Exploiting Bilateral Symmetry in Brain Lesion Segmentation with

Reﬂective Registration

Kevin Raina

, Uladzimir Yahorau

and Tanya Schmah

Department of Mathematics and Statistics, University of Ottawa, Ontario, Canada

Keywords:

Stroke, Brain Lesions, Lesion Mapping, Image Segmentation, MRI, Convolutional Neural Network.

Abstract:

Brain lesions, including stroke lesions and tumours, have a high degree of variability in terms of location,

size, intensity and form, making automatic segmentation difﬁcult. We propose an improvement to existing

segmentation methods by exploiting the bilateral quasi-symmetry of healthy brains, which breaks down when

lesions are present. Speciﬁcally, we use nonlinear registration of a neuroimage to a reﬂected version of itself

(“reﬂective registration”) to determine for each voxel its homologous (corresponding) voxel in the other hemi-

sphere. A patch around the homologous voxel is added as a set of new features to the segmentation algorithm.

To evaluate this method, we implemented two different CNN-based multimodal MRI stroke lesion segmenta-

tion algorithms, and then augmented them by adding extra symmetry features using the reﬂective registration

method described above. For each architecture, we compared the performance with and without symmetry

augmentation, on the SISS Training dataset of the Ischemic Stroke Lesion Segmentation Challenge (ISLES)

2015 challenge. Using linear reﬂective registration improves performance over baseline, but nonlinear reﬂec-

tive registration gives signiﬁcantly better results: an improvement in Dice coefﬁcient of 13 percentage points

over baseline for one architecture and 9 points for the other. We argue for the broad applicability of adding

symmetric features to existing segmentation algorithms, speciﬁcally using the proposed nonlinear, template-

free method.

1 INTRODUCTION

Segmentation of lesions in neuroimages, also called

lesion mapping, can give valuable information for

prognosis, treatment planning and monitoring of dis-

ease progression. The “gold standard” for lesion seg-

mentation is still manual delineation by a human ex-

pert, going through each of the horizontal slices of

the three-dimensional image and labeling each sepa-

rate voxel as either healthy or belonging to a lesion.

This is tedious, time-consuming, and often imprac-

tical, and therefore in practice, a human expert usu-

ally gives only a qualitative assessment of lesions.

Further, there is inter-observer variability; the size

of this variability varies signiﬁcantly by task, but we

note that an average Dice score of 0.58 for overlap

of manually-outlined lesions by two raters was re-

ported for the ISLES2016 challenge (Winzeck et al.,

2018). These observations indicate a need for auto-

matic brain lesion segmentation algorithms. How-

https://orcid.org/0000-0002-6240-9675

https://orcid.org/0000-0002-0522-4148

https://orcid.org/0000-0002-0404-8824

ever, accurate lesion segmentation is a challenging

task for many reasons, including large variability in

location, size, shape and frequency of lesions across

patients.

While a plethora of automatic lesion segmenta-

tion methods has been proposed, most of the currently

leading methods are based on convolutional neural

networks (CNN) (Winzeck et al., 2018). Many of

these use 2D CNNs, where the 3D neuroimage is pro-

cessed as a sequence of independent 2D slices. These

approaches are arguably suboptimal, since they do not

take into account the 3D spatial structure of the data.

Nonetheless, many 2D methods have shown promis-

ing results, including the methods of Havaei et al.

(Havaei et al., 2017) and Kamnitsas et al. (Kamnitsas

et al., 2017), which we use as baseline architectures in

the present paper. Some other works have used CNNs

with an input of three orthogonal patches around each

voxel being classiﬁed, thus incorporating some 3D in-

formation, however this signiﬁcantly increased mem-

ory requirements and computational complexity. The

technique of dense inference greatly sped up infer-

ence time, and led to several successful 3D segmenta-

tion methods, see discussion and references in Kam-

116

Raina, K., Yahorau, U. and Schmah, T.

Exploiting Bilateral Symmetry in Brain Lesion Segmentation with Reﬂective Registration.

DOI: 10.5220/0008912101160122

In Proceedings of the 13th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2020) - Volume 2: BIOIMAGING, pages 116-122

ISBN: 978-989-758-398-8; ISSN: 2184-4305

nitsas et al. (Kamnitsas et al., 2017).

We propose a general method for improving ex-

isting segmentation algorithms, including all of the

CNN-based methods mentioned above, by exploit-

ing the bilateral quasi-symmetry of healthy brains.

This symmetry often breaks down when lesions are

present. In particular, stroke lesions usually affect

only one hemisphere; while for some other lesion

types such as tumours, lesions may be present in both

hemispheres but any symmetry is coincidental and

rare. The basic idea is illustrated in Figure 1. The

ﬁrst subﬁgure shows an axial slice of a brain with a

stroke lesion in one hemisphere, and two homologous

(“mirror”) voxels, i.e voxels in corresponding parts

of the brain but in opposite hemispheres. Through-

out this paper, homologous voxels are determined us-

ing “reﬂective registration”, which is registration of

an image to its own reﬂection, as detailed in the next

section. In healthy normal brains, there is a strong

correlation between intensities of homologous vox-

els. In lesioned brains, voxels in a lesion often have

intensities very different from the intensities of their

homologous voxels, as shown in Fig. 1 (b). On the

other hand, lesions typically represent a small propor-

tion of total brain volume, so non-lesion voxels typi-

cally have non-lesion mirror voxels as well, typically

resulting in small intensity differences (if the mirror

voxels have been accurately located). The distribu-

tion of these intensity differences, for both lesion and

non-lesion voxels, is illustrated in Figures 1 (c) and

(d). The difference between these two subﬁgures is in

the nature of the reﬂective registration used to iden-

tify homologous voxels: in (c), “linear” (afﬁne) reg-

istration is used, while in (d), nonlinear registration

is used. The increase in mass around zero for non-

lesions when using nonlinear registration (d) in com-

parison to linear registration (c) suggests the supe-

riority of nonlinear reﬂective registration in locating

mirror voxels. In both cases, the pattern of intensity

differences can be used to aid the classiﬁcation of a

voxel as lesion or non-lesion. This method is inspired

by the clinical practice of radiologists, who make fre-

quent use of comparisons with homologous areas to

detect abnormalities.

Our method, explained in more detail below, uses

3D nonlinear registration of a neuroimage with a re-

ﬂected version of itself to determine for each voxel its

homologous voxel in the other hemisphere. A patch

around the homologous voxel is added as a set of

new features to the segmentation algorithm. To eval-

uate this method, we implemented two baseline mul-

timodal MRI stroke lesion segmentation algorithms,

both based on 2D CNNs, following Havaei et al.

(Havaei et al., 2017) and Kamnitsas et al. (Kamnit-

sas et al., 2017), and then augmented them by adding

extra symmetry features as described above. For each

architecture, we evaluated the baseline method and

two versions of symmetry augmentation: one using

“linear” (afﬁne) registration only, and one using non-

linear registration. We compared the performance of

these three segmentation methods on the SISS Train-

ing dataset of the Ischemic Stroke Lesion Segmen-

tation Challenge (ISLES 2015)(Maier et al., 2017).

Though our experiments use 2D CNNs, our method

can be applied without modiﬁcation to 3D CNNs.

We are aware of three prior works that have also

used brain quasi-symmetry to improve the perfor-

mance of CNN based methods: (Shen et al., 2017),

(Wang et al., 2016) and (Zhang et al., 2017). Shen

et al. use the SIFT-based method of Loy and Eklundh

(Loy and Eklundh, 2006) to identify homologous vox-

els, and report a mean improvement of 3% in Dice

scores on the high-grade (HG) BraTS2013 (brain tu-

mor) Training set. Wang et al. report a higher in-

crease in mean Dice scores, from 0.63 to 0.78, on

a private test set of 8 brains with chronic stroke le-

sions. However the method they use is unclear; the

absence of such an explanation suggests a simple lin-

ear transformation, perhaps a reﬂection in the medial

(mid-sagittal) plane. Zhang et al. uses a reﬂection

in the medial plane of images that have already been

registered to a template (with the registration method

unspeciﬁed). Evaluations on the BraTS2017 dataset

show that adding symmetry features speeds conver-

gence of the algorithm, however there is no consis-

tent improvement in segmentation accuracy. All three

methods require homologous voxels to be in the same

axial plane, a restriction that our method does not

have.

We note that the idea of using symmetry also

appeared in early literature, prior to the widespread

use of neural networks, for example in (Meier et al.,

2014), (Schmidt et al., 2005) and (Dvorak et al.,

2013). These works are based on an initial linear reg-

istration of each subject’s brain to a template. This

is also true of (Tustison et al., 2015), with the ma-

jor differences being the use of multiple modalities

and nonlinear registration. Our contributions are thus

two-fold: (1) Using template-free registration of an

image with a reﬂected version of itself, called reﬂec-

tive registration, to produce supplementary features

for use by segmentation algorithms; and (2) demon-

strating that nonlinear reﬂective registration is better

than linear reﬂective registration for locating mirror

voxels, as judged by improved segmentation results.

Exploiting Bilateral Symmetry in Brain Lesion Segmentation with Reﬂective Registration

117

Figure 1: The quasi-symmetry property of the normal brain can be used to aid lesion segmentation. Subﬁgure (a) shows a

lesion voxel (red) and its non-lesion (blue) mirror voxel (projected onto the same axial slice); (b) shows the voxel intensity

plotted against the intensity of its mirror for a sample of 600 voxels taken from the same brain and equally divided between

lesions and non-lesions; (c) and (d) show superimposed densities for the difference between standardized intensities of a

voxel and its mirror, based on a sample of 2000 voxels taken from the same brain and equally divided amongst lesions and

non-lesions, where homologous voxels have been identiﬁed using linear and nonlinear reﬂective registration in (c) and (d)

respectively.

2 METHODS

Our two baseline algorithms are slight modiﬁcations

of: (i) TwoPathCNN by Havaei et al. (Havaei et al.,

2017), see Figure 2; and (ii) Wider2dSeg by Kamnit-

sas et al. (Kamnitsas et al., 2017). All of the 2D archi-

tectures in Kamnitsas et al. (Kamnitsas et al., 2017),

including Wider2dSeg, are two dimensional variants

of their 3D deepMedic architecture. The 2D archi-

tectures vary in the number of layers, feature maps

(FMs) per layer, and FMs per hidden layer. See Table

B.1 in Appendix B of Kamnitsas et al. (Kamnitsas

et al., 2017) for more details. We ﬁrst describe the ar-

chitectures and training of these baseline models, and

then describe how to compute and append symmet-

ric features so as to preserve dense inference on 2D

images of arbitrary size.

Figure 2: TwoPathCNN architecture, reproduced from

Havaei et al. (Havaei et al., 2017) with permission of the

authors. Note that the ﬁnal output is a 2 × 1 × 1 tensor for

our application.

2.1 Baseline Models

TwoPathCNN and Wider2dSeg are convolutional

neural networks. Both architectures take as input

one or, as in the case of Kamnitsas’ 2D architec-

tures, two stacks of four patches from different MRI

modalities. The networks branch into two path-

ways. TwoPathCNN consists of three convolutional

blocks of sizes shown in Fig. 2, which also shows

the locations of maxout and max pooling operations.

Wider2dSeg is a deeper architecture with 16 convo-

lutional blocks that makes use of multiscale features

through downsampling, convolution and upsampling

back to the original scale. This allows for a larger area

of information to be used. For more details on the ar-

chitectures, see (Havaei et al., 2017) and (Kamnitsas

et al., 2017). An important feature of these architec-

tures is that all the layers of the network are convo-

lutional, enabling dense inference on full images or

image segments.

2.1.1 Training

We interpret the output of the CNNs as predicted la-

bel probabilities for lesion, and deﬁne a training loss

function consisting of the negative log likelihood with

both L1 and L2 regularization. This loss is minimised

by following a stochastic gradient descent approach

on randomly selected minibatches of patches within

each brain.

Performance of CNNs depends greatly on the dis-

tribution of the training samples. A commonly used

approach is to train a classiﬁer on the same number

of image patches from each of the classes per mini-

batch. However since the classes are imbalanced, this

approach biases the classiﬁer towards making false

positive predictions.

In TwoPathCNN, we follow the two-phase train-

ing proposed in [3] with minibatches of one labeled

sample per training instance. In the ﬁrst phase, the

mini-batches are equally divided among lesions and

BIOIMAGING 2020 - 7th International Conference on Bioimaging

118

backround. In the second phase, we keep the weights

of all layers ﬁxed and retrain only the ﬁnal layer on

patches uniformly extracted to be closer to the true

data distribution.

For Wider2dSeg, the patch size fed down each

pathway is larger than the network’s receptive ﬁeld

(Kamnitsas et al., 2017). This technique, called dense

training, increases the effective batch size by con-

structing mini-batches of more than one labeled sam-

ple per training instance since it allows the network

to segment a neighborhood of voxels sorrounding the

central voxel. (Kamnitsas et al., 2017). This makes

the patch size, also called image segment size, an im-

portant parameter to tune since larger patch sizes cap-

ture more background voxels than smaller patch sizes.

2.2 Implementation Details

We implemented the models using Tensorﬂow (Abadi

et al., 2016). We apply only minimal pre-processing:

we normalize within each input channel by subtract-

ing its mean and dividing by its standard deviation.

Unlike (Havaei et al., 2017), we did not use N41TK

bias correction and we did not remove the 1% highest

and lowest intensities. Similarily we didn’t use batch

normalization as proposed in (Kamnitsas et al., 2017)

since it is more of a requirement for 3D architectures.

We use standard momentum and ﬁx the momentum

coefﬁcient µ = 0.6 throughout training for all archi-

tectures.

2.2.1 TwoPathCNN

Weights are initialized from a uniform distribution on

(−0.005, 0.005), as in (Havaei et al., 2017), and bi-

ases are initialized to zero. Every epoch consists of

10,000 iterations of stochastic gradient descent with

momentum on mini-batches of 10 labeled samples.

Each sample consists of 4 stacked patches of size

33 × 33, each patch correspond to a different MRI

modality, and a label, which is the ground truth la-

bel for the central voxel in the patch. The ﬁrst phase

of training consists of 50,000 iterations or 5 epochs.

The minibatches at this stage contain equal number of

positive and negative examples. The learning rate is

set to 0.001 decays by a factor of 0.1 (Havaei et al.,

2017) starting from the third epoch. The second phase

of training consists of another 4 epochs of 10,000 it-

erations each. The minibatches at this stage have the

property that approximately 2% of samples presented

in them are labeled as negative. The learning rate is

reset to 0.001 and decays by a factor of 0.1 after each

epoch. Thus, in total, the model is trained on 900,000

samples. The L

regularization constant is 10

−6

and

the L

regularization constant is 10

−4

. For further

regularization, dropout at a rate of 0.5 was applied on

hidden layers of the local pathway. In all of these im-

plementation details, we follow (Havaei et al., 2017),

except that the regularization constants were inspired

from (Kamnitsas et al., 2017) (which used the same

dataset as we do), but changed to account for the in-

creased number of parameters.

2.2.2 Wider2dSeg

As in (Kamnitsas et al., 2017), we use the weight ini-

tiliazation method of He et al. (He et al., 2015), since

deeper architectures are prone to larger signal vari-

ance. The bias terms are initialized to zero. We used

the RMSProp optimizer for a total of 80,000 iterations

with a learning rate of 0.001 and decayed it by a factor

of 0.5 (Kamnitsas et al., 2017) at the the following it-

erations: 25,000, 39,000, 49,000, 59,000, 71,000, and

75,000. In contrast to (Kamnitsas et al., 2017), we

use an image segment size of 43 for the ﬁrst pathway

and 75 for the second pathway, which segments the

neighborhood around the central voxel per train-

ing instance. Mini-batches are of size 12 and equally

divided amongst lesions and background. The batch

and image segment sizes were chosen to achieve the

same effective batch size shown in Table B.1 of Ap-

pendix B from (Kamnitsas et al., 2017). To regularize

the network we follow (Kamnitsas et al., 2017) and

set the L

constant to 10

−8

, the L

constant to 10

−6

and apply dropout at a rate of 0.5 on the last two hid-

den layers.

2.3 Symmetry-augmented Methods

(LSymm and NLSymm)

For each subject, we augment the four image modali-

ties by four “mirror” images produced as follows. We

begin by producing a reﬂected FLAIR image by re-

versing the orientation of the x (left-right) axis, using

the fslorient tool of FSL (Smith et al., 2001). Since

the original images are linearly co-registered, this step

is approximately a reﬂection in the median, i.e. mid-

sagittal, plane. (FLAIR was chosen due to its frequent

use in lesion segmentation, however we intend in later

work to compare the use of T1 or multiple modali-

ties in this step.) We align the result with the original

FLAIR image using either “linear”/afﬁne (LSymm) or

nonlinear (NLSymm) registration. This step uses the

SynQuick method in the ANTs package (Avants et al.,

2009). For LSymm, the “-t a” option was used, giv-

ing a 2-stage rigid+afﬁne registration. For NLSymm,

the default options were used, giving a 3-stage rigid +

afﬁne + nonlinear (“SyN”) registration. In either case,

the resulting transformation, composed with a reﬂec-

Exploiting Bilateral Symmetry in Brain Lesion Segmentation with Reﬂective Registration

119

Figure 3: An example of different segmentations of Brain #8 of the SISS2015 Training set from the ISLES2015 Chal-

lenge. From left to right, the columns show segmentations produced by the TwoPathCNN, TwoPathCNN+LSymm and

TwoPathCNN+NLSymm methods, and the ground truth.

Table 1: Performance of TwoPathCNN and Wider2dSeg based on a 7-fold cross-validation for baseline, NLSymm and LSymm

on the ISLES2015 (SISS) training data. Results for Dice, Recall and Precision (Havaei et al., 2017) are shown as mean (std.

dev.).

Architecture Dice Recall Precision

TwoPathCNN 0.45(0.25) 0.59(0.22) 0.45(0.29)

TwoPathCNN+LSymm 0.52(0.23) 0.63(0.23) 0.50(0.28)

TwoPathCNN+NLSymm 0.54(0.21) 0.65(0.22) 0.52(0.26)

Wider2dSeg 0.49(0.25) 0.53(0.28) 0.54(0.25)

Wider2dSeg+LSymm 0.61(0.22) 0.58(0.25) 0.67(0.22)

Wider2dSeg+NLSymm 0.62(0.22) 0.60(0.25) 0.68(0.21)

tion, produces a symmetry transformation T (x, y, z)

that associates to each voxel its corresponding “mir-

ror” voxel in the opposite hemisphere.

Once we have obtained our linear or nonlinear

symmetry transformation for each subject, we use it

to construct a Symmetry Difference Image for each

modality, by subtracting from each voxel’s standard-

ized intensity the standardized intensity of the “mir-

ror” voxel, S

(x, y, z) = I

(x, y, z) − I

(T (x, y, z)).

This results in 4 Symmetry Difference Images

(SDIs), one for each modality, which we use to aug-

ment the original 4 images. For instance in the base-

line TwoPathCNN model, for each voxel, one 33×33

patch is extracted from each of the 4 MR images and

combined into a 4 × 33 × 33 tensor; in LSymm and

NLSymm, one 33 × 33 patch is extracted from each

of the 8 MR images (originals plus SDIs) and com-

bined into a 8 × 33 × 33 tensor. This double-size ten-

sor is fed into the both the local and global pathways

of the TwoPathCNN architecture. Apart from dou-

bling the number of images from 4 to 8, all details of

architecture and training are exactly as in the baseline

methods.

3 EXPERIMENTAL RESULTS

3.1 Dataset

We evaluated our methods on the ISLES2015 (SISS)

training data. The training data consists of FLAIR,

DWI ,T1 and T1-contrast images of size 230 × 230 ×

154, for each of 28 patients with sub-acute ischemic

stroke lesions. All images are skull-stripped and have

isotropic 1mm

voxel resolution.

BIOIMAGING 2020 - 7th International Conference on Bioimaging

120

3.2 Experiment

For each of the two architectures, we compare the

three methods described above: baseline, baseline

with LSymm, and baseline with NLSymm, using 7-

fold cross-validation on the Training dataset of 28

subjects. All methods are run with the same hyper-

parameters, on the same pseudo-random sequence of

training patches.

3.3 Results

Example segmentations produced by the three meth-

ods on TwoPathCNN are shown in Fig. 3. The main

results are summarised in Table 1. Adding linearly or

nonlinearly registered symmetry features (LSymm or

NLSymm) to the baseline architectures consistently

improves mean Dice coefﬁcient, Recall and Precision,

showing the effectiveness of reﬂective registration.

For the Dice coefﬁcient, we performed one-sided

paired t-tests for symmetry-augmented vs. baseline

methods, and found that the resulting p-values were

always less than 10

−5

, for both LSymm and NLSymm

and for both baseline architectures. Moreover, nonlin-

early registered symmetry features (NLSymm) con-

sistently produced higher Dice, Recall and Precision

scores compared to linearly registered symmetry fea-

tures (p < 0.001 for the TwoPathCNN architecture,

and p = 0.08 for the Wider2dSeg architecture). Of

the two architectures evaluated, Wider2dSeg beneﬁt-

ted more from the symmetry augmentation, however

the difference between LSymm and NLSymm was

not signiﬁcant; both differences were perhaps due to

its deeper architecture.

4 CONCLUSIONS

We have proposed an improvement to existing seg-

mentation methods by exploiting the bilateral quasi-

symmetry of healthy brains. Our method, which does

not require a template, consists of augmenting the in-

put images to a CNN with extra Symmetry Difference

Images, which are intensity differences between ho-

mologous (“mirror”) voxels in different hemispheres.

We showed how to incorporate these symmetric fea-

tures into the increasingly popular patch-based CNNs

so as to preserve dense inference. In an experi-

ment on the ISLES2015 SISS dataset, we found that

adding symmetric features generated using nonlinear

reﬂective registration (the “NLSymm” method) con-

sistently resulted in a mean improvement in Dice co-

efﬁcient, Recall and Precision. Using linear reﬂec-

tive registration instead gave consistently smaller im-

provements over baseline, showing that nonlinear reg-

istration is superior in this application. For the Dice

coefﬁcient, improvement over baseline was signiﬁ-

cant (p < 10

−5

) for both linear and nonlinear sym-

metric features. The nonlinear method was signiﬁ-

cantly better than the linear one (p < 0.001) for one

baseline architecture (TwoPathCNN) but not the other

(Wider2dSeg).

While our numerical results are not directly com-

parable with those of the three preceding studies of

symmetric feature augmentation for CNNs mentioned

in the Introduction, we note that our improvements in

Dice scores of 9 to 13% on an open dataset compare

favourably to earlier results.

We have shown that the brain’s quasi-symmetry

property is a valuable tool for brain lesion segmenta-

tion. The ease of application of symmetry augmen-

tation to most existing CNN methods and many other

methods suggests a potentially wide-ranging utility of

the method.

REFERENCES

Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean,

J., Devin, M., Ghemawat, S., Irving, G., Isard, M.,

et al. (2016). Tensorﬂow: a system for large-scale

machine learning. In OSDI, volume 16, pages 265–

283.

Avants, B. B., Tustison, N. J., Song, G., and Gee, J. C.

(2009). ANTS: Open-source tools for normalization

and neuroanatomy. Transac Med Imagins Penn Image

Comput Sci Lab.

Dvorak, P., Bartusek, K., and Kropatsch, W. (2013). Au-

tomated segmentation of brain tumour edema in ﬂair

mri using symmetry and thresholding. PIERS Pro-

ceedings, Stockholm, Sweden.

Havaei, M., Davy, A., Warde-Farley, D., Biard, A.,

Courville, A., Bengio, Y., Pal, C., Jodoin, P.-M., and

Larochelle, H. (2017). Brain tumor segmentation

with deep neural networks. Medical image analysis,

35:18–31.

He, K., Zhang, X., Ren, S., and Sun, J. (2015). Delv-

ing deep into rectiﬁers: Surpassing human-level per-

formance on imagenet classiﬁcation. In Proceedings

of the IEEE international conference on computer vi-

sion, pages 1026–1034.

Kamnitsas, K., Ledig, C., Newcombe, V. F., Simpson,

J. P., Kane, A. D., Menon, D. K., Rueckert, D., and

Glocker, B. (2017). Efﬁcient multi-scale 3D CNN

with fully connected CRF for accurate brain lesion

segmentation. Medical image analysis, 36:61–78.

Loy, G. and Eklundh, J.-O. (2006). Detecting symmetry

and symmetric constellations of features. In Euro-

pean Conference on Computer Vision, pages 508–521.

Springer.

Maier, O., Menze, B. H., von der Gablentz, J., H

ani, L.,

Heinrich, M. P., Liebrand, M., Winzeck, S., Basit, A.,

Exploiting Bilateral Symmetry in Brain Lesion Segmentation with Reﬂective Registration

121

Bentley, P., Chen, L., et al. (2017). ISLES 2015 - a

public evaluation benchmark for ischemic stroke le-

sion segmentation from multispectral MRI. Medical

image analysis, 35:250–269.

Meier, R., Bauer, S., Slotboom, J., Wiest, R., and Reyes, M.

(2014). Appearance-and context-sensitive features for

brain tumor segmentation. Proceedings of MICCAI

BRATS Challenge, pages 020–026.

Schmidt, M., Levner, I., Greiner, R., Murtha, A., and

Bistritz, A. (2005). Segmenting brain tumors us-

ing alignment-based features. In Fourth International

Conference on Machine Learning and Applications

(ICMLA’05), pages 6–pp. IEEE.

Shen, H., Zhang, J., and Zheng, W. (2017). Efﬁcient

symmetry-driven fully convolutional network for mul-

timodal brain tumor segmentation. In 2017 IEEE In-

ternational Conference on Image Processing (ICIP),

pages 3864–3868. IEEE.

Smith, S., Bannister, P. R., Beckmann, C., Brady, M., Clare,

S., Flitney, D., Hansen, P., Jenkinson, M., Leibovici,

D., Ripley, B., et al. (2001). FSL: New tools for func-

tional and structural brain image analysis. NeuroIm-

age, 13(6):249.

Tustison, N. J., Shrinidhi, K., Wintermark, M., Durst, C. R.,

Kandel, B. M., Gee, J. C., Grossman, M. C., and

Avants, B. B. (2015). Optimal symmetric multimodal

templates and concatenated random forests for su-

pervised brain tumor segmentation (simpliﬁed) with

antsr. Neuroinformatics, 13(2):209–225.

Wang, Y., Katsaggelos, A. K., Wang, X., and Parrish, T. B.

(2016). A deep symmetry convnet for stroke lesion

segmentation. In 2016 IEEE International Conference

on Image Processing (ICIP), pages 111–115. IEEE.

Winzeck, S., Hakim, A., McKinley, R., Pinto, J. A., Alves,

V., Silva, C., Pisov, M., Krivov, E., Belyaev, M.,

Monteiro, M., et al. (2018). ISLES 2016 and 2017-

benchmarking ischemic stroke lesion outcome predic-

tion based on multispectral MRI. Frontiers in neurol-

ogy, 9.

Zhang, H., Zhu, X., and Willke, T. L. (2017). Seg-

menting brain tumors with symmetry. arXiv preprint

arXiv:1711.06636. NIPS ML4H Workshop 2017.

BIOIMAGING 2020 - 7th International Conference on Bioimaging

122