Journal of Neural Systems, 4(1):91–99, 1997.
[7] L. Fletcher-Heath, L. Hall, D. Goldgof, and F. R. Murtagh.
Automatic segmentation of non-enhancing brain tumors
in magnetic resonance images. Artificial Intelligence in
Medicine, 21:43–63, 2001.
[8] C. Garcia and J. Moreno. Kernel based method for segmentation and modeling of magnetic resonance images. Lecture
Notes in Computer Science, 3315:636–645, Oct 2004.
[9] G. Gerig, M. Jomier, and M. Chakos. ”valmet”: a new validation tool for assessing and improving 3d object segmentation. In MICCAI, pages 516–523, 2001.
[10] D. Gering. Diagonalized nearest neighbor pattern matching
for brain tumor segmentation. MICCAI, 2003.
[11] D. Gering. Recognizing Deviations from Normalcy for Brain
Tumor Segmentation. PhD thesis, MIT, 2003.
[12] P. Gibbs, D. Buckley, S. Blackb, and A. Horsman. Tumour
volume determination from MR images by morphological
segmentation. Physics in Medicine and Biology, 41:2437–
2446, 1996.
[13] S. Ho, E. Bullitt, and G. Gerig. Level set evolution with
region competition: automatic 3D segmentation of brain tumors. In 16th International Conference on Pattern Recognition, pages 532–535, 2002.
[14] Icbm view: an interactive web visualization tool for
stereotaxic data from the icbm and other projects,
http://www.bic.mni.mcgill.ca/icbmview/, Online.
[15] M. Kaus, S. Warfield, A. Nabavi, P. Black, F. Jolesz, and
R. Kikinis. Automated segmentation of MR images of brain
tumors. Radiology, 218:586–591, 2001.
[16] C.-H. Lee, M. Schmidt, A. Murtha, A. Biaritz, J. Sander,
and R. Greiner. Segmenting brain tumors with conditional
random fields and support vector machines. In Workshop
on Computer Vision for Biomedical Image Applications at
ICCV, 2005.
[17] K. Leemput, F. Maes, D. Vandermeulen, and P. Suetens.
Automated model-based tissue classification of MR images
of the brain. IEEE Transactions on Medical Imaging,
18(10):897–908, October 1999.
[18] T. Leung and J. Malik. Representing and recognizing the
visual appearance of materials using three-dimensional textons. Interational Journal of Computer Vision, 43(1):29–44,
Jun 2001.
[19] J. Liu, J. Udupa, D. Odhner, D. Hackney, and G. Moonis.
A system for brain tumor volume estimation via mr imaging
and fuzzy connectedness. Comput. Medical Imaging Graph.,
21(9):21–34, 2005.
[20] J. Malik, S. Belongie, T. Leung, and J. Shi. Color and texture
analisys for image segmentation. IJCV, 43(1):7–27, 2001.
[21] N. Paragios and R. Deriche. Geodesic active regions: A new
paradigm to deal with frame partition problems in computer
vision. Visual Communication and Image Representation,
13:249–268, 2002.
[22] M. R. Patel and V. Tse. Diagnosis and staging of brain tumors. Semonars in Roentgenology, 39(3):347–360, 2004.
[23] M. Prastawa, E. Bullitt, S. Ho, and G. Gerig. A brain tumor
segmentation framework based on outlier detection. Medical
Image Analysis, 8(3):275–283, September 2004.
[24] M. Prastawa, E. Bullitt, N. Moon, K. Leemput, and
G. Gerig. Automatic brain tumor segmentation by subject
specific modification of atlas priors. Academic Radiology,
10(12):1341–1348, December 2003.
[25] M. Rousson, T. Brox, and R. Deriche. Active unsupervised
texture segmentation on a diffusion based feature space. In
CVPR, 2003.
[26] M. Rousson and R. Deriche. A variational framework for
active and adaptative segmentation of vector valued images.
In MOTION ’02: Proceedings of the Workshop on Motion
and Video Computing, page 56, 2002.
[27] M. Rousson and N. Paragios. Shape priors for level set representations. In ECCV (2), pages 78–92, 2002.
[28] M. Schmidt. Automatic brain tumor segmentation. Master’s
thesis, University of Alberta, 2005.
[29] J. Sethian. Level Set Methods. Cambridge University Press,
1996.
[30] J. Sled, A. Zijdenbos, and A. Evans. A nonparametric
method for automatic correction of intensity nonuniformity
in MRI data. IEEE Transactions on Medical Imaging, 17:87–
97, February 1999.
[31] S. Smith, M. Jenkinson, M. Woolrich, C. Beckmann,
T. Behrens, H. Johansen-Berg, P. Bannister, M. D. Luca,
I. Drobnjak, D. Flitney, R. Niazy, J. Saunders, J. Vickers,
Y. Zhang, N. D. Stefano, J. Brady, and P. Matthews. Advances in functional and structural mr image analysis and
implementation as fsl. NeuroImage, 23(1):208–219, 2004.
[32] Statistical
parametric
mapping,
http://www.fil.ion.bpmf.ac.uk/spm/, Online.
[33] W. Wells, P. Viola, and R. Kikinis. Multi-modal volume registration by maximization of mutual information. In Medical
Robotics and Computer Assisted Surgery, pages 55–62. Wiley, 1995.
[34] A. Zijdenbos, R. Forghani, and A. Evans. Automatic
“pipeline” analysis of 3-d mri data for clinical trials: application to multiple sclerosis. IEEE Transactions on Medical
Imaging, 21(10):1280–1291, Oct 2002.