Two-Stage Cascaded U-Net: 1st Place

Solution to BraTS Challenge 2019

Segmentation Task

Zeyu Jiang1 , Changxing Ding1(B) , Minfeng Liu2 , and Dacheng Tao3

1
School of Electronic and Information Engineering,
South China University of Technology, Guangzhou, China
[email protected]
2
Nanfang Hospital, Southern Medical University,
Guangzhou 510515, Guangdong, China
3
UBTECH Sydney AI Centre, SIT, FEIT, University of Sydney, Sydney, Australia

Abstract. In this paper, we devise a novel two-stage cascaded U-Net

to segment the substructures of brain tumors from coarse to fine. The
network is trained end-to-end on the Multimodal Brain Tumor Segmen-
tation Challenge (BraTS) 2019 training dataset. Experimental results on
the testing set demonstrate that the proposed method achieved average
Dice scores of 0.83267, 0.88796 and 0.83697, as well as Hausdorff dis-
tances (95%) of 2.65056, 4.61809 and 4.13071, for the enhancing tumor,
whole tumor and tumor core, respectively. The approach won the 1st
place in the BraTS 2019 challenge segmentation task, with more than 70
teams participating in the challenge.

Keywords: Deep learning · Brain tumor segmentation · U-Net

1 Introduction
Gliomas are the most common type of primary brain tumors. Automatic three-
dimensional brain tumor segmentation can save doctors time and provide an
appropriate method of additional tumor analysis and monitoring. Recently, deep
learning approaches have consistently outperformed traditional brain tumor seg-
mentation methods [6,10,17,20,24,27].
The multimodal brain tumor segmentation challenge (BraTS) is aimed at
evaluating state-of-the-art methods for the segmentation of brain tumors [1–
4,13]. The BraTS 2019 training dataset, which comprises 259 cases of high-grade
gliomas (HGG) and 76 cases of low-grade gliomas (LGG), is manually annotated
by both clinicians and board-certified radiologists. For each patient, a native pre-
contrast (T1), a post-contrast T1-weighted (T1Gd), a T2-weighted (T2) and a
T2 Fluid Attenuated Inversion Recovery (T2-FLAIR) are provided. An example
image set is presented in Fig. 1. Each tumor is segmented into enhancing tumor,
the peritumoral edema, and the necrotic and non-enhancing tumor core. A num-
ber of metrics (Dice score, Hausdorff distance (95%), sensitivity and specificity)
c Springer Nature Switzerland AG 2020
A. Crimi and S. Bakas (Eds.): BrainLes 2019, LNCS 11992, pp. 231–241, 2020.
https://doi.org/10.1007/978-3-030-46640-4_22
232 Z. Jiang et al.

(a) T2 Flair (b) T1 weighted

(c) T1 post-contrast (d) T2 weighted

Fig. 1. Example of image modalities in the BraTS 2019 dataset.

are used to measure the segmentation performance of the algorithms proposed

by participants.
In BraTS 2017, Kamnitsas et al. [9], who was the first-place winner of the
challenge, proposed Ensembles of Multiple Models and Architectures (EMMA)
for robust segmentation, which was achieved by combining several network archi-
tectures including DeepMedic [10], 3D U-Net [18] and 3D FCN [12]. These
networks were trained with different optimization processes via diverse loss
functions such as Dice loss [14] and cross-entropy loss. In BraTS 2018, Myro-
nenko [15], who achieved the best performance on the testing dataset, utilized
an asymmetrical U-Net with a larger encoder to extract image features, along
with a smaller decoder to reconstruct the label. He fed a very large patch size
(160 × 192 × 128 voxels) into the network, and also added a variational autoen-
coder (VAE) branch in order to regularize the shared encoder.
In this work, inspired by the cascaded strategy [19,22,25,26], we propose a
novel two-stage cascaded U-Net. In the first stage, we use a variant of U-Net
 Two-Stage Cascaded U-Net: 1st Place Solution to BraTS Challenge 2019 233

as the first stage network to train a coarse prediction. In the second stage, we
increase the width of the network and use two decoders so as to boost perfor-
mance. The second stage is added to refine the prediction map by concatenating
a preliminary prediction map with the original input to utilize auto-context. We
do not use any additional training data and only participate in the segmentation
task in testing phase.

2 Methods
Myronenko [15] proposed an asymmetrical U-Net with a variational autoencoder
branch [5,11]. In this paper, we take a variant of this approach as the basic
segmentation architecture. We further propose a two-stage cascaded U-Net. The
details are illustrated as follows.

Fig. 2. Overview of the two-stage cascaded network.

2.1 Model Cascade

As can be seen in Fig. 2, in the first stage, multi-modal magnetic resonance
images (4 × 128 × 128 × 128) are passed into the first stage U-Net and predict a
segmentation map roughly. The coarse segmentation map is fed together with
the raw images into the second stage U-net. The second stage can provide a
more accurate segmentation map with more network parameters. The two-stage
cascaded network is trained in an end-to-end fashion.
234 Z. Jiang et al.

X2 X2 X2
X2
X2

X2

Fig. 3. The first stage network architecture.

2.2 The First Stage Network Architecture

Due to GPU memory limitations, our networks is designed to take input patches
of size 128 × 128 × 128 voxels and to use a batch size of one. The network archi-
tecture consists of a larger encoding path, to extract complex semantic features,
and a smaller decoding path, to recover a segmentation map with the same input
size. The architecture of the first stage network is presented in Fig. 3.
The 3D U-Net has an encoder and a decoder path, each of which have four
spatial levels. At the beginning of the encoder, patches of size 128 × 128 × 128
voxels with four channels are extracted from the brain tumor images as input,
followed by an initial 3 × 3 × 3 3D convolution with 16 filters. We also use a
dropout with a rate of 0.2 after the initial encoder convolution. The encoder
part uses a pre-activated residual block [7,8]. Each of these blocks consists of
two 3 × 3 × 3 convolutions with Group Normalization [23] with group size of
8 and Rectified Linear Unit (ReLU) activation, followed by additive identity
skip connection. The number of pre-activated residual blocks is 1, 2, 2, and 4
within each spatial level. Moreover, a convolution layer with a 3 × 3 × 3 filter
and a stride of 2 is used to reduce the resolution of the feature maps by 2 and
simultaneously increase the number of feature channels by 2.
Unlike the encoder, the decoder structure uses a single pre-activated residual
block for each spatial level. Before up-sampling, we use 1 × 1 × 1 convolutions
to reduce the number of features by a factor of 2. Compared with [15], we use
a deconvolution with kernel size 2 × 2 × 2 and a stride of 2 rather than trilin-
ear interpolation in order to double the size of the spatial dimension. The net-
work features shortcut connections between corresponding layers with the same
 Two-Stage Cascaded U-Net: 1st Place Solution to BraTS Challenge 2019 235

resolution in the encoder and decoder by elementwise summation. At the end

of the decoder, a 1 × 1 × 1 convolution is used to decrease the number of output
channels to three, followed by a sigmoid function. The detail of the structure is
shown in Table 1.

2.3 The Second Stage Network Architecture

Different from the network in the first stage, we double the number of filters
in the initial 3D convolution in order to increase the network width. What’s
more, we use two decoders. The structure of the two decoders is the same except
that one uses a deconvolution and the other uses trilinear interpolation. The
interpolation decoder is used only during training. Because the performance of
the decoder used deconvolution is better than used trilinear interpolation and
add a decoder used trilinear interpolation to regularize the shared encoder can
improve the performance in our experiment. The architecture of the second stage
network is presented in Fig. 4 and the detail of the structure is shown in Table 2.

X2
十

X2 X2

X2 X2
X2
十
X2
X2
X2
十

Fig. 4. The second stage network architecture.

2.4 Loss
The Dice Similarity Coefficient measures (DSC) the degree of overlap between
the prediction map and ground truth. The DSC is calculated by Eq. 1, where S
is the output of network, R is the ground truth label and | · | denotes the volume
of the region.
2|S ∩ R|
DSC = (1)
|S| + |R|
236 Z. Jiang et al.

The soft Dice loss is designed as following:


2∗ S∗R
Ldice =  2  2 (2)
S + R +
Instead of learning the labels (e.g. enhancing tumor, edema, necrosis and non-
enhancing), we directly optimize the three overlapping regions (whole tumor,
tumor core and enhancing tumor) with the Dice loss, then simply add the Dice
loss functions of each region together. We also add the loss of each stage together
to arrive at the final loss.

Table 1. The first stage network structure, where + stands for additive identity skip
connection, Conv3 - 3 × 3 × 3 convolution, Conv1 - 1 × 1 × 1 convolution, GN - group
normalization, ConvTranspose - deconvolution with kernel size 2 × 2 × 2.

U-Net 1
Name Details Repeat Size
Input 4 × 128 × 128 × 128
InitConv Conv3, Dropout 1 16 × 128 × 128 × 128
EnBlock1 GN, ReLU, Conv3, GN, ReLU, Conv3, + 1 16 × 128 × 128 × 128
Encoder EnDown1 Conv3 stride 2 1 32 × 64 × 64 × 64
EnBlock2 GN, ReLU, Conv3, GN, ReLU, Conv3, + 2 32 × 64 × 64 × 64
EnDown2 Conv3 stride 2 1 64 × 32 × 32 × 32
EnBlock3 GN, ReLU, Conv3, GN, ReLU, Conv3, + 2 64 × 32 × 32 × 32
EnDown3 Conv3 stride 2 1 128 × 16 × 16 × 16
EnBlock4 GN, ReLU, Conv3, GN, ReLU, Conv3, + 4 128 × 16 × 16 × 16
DeUp3 Conv1, ConvTranspose, +EnBlock3 1 64 × 32 × 32 × 32
DeBlock3 GN, ReLU, Conv3, GN, ReLU, Conv3, + 1 64 × 32 × 32 × 32
DeUp2 Conv1,ConvTranspose,+EnBlock2 1 32 × 64 × 64 × 64
Decoder DeBlock2 GN, ReLU, Conv3, GN, ReLU, Conv3, + 1 32 × 64 × 64 × 64
DeUp2 Conv1, ConvTranspose, +EnBlock1 1 16 × 128 × 128 × 128
DeBlock1 GN, ReLU, Conv3, GN, ReLU, Conv3, + 1 16 × 128 × 128 × 128
EndConv Conv1 1 3 × 128 × 128 × 128
Sigmoid Sigmoid 1 3 × 128 × 128 × 128

3 Experiments
3.1 Data Pre-processing and Augmentation
Before feeding the data into the deep learning network, a preprocessing method
is used to process the input data. Since the MRI intensity values are non-
standardized, we apply intensity normalization to each MRI modality from each
patient independently by subtracting the mean and dividing by the standard
deviation of the brain region only.
Moreover, to prevent an overfitting issue from arising, we deploy three
types of data augmentation. Firstly, we apply a random intensity shift between
 Two-Stage Cascaded U-Net: 1st Place Solution to BraTS Challenge 2019 237

Table 2. The second stage network structure, where + stands for additive identity skip
connection, Conv3 - 3 × 3 × 3 convolution, Conv1 - 1 × 1 × 1 convolution, GN - group
normalization, ConvTranspose - deconvolution with kernel size 2 × 2 × 2, Upsampling
- trilinear interpolation, Decoder2 is used only during training.

U-Net 2
Name Details Repeat Size
Input 7 × 128 × 128 × 128
InitConv Conv3, Dropout 1 32 × 128 × 128 × 128
EnBlock1 GN, ReLU, Conv3, GN, ReLU, Conv3, + 1 32 × 128 × 128 × 128
Encoder EnDown1 Conv3 stride 2 1 64 × 64 × 64 × 64
EnBlock2 GN, ReLU, Conv3, GN, ReLU, Conv3, + 2 64 × 64 × 64 × 64
EnDown2 Conv3 stride 2 1 128 × 32 × 32 × 32
EnBlock3 GN, ReLU, Conv3, GN, ReLU, Conv3, + 2 128 × 32 × 32 × 32
EnDown3 Conv3 stride 2 1 256 × 16 × 16 × 16
EnBlock4 GN, ReLU, Conv3, GN, ReLU, Conv3, + 4 256 × 16 × 16 × 16
DeUp3 Conv1, ConTranspose, +EnBlock3 1 128 × 32 × 32 × 32
DeBlock3 GN, ReLU, Conv3, GN, ReLU, Conv3, + 1 128 × 32 × 32 × 32
DeUp2 Conv1, ConTranspose, +EnBlock2 1 64 × 64 × 64 × 64
Decoder1 DeBlock2 GN, ReLU, Conv3, GN, ReLU, Conv3, + 1 64 × 64 × 64 × 64
DeUp2 Conv1, ConTranspose, +EnBlock1 1 32 × 128 × 128 × 128
DeBlock1 GN, ReLU, Conv3, GN, ReLU, Conv3, + 1 32 × 128 × 128 × 128
EndConv Conv1 1 3 × 128 × 128 × 128
Sigmoid Sigmoid 1 3 × 128 × 128 × 128
DeUp3 1 Conv1, Upsampling, +EnBlock3 1 128 × 32 × 32 × 32
DeBlock3 1 GN, ReLU, Conv3, GN, ReLU, Conv3, + 1 128 × 32 × 32 × 32
DeUp2 1 Conv1, Upsampling, +EnBlock2 1 64 × 64 × 64 × 64
Decoder2 DeBlock2 1 GN, ReLU, Conv3, GN, ReLU, Conv3, + 1 64 × 64 × 64 × 64
(Used only
during
training)
DeUp2 1 Conv1, Upsampling, +EnBlock1 1 32 × 128 × 128 × 128
DeBlock1 1 GN, ReLU, Conv3, GN, ReLU, Conv3, + 1 32 × 128 × 128 × 128
EndConv 1 Conv1 1 3 × 128 × 128 × 128
Sigmoid 1 Sigmoid 1 3 × 128 × 128 × 128

[−0.1−0.1] of the standard deviation of each channel, as well as a random scaling

intensity of the input between scales [0.9−1.1]. Secondly, we train our network by
randomly cropping the MRI data from 240 × 240 × 155 voxels to 128 × 128 × 128
voxels due to memory limitation. Finally, we use random flipping along each 3D
axis with a probability of 50%.

3.2 Training Details

The implementation of our network is based on PyTorch 1.1.0 [16]. The maxi-
mum number of training iterations is set to 405 epochs with 5 epochs of linear
warmup. We use Adam optimizer to update the weights of the network, with a
batch size of 1 and an initial learning rate of α0 = 1e − 4 at the very beginning
and decays it as following:
 0.9
e
α = α0 × 1 − (3)
Ne
238 Z. Jiang et al.

where e is an epoch counter, and Ne is a total number of epochs. We regu-

larize using an l2 weight decay of 1e − 5. Training is performed on a Nvidia
Titan V GPU with 12 Gb memory. However, our method requires slightly more
than 12 Gb memory in our experiment. We utilize gradient checkpointing [21]
by PyTorch to reduce the memory consumption.

3.3 Augmentation for Inference

At testing time, we segment the whole brain region at once instead of using a
sliding window. The interpolation decoder is not used during the inference phase.
To obtain a more robust prediction, we preserve eight weights of the model in
the last time of the training progress for prediction. For each snapshot, the input
images are used different flipping before being fed into the network. Finally, we
average the output of the resulting eight segmentation probability maps.

3.4 Post-processing

We replace enhancing tumor with necrosis when the volume of predicted enhanc-
ing tumor is less than the threshold to post-process our segmentation results
(The threshold is chosen for each experiment independently, depending on the
performance of BraTS 2019 validation dataset).

4 Results
The variability of a single model can be quite high. We use total five net-
works from the 5-fold cross-validation as an ensemble to predict segmentation for
BraTS 2019 validation dataset. Also, we use an ensemble of a set of 12 models,
which are trained from scratch using the entire training dataset. The best single
model is chosen from the set of 12 models.
We report the results of our approach on the BraTS 2019 validation dataset,
which contains 125 cases with unknown glioma grade and unknown segmen-
tation. All reported values are computed via the online evaluation platform
(https://ipp.cbica.upenn.edu/) for evaluation of Dice score, sensitivity, speci-
ficity and Hausdorff distance (95%). Validation set results can be found in
Table 3. The performance of the best single model is slightly better than ensem-
ble of 5-fold cross-validation. The ensemble of 12 models results in a minor
improvement compared with the best single model.
Testing set results are presented in Table 4. Our algorithm achieved the first
place out of more than 70 participating teams.
 Two-Stage Cascaded U-Net: 1st Place Solution to BraTS Challenge 2019 239

Table 3. Mean Dice and Hausdorff measurements of the proposed segmentation

method on BraTS 2019 validation set. DSC - dice similarity coefficient, HD95 - Haus-
dorff distance (95%), WT - whole tumor, TC - tumor core, ET - enhancing tumor
core.

Method DSC HD95

Validation WT TC ET WT TC ET
Ensemble of 5-fold 0.90797 0.85888 0.79667 4.35413 5.69195 3.12642
Best single model 0.90819 0.86321 0.80199 4.44375 5.86201 3.20551
Ensemble of 12 models 0.90941 0.86473 0.80211 4.26398 5.43931 3.14581

Table 4. Mean Dice and Hausdorff measurements of the proposed segmentation

method on BraTS 2019 testing set. DSC - dice similarity coefficient, HD95 - Haus-
dorff distance (95%), WT - whole tumor, TC - tumor core, ET - enhancing tumor
core.

Method DSC HD95

Testing WT TC ET WT TC ET
Ensemble of 12 models 0.88796 0.83697 0.83267 4.61809 4.13071 2.65056

5 Conclusion
In this paper, we propose a two-stage cascaded U-Net. Our approach refines the
prediction through a progressive cascaded network. Experiments on the BraTS
2019 validation set demonstrate that our method can obtain very competitive
segmentation even though using single model. The testing results show that our
proposed method can achieve excellent performance, winning the first position
in the BraTS 2019 challenge segmentation task among 70+ participating teams.

Acknowledgements. Changxing Ding was supported in part by the National Natural

Science Foundation of China (Grant No.: 61702193), Science and Technology Program
of Guangzhou (Grant No.: 201804010272), and the Program for Guangdong Introducing
Innovative and Entrepreneurial Teams (Grant No.: 2017ZT07X183). Dacheng Tao was
supported by Australian Research Council Projects (FL-170100117, DP-180103424 and
LP-150100671).

References
1. Bakas, S., et al.: Segmentation labels and radiomic features for the pre-operative
scans of the TCGA-GBM collection. Cancer Imaging Archive (2017). https://doi.
org/10.7937/K9/TCIA.2017.KLXWJJ1Q
2. Bakas, S., et al.: Segmentation labels and radiomic features for the pre-operative
scans of the TCGA-LGG collection. Cancer Imaging Archive 286 (2017)
3. Bakas, S., et al.: Advancing the cancer genome atlas glioma MRI collections with
expert segmentation labels and radiomic features. Sci. Data 4, 170117 (2017).
https://doi.org/10.1038/sdata.2017.117
240 Z. Jiang et al.

4. Bakas, S., et al.: Identifying the best machine learning algorithms for brain tumor
segmentation, progression assessment, and overall survival prediction in the brats
challenge. arXiv preprint arXiv:1811.02629 (2018)
5. Doersch, C.: Tutorial on variational autoencoders. arXiv preprint arXiv:1606.05908
(2016)
6. Havaei, M., et al.: Brain tumor segmentation with deep neural networks. Med.
Image Anal. 35, 18–31 (2017)
7. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In:
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
pp. 770–778 (2016)
8. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In:
Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) ECCV 2016. LNCS, vol. 9908, pp.
630–645. Springer, Cham (2016). https://doi.org/10.1007/978-3-319-46493-0 38
9. Kamnitsas, K., et al.: Ensembles of multiple models and architectures for robust
brain tumour segmentation. In: Crimi, A., Bakas, S., Kuijf, H., Menze, B., Reyes,
M. (eds.) BrainLes 2017. LNCS, vol. 10670, pp. 450–462. Springer, Cham (2018).
https://doi.org/10.1007/978-3-319-75238-9 38
10. Kamnitsas, K., et al.: Efficient multi-scale 3D CNN with fully connected CRF for
accurate brain lesion segmentation. Med. Image Anal. 36, 61–78 (2017)
11. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint
arXiv:1312.6114 (2013)
12. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic
segmentation. In: Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, pp. 3431–3440 (2015)
13. Menze, B.H., et al.: The multimodal brain tumor image segmentation benchmark
(BRATS). IEEE Trans. Med. Imaging 34(10), 1993–2024 (2015). https://doi.org/
10.1109/tmi.2014.2377694
14. Milletari, F., Navab, N., Ahmadi, S.A.: V-net: fully convolutional neural networks
for volumetric medical image segmentation. In: 2016 Fourth International Confer-
ence on 3D Vision (3DV), pp. 565–571. IEEE (2016)
15. Myronenko, A.: 3D MRI brain tumor segmentation using autoencoder regular-
ization. In: Crimi, A., Bakas, S., Kuijf, H., Keyvan, F., Reyes, M., van Walsum,
T. (eds.) BrainLes 2018. LNCS, vol. 11384, pp. 311–320. Springer, Cham (2019).
https://doi.org/10.1007/978-3-030-11726-9 28
16. Paszke, A., et al.: Automatic differentiation in PyTorch (2017)
17. Pereira, S., Pinto, A., Alves, V., Silva, C.A.: Brain tumor segmentation using
convolutional neural networks in mri images. IEEE Trans. Med. Imaging 35(5),
1240–1251 (2016)
18. Ronneberger, O., Fischer, P., Brox, T.: U-Net: convolutional networks for biomed-
ical image segmentation. In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F.
(eds.) MICCAI 2015. LNCS, vol. 9351, pp. 234–241. Springer, Cham (2015).
https://doi.org/10.1007/978-3-319-24574-4 28
19. Roth, H.R., et al.: A multi-scale pyramid of 3D fully convolutional networks for
abdominal multi-organ segmentation. In: Frangi, A.F., Schnabel, J.A., Davatzikos,
C., Alberola-López, C., Fichtinger, G. (eds.) MICCAI 2018. LNCS, vol. 11073, pp.
417–425. Springer, Cham (2018). https://doi.org/10.1007/978-3-030-00937-3 48
20. Shen, H., Wang, R., Zhang, J., McKenna, S.J.: Boundary-aware fully convolu-
tional network for brain tumor segmentation. In: Descoteaux, M., Maier-Hein, L.,
Franz, A., Jannin, P., Collins, D.L., Duchesne, S. (eds.) MICCAI 2017. LNCS, vol.
10434, pp. 433–441. Springer, Cham (2017). https://doi.org/10.1007/978-3-319-
66185-8 49
 Two-Stage Cascaded U-Net: 1st Place Solution to BraTS Challenge 2019 241

21. Siskind, J.M., Pearlmutter, B.A.: Divide-and-conquer checkpointing for arbitrary

programs with no user annotation. Optim. Methods Softw. 33(4–6), 1288–1330
(2018)
22. Tu, Z., Bai, X.: Auto-context and its application to high-level vision tasks and
3D brain image segmentation. IEEE Trans. Pattern Anal. Mach. Intell. 32(10),
1744–1757 (2009)
23. Wu, Y., He, K.: Group normalization. In: Ferrari, V., Hebert, M., Sminchisescu, C.,
Weiss, Y. (eds.) ECCV 2018. LNCS, vol. 11217, pp. 3–19. Springer, Cham (2018).
https://doi.org/10.1007/978-3-030-01261-8 1
24. Zhao, X., Wu, Y., Song, G., Li, Z., Zhang, Y., Fan, Y.: A deep learning model
integrating FCNNS and CRFS for brain tumor segmentation. Med. Image Anal.
43, 98–111 (2018)
25. Zhou, C., Chen, S., Ding, C., Tao, D.: Learning contextual and attentive informa-
tion for brain tumor segmentation. In: Crimi, A., Bakas, S., Kuijf, H., Keyvan, F.,
Reyes, M., van Walsum, T. (eds.) BrainLes 2018. LNCS, vol. 11384, pp. 497–507.
Springer, Cham (2019). https://doi.org/10.1007/978-3-030-11726-9 44
26. Zhou, C., Ding, C., Lu, Z., Wang, X., Tao, D.: One-pass multi-task convolutional
neural networks for efficient brain tumor segmentation. In: Frangi, A.F., Schn-
abel, J.A., Davatzikos, C., Alberola-López, C., Fichtinger, G. (eds.) MICCAI 2018.
LNCS, vol. 11072, pp. 637–645. Springer, Cham (2018). https://doi.org/10.1007/
978-3-030-00931-1 73
27. Zhou, C., Ding, C., Wang, X., Lu, Z., Tao, D.: One-pass multi-task networks
with cross-task guided attention for brain tumor segmentation. arXiv preprint
arXiv:1906.01796 (2019)