A New Algorithm for Fully Automatic Brain Tumor Segmentation with 3-D Convolutional Neural Networks

A New Algorithm for Fully Automatic Brain Tumor
Segmentation with 3-D Convolutional Neural Networks
Christopher Elamri
Stanford University
mcelamri@stanford.edu
Teun de Planque
Stanford University
teun@stanford.edu
Abstract
Glioblastoma (GBM) is the most common primary tumor
of the central nervous system in the United States, with 3 in
100,000 people diagnosed every year [1]. In this paper,
we present a new algorithm for fully automatic brain tu-
mor segmentation based on 3-D Convolutional Neural Net-
works. Segmentation of the GBM region of the surround-
ing brain makes it easier to access the image data within
MR scans, and consequently can help us better understand
GBMs. Most methods for volumetric image data use 2-
D convolutions; we present a true generalization of CNNs
to 3-D filters to preserve spatial information and increase
robustness. In addition, we present a relatively high bias
CNN architecture that enables us to both expand the effec-
tive data size and reduce the variance of our model. Our
median Dice score accuracy is around 89 percent in the
whole tumor segmentation. This result represents a signif-
icant improvement over past algorithms and demonstrates
the power of our approach in generalizing low-bias high-
variance methods like CNNs to learn from medium-size data
sets.
1. Introduction
Glioblastoma (GBM) is the most prevalent and most ag-
gressive type of brain cancer [2]. As a result of its resistance
to existing therapeutic methods, the most common length
of survival following diagnosis is 12 to 15 months, and less
than 3-5% of patients survive more than five years [1][3].
Much is still being done to understand GBM and its two
subtypes, high grade glioma (HGG) and low grade glioma
(LGG), but there is still a gold mine of untapped data. Chief
amongst these is imaging data in the form of magnetic res-
onance (MR) scans. Most methods of analyzing and ex-
tracting quantitative information from this imaging data re-
quires some form of segmentation of the GBM, and bet-
ter yet, classification of the tumor into four sub-categories:
necrosis, edema, non-enhancing tumor, and enhancing tu-
mor. To date, the gold standard of segmentation is still hu-
man radiologist segmentation. However, with the pure size
of imaging data being accrued, manual segmentation of all
images is no longer a sustainable system.
Nonetheless, existing automated segmentation algo-
rithms are based on high-bias learning algorithms and have
failed to give substantial improvement in classification ac-
curacy as described by Subbanna et al., Menze et al., and
Bauer et al. [4]. As a result, these high-bias systems are in-
creasingly being replaced with data-driven low-bias models.
The development of low-bias systems such as convolutional
neural networks for medical purposes is, nevertheless, chal-
lenging due to their high-variance, as many medical datasets
are limited in size.
In this paper, we propose a novel approach to automatic
brain tumor segmentation. We take as input a 3-D MR im-
age and output one of five labels (normal or one of 4 tu-
mor subtypes) for each voxel (the 3-D analog of pixel).
Our new approach surpasses current methods by reducing
the variance of CNNs applied to medium datasets. First,
where most methods use 2-D convolutions for 3-D data, we
introduce a CNN that uses 3-D filters, made computation-
ally efficient by the transformation to Fourier space. Con-
sequently, our system is more robust and minimizes loss of
spatial information. Second, we present a novel CNN archi-
tecture which differs from those traditionally used in brain
segmentation and computer vision [4][5]; our CNN archi-
tecture uses a higher-bias system than CNNs using pixel-
wise neural networks that enables us to both expand the
effective data size from patient number to total pixel and
reduce the variance of CNNs. Our algorithm for GBM seg-
mentation and characterization exploits the CNNs frame-
work and the structure of medical imaging data to train a
model that not only is highly robust but also generalizes
data-driven low-bias systems to medium data while mini-
mizing variance.
1

2. Problem Statement
Our problem amounts to efficiently classifying brain
voxels into five categories: non-tumor, necrosis, edema,
non-enhancing, and enhancing. The input to our system
consists of a preprocessed 3-D image of a brain. Each 3-
D image is of size 240 × 240 × 155 × 4, where the depth
4 represents the four modalities of our input images. The
goal is to for each voxel correctly and efficiently output its
category (non-tumor, necrosis, edema, non-enhancing, and
enhancing).
2.1. Dataset
We use the data provided by the MICCAI Multimodal
Brain Tumor Image Segmentation Challenge (BRATS) [4].
The 2015 BRATS dataset contains 274 patients: 220 pa-
tients with high grade glioblastomas and 54 patients with
low grade glioblastomas. For each patient our data contains
four MR modalities: T1 post-contrast, T1 pre-contrast, T2
weighted, and FLAIR. In addition, the data contains an ex-
pert segmentation for each patient that we will treat as the
ground truth. The expert segmentation was made by pool-
ing together the segmentation of eleven different radiolo-
gists and includes pixel-wide labeling as one of five cat-
egories: non-tumor, necrosis, edema, non-enhancing, en-
hancing.
All the data has been preprocessed, involving skull strip-
ping, co-registration of the images, and interpolation of the
images so that all images are of the same size (240 × 240 ×
155 pixels). We were not given any information about the
spatial scale of the pixels; and thus, all units for our current
method will be in the given pixel-space.
3. Technical Approach
3.1. Convolutional Neural Network Motivation
Convolutional neural networks (CNNs) are powerful
low-bias systems in deep learning with remarkable perfor-
mance in complex tasks ranging from facial recognition to
image classification [7][8]. CNNs use 2D convolutions as
filters for feature learning and feed the results of these con-
volution layers into a fully connected neural network. Our
algorithm builds on the idea of a CNN as a fully connected
neural network trained on sequential convolution layers to
improve accuracy and computational efficiency in learning
from 3D imaging data. Specifically, we generalize the idea
of 2D convolutions to 3D filters that retain the full spatial
information of the original stack of images. We then decou-
ple the pixels when providing the training data for a fully
connected neural network, thus increasing the size of our
training set and providing a computational speedup.
(a) Modalities (b) Labels
(c) Modalities (d) 3D-Rendering of Tumor
Figure 1: Data Visualization. (a) All four modalities are
co-registered. The modalities are ul: T1-pre, ur: T1-post,
dl: T2W, dr: FLAIR. (b) Corresponds to (a). Better visual-
ization of the four subregions of the tumor (Y: necrosis, G:
Edema, B: Non-Enhancing, P: Enhancing). (c) Visualiza-
tion of the colored labels on top of the modality images for
a different patient. (ul: T1-pre, ur: T1-post, dl: T2W, dr:
FLAIR) (d) 3-D rendering of the tumor within the brain.
3.2. 3-D Convolutions
We can reduce the bias of our system by considering the
input image as a 3-dimensional space of pixels. Expanding
our convolution from our usual 2-dimensional convolutions
into 3-dimensional convolutions, we get the following equa-
tion:
(I ∗ f)[x, y, z] =
240
τx=1
240
τy=1
155
τz=1
I[τx, τy, τz] ˙f[x − τx, y − τy, z − τz]
where I is the 3-dimensional image and f is the 3-
dimensional filter. Given an image of size n × n × n and
a filter of size m × m × m the time complexity of the con-
volution operation is O(m3
n3
). We can reduce the time
complexity of the convolution operation to O(n3
log(n)) by
implementing our convolution as an element-wise multipli-
cation in Fourier space. We use a 3-dimensional Difference
of Gaussian filter:
DoG(σ) =
1
(2πσ2)3/2
exp
−x2 + y2 + z2
2σ2
−
1
(πσ2)3/2
exp
−x2 + y2 + z2
σ2
Each filter is a difference of two 3-dimensional Gaussians
(normalized) with scales
√
2σ and σ. We create 8 of these
filters with the scales [
√
2, 2, 2
√
2, ..., 16]. Difference of
2

Figure 2: Visualization of 3-D Convolution. Similar to the
2-dimensional equivalent, in a 3-dimensional convolution,
we create an output image from the dot-product of our 3-
dimensional filter and a sub-image centered at every pixel.
We represent the filter with the red-box, our image with blue
box, and the convolutional product with our white box
Gaussian is an effective blob detector since it is rotationally
symmetric.
To create our final convolutional products, we will con-
sider all four modalities, and the magnitude of the gradient
of those images. In addition to the eight filter products, we
will include the original pixel intensity and the magnitude
of the gradient value. Thus, for each of the four modalities,
we will get an expansion into 18 feature images, giving us a
total feature space of 72-dimensions for each of our pixels.
We choose a filter of size 33 × 33 × 33. Basically, we get
for each voxel [x, y, z] the following set of features:
• 1 voxel intensity V [x, y, z]
• 1 voxel intensify gradient V [x, y, z]
• 8 DoG convolutions (V ∗ DoG)[x, y, z]
• 8 DoG convolutions in gradient space ( V ∗
DoG)[x, y, z]
3.3. 3-D Convolutional Neural Network Architec-
ture
Our input consists of a 3-dimensional image with four
channels (one for each modality). Our first layer is a non-
trained convolutional layer with 72 240×240×155×4 fil-
ters, 3-dimensionally convolved over all channels with our
input image. In this layer, our filters are the hand-selected
Difference of Gaussian filters for a specified modality with
all other values in the other three channels being zero. Our
training does not go through this layer.
The following convolutional layers use convolutions of
filters sized 1 × 1 × 1 over the number of channels in the
preceding layer (either 72 for the first layer or 100 for all
Figure 3: Main Methodology of Learning Pipeline. We
start with our four modalities of input images. To that,
we will convolve our 72 sparse 3-dimensional Difference
of Gaussian filters to get our first convolution layer. From
that point on, we will be convolving a trivial 1 × 1 × 1 (so a
scalar) filter over all channels to get the following convolu-
tional layer. The convolution of a 1 × 1 × 1 filter essentially
decouples each pixels information from each other after the
initial layers convolution. On the last layer, we will get five
channels, each channel giving us a probability score for ev-
ery pixel in our image for each of the five categories.
other layers). The 1×1×1 filter is a scalar that we will con-
stantly multiply over the whole image. In a sense, by only
wiring the convolution layers onward with these 1 × 1 × 1
filters, we are decoupling all of the pixels in our image. Be-
cause the first convolution layer couples neighborhood in-
formation via the 33×33×33 convolutions, all subsequent
layers pixels will maintain some level of information of its
original neighborhood. Thus, we are able to decrease the
number of weights we would have to train while drastically
increasing our effective training data size.
On the last layer, we will get five channels, each channel
giving us a probability score for every pixel in our image
along the five sub-regions (0 = non-tumor, 1 = necrosis, 2 =
edema, 3 = non-enhancing, and 4 = enhancing). We use the
softmax function as the loss function. To train the convolu-
tional neural network we use stochastic gradient descent by
patient. During test time, we classify using a voting met-
ric that assigns a voxel to one of the five categories based
on its classification scores (see algorithm 1). This method
is equivalent to how the expert segmentations were pooled
from 11 different radiologist’s segmentations.
Essentially, our classification priority in descending or-
der goes: enhancing, necrosis, non-enhancing, and edema.
If our neural nets returns positive classifications for multiple
tumor subtypes, we classify to the positive subtype with the
3

highest priority. This hierarchical design is based off of the
hierarchical majority vote used to combine several different
algorithmic results [4].
This seemingly arbitrary methodology makes perfect
sense in the context of our classification problem. Tumor
segmentations are judged generally in terms of three accu-
racies: whole tumor accuracy, tumor core accuracy, and en-
hancing tumor accuracy. Thus, because they have their own
accuracy scores, we must prioritize classification of the core
over the non-core (edema), and then also the enhancing core
over the other core. The enhancing core generally covers a
smaller area of the brain, which lends even more reason to
be more sensitive to its detection.
Results are reported as the standard Dice score calculated
via 10-fold cross validation (see: beginning of next section).
We do not use cross validation to select parameters, decid-
ing to keep our neural net parameters set to default values.
This is both because the additional computation time would
be prohibitive, and also because our Dice scores (which are
also calculated from the cross validation) would become bi-
ased upwards.
Algorithm 1 Hierarchical majority decision Our neural
network outputs a number between 0 and 1 per voxel for
each of the four tumor structures (edema, non-enhancing,
necrosis, enhancing), respectively indicated by pedema,
pnonenh, pnecrosis, penh
1: procedure HIERARCHICALDECISION(Score)
2: label ← NormalTissue
3: if pedema >= 0.5 then
4: label ← Edema
5: end if
6: if pnonenh >= 0.5 then
7: label ← Non − Enhancing
8: end if
9: if pnecrosis >= 0.5 then
10: label ← Necrosis
11: end if
12: if penh >= 0.5 then
13: label ← Enhancing
14: end if
15: end procedure
4. Results
We use Dice Coefficient to compute our results (also
known as the Sørensen-Dice Index) [9].
DiceScore =
2|Pred ∩ Ref|
|Pred| + |Ref|
where Pred is the set of voxels designated as the region of
interest in the prediction and Ref is the set of voxels desig-
Figure 4: Performance Comparison. Comparison of our
3D CNN method to the leading methods of both 2013, 2014,
radiologists, and our own 2D implementation of our algo-
rithm. The main metric used for our accuracy is Dice Score.
Whole = (Edema, Non-enhancing, Necrosis, Enhancing)
Figure 5: Histogram of Dice Score Accuracies
nated as the region of interest in the reference for compari-
son (in our case, the expert segmentation).
The median Dice performance on the whole tumor de-
tection is 89%! To wit, the inter-radiologist repeatability
is only 85%, so our accuracy has saturated with respect to
the ground truth. Figure 4 shows a comparison amongst the
leading algorithms in the literature. The first row shows the
performance of human radiologists. The full distribution of
accuracy of the algorithm among the 274 patient is in figure
5. One particularly successful segmentation can be seen in
figure 6.
5. Discussion
Our median performance on the whole tumor detection
revolves around a dice score of 89, which is right before
losing clinical significance. This is very competitive with
previous methods. One clear issue with our algorithm is
dealing with outliers as can be seen in figure 5. More than
half of our segmentations are wildly successful, but some
segmentations return sub-50% scores, which you would not
typically see with a radiologist.
4

Figure 6: Slices-by-Slices Result A representative slice
view of a test case giving 98% Dice score accuracy. Each
of the 12 images has three parts. On the left is the T1 post-
contrast image at the given slice, the middle image is the
expert segmentation labeling, and the right-hand side image
is the labelling at the slice by our algorithm. The label color
scheme follows that which was introduced in figure 1.
figure 4 shows a comparison amongst the top results in
the literature. Some notable ones in 2013 include ones by
Zhao and Subbanna which incorporated Markov Random
Fields (MRF), achieving a Dice accuracy of 82, respectively
[4]. Festa from 2013 used random forests to achieve a Dice
of 62 [4]. In 2014, groups used deep learning and convolu-
tion neural nets (CNNs) to achieve accuracies of 85 (Davy)
and 88 (Urban) [10]. Our method improves upon the ex-
isting vanilla convolutional neural network architecture by
making some high biased assumptions that allow us to dras-
tically reduce our variance.
Basically, one of our assumption is that voxels are inde-
pendent, only coupled by information tied to their feature
vectors. Certainly, this is a high-bias assumption, but it al-
lows us to use n = 4.5 billion training examples rather than
only n = 274 patient samples. Neural nets are ultimately
low-bias systems, but our hope is that the improved vari-
ance caused by the enlarged sample space will overcom-
pensate for our high-bias assumptions. Furthermore, we
contrast our method with standard deep learning algorithms
like CNN, which learn the features using convolution ker-
nels on the inputs. Vanilla CNNs use each patient as a single
training sample. We, on the other hand, select Difference of
Gaussian (DoG) convolution filters to relate voxels to their
neighborhoods. Although we are highly biasing our system
compared to usual deep learning framework, our features
may be more adequate since our subsequent feed-forward
neural net can then learn higher level features from each
pixel’s lower level features.
6. Conclusion
The higher-biased CNN 3-D architecture performs re-
markably well on the glioblastoma segmentation problem.
The segmentation results are competitive with those using
much more complex methods, and we argue our success
is due to our smart choice of features along with a greatly
enlarged sample space and flexible training method (neural
nets). Our algorithm is powerful despite its relatively high
bias, and we hope that it may serve the medical community
in their work.
The natural next step of this project is a thorough anal-
ysis of our models asymptotes. We have claimed that our
large data set has significantly reduced model variance, but
it is unknown whether we can further reduce variance with
more data. Given that our segmentation algorithm is al-
ready on par with our reference expert segmentations, we
suspect but would like to confirm that our model has already
reached its large-data asymptotic performance.
7. Acknowledgments
This project would not have been possible without the
help of our research mentor Darvin Yi and the Gevaert
Biomedical Informatics Lab. In addition, we would like to
thank the CS231N course staff for their ongoing support.
References
[1] World Cancer Report 2014. World Health Organization.
2014. pp. Chapter 5.16.
[2] Bleeker, Fonnet E., Remco J. Molenaar, and Sieger
Leenstra. ”Recent Advances in the Molecular
Understanding of Glioblastoma.” J Neurooncol Journal
of Neuro-Oncology 108.1 (2012): 11-27. Web.
[3] Gallego, O. ”Nonsurgical Treatment of Recurrent
Glioblastoma.” Current Oncology Curr. Oncol. 22.4
(2015): 273. Web.
[4] Menze, B. H. The Multimodal Brain Tumor Image
Segmentation Bench-mark (BRATS). IEEE Transactions
on Medical Imaging 2014 (2013).
[5] Njeh, Ines, Lamia Sallemi, Ismail Ben Ayed, Khalil
Chtourou, Stephane Lehericy, Damien Galanaud, and
Ahmed Ben Hamida. ”3D Multimodal MRI Brain
Glioma Tumor and Edema Segmentation: A Graph
Cut Distribution Matching Approach.” Computerized
Medical Imaging and Graphics 40 (2015): 108-19. Web.
[6] Smirniotopoulos, James G., Frances M. Murphy,
Elizabeth J. Rushing, John H. Rees, and Jason W.
Schroeder. ”Patterns of Contrast Enhancement in the
Brain and Meninges1.” RadioGraphics 27.2 (2007):
525-51. Web.
5

[7] Lawrence, S., C.l. Giles, Ah Chung Tsoi, and A.d.
Back. ”Face Recognition: A Convolutional Neural-
network Approach.” IEEE Trans. Neural Netw. IEEE
Transactions on Neural Networks 8.1 (1997): 98-113.
Web.
[8] Krizhevsky, A., Sutskever, I., and Hinton, G.E.
”ImageNet Classiﬁcation with Deep Convolutional
Neural Networks.” Adv. Neural Inf. Process. Syst.
(2012) 1-9. Web.
[9] Srensen, T. A method of establishing groups of equal
amplitude in plant sociology based on similarity of
species and its application to analyses of the vegetation
on Danish commons. Kongelige Danske Videnskabernes
Selskab. (1948) 5 (4):1-34.
[10] Challenge Manuscripts.” (2014) MICCAI 2014.
Harvard Medical School, Boston, Massachusetts
(2014) .
6

A New Algorithm for Fully Automatic Brain Tumor Segmentation with 3-D Convolutional Neural Networks

Recommended

More Related Content

What's hot (20)

Similar to A New Algorithm for Fully Automatic Brain Tumor Segmentation with 3-D Convolutional Neural Networks (20)

Recently uploaded (20)

A New Algorithm for Fully Automatic Brain Tumor Segmentation with 3-D Convolutional Neural Networks