You are on page 1of 7

EncapNet-3D and U-EncapNet

for Cell Segmentation


Takumi Sato Kazuhiro Hotta
Department of Electrical and Electronic Engineering Department of Electrical and Electronic Engineering
Meijo University Meijo University
1-501 Shiogamaguchi, Tempaku-ku, Nagoya 468-8502, Japan 1-501 Shiogamaguchi, Tempaku-ku, Nagoya 468-8502, Japan
193427015@ccalumni.meijo-u.ac.jp kazuhotta@meijo-u.ac.jp

Abstract—EncapNet is a kind of Capsule network that accuracy on binary segmentation tasks but not to multi
significantly improved routing problems that has thought to be segmentation task. We propose U-EncapNet which is based
the main bottleneck of capsule network. In this paper, we on EncapNet that has achieved high accuracy on multi-classes
propose EncapNet-3D that has stronger connection between classification such as ImageNet-1k. U-EncapNet uses U-net
master and aide branch, which original EncapNet has only architecture which has achieved high performance in semantic
single co-efficient per capsule. We achieved this by adding 3D segmentation. We also used dropout for reducing overfitting.
convolution and Dropout layers to connection between them. 3D
convolution makes connection between capsules stronger. We We consider that capsule is very efficient way to express
also propose U-EncapNet, which uses U-net architecture to feature map. Before EncapNet was introduced, routing-by-
achieve high accuracy in semantic segmentation task. agreement had very huge parameters and efficiency was
EncapNet-3D has successfully accomplished to reduce network hidden behind it. Using approximate routing, we can now
parameters 321 times smaller compared to U-EncapNet, 52 reduce network parameters, with very small decrease of
times smaller than U-net. We show the result on segmentation accuracy. Recently, network architecture which can produce
problem of cell images. U-EncapNet has advanced performance high accuracy has been proposed such as ENET[14],
of 1.1% in cell mean IoU in comparison with U-net. EncapNet- MobileNet[15] and ShuffleNet[16]. Proving that Capsule
3D has achieved 3% increase in comparison with ResNet-6 in Network can achieve high accuracy in low parameters will
cell membrane IoU.
expand the possibilities of Capsule Network which only few
Keywords—Capsule Network, EncapNet, Cell Segmentation
paper [17] has mentioned about it.
We also propose EncapNet-3D, which has stronger
I. INTRODUCTION connection between master and aide branch compared to
Convolutional neural network (CNN) has achieved high EncapNet. EncapNet uses group convolution to separate each
accuracy in many tasks. But CNN has some problems such as capsule. EncapNet does not have enough connection between
the lack of location information by connecting many capsules since most of them are divided by groups. EncapNet-
convolution layers. Capsule Network [1] has been introduced 3D used 3D convolution on co-efficient between master and
to solve the problem. Capsule Network has two key features; aide branch to increase connection between capsule. This
capsule and routing-by-agreement. Capsule is a group of makes EncapNet-3D perform very close accuracy with only
feature represented as a vector instead of scalar. A single one block. EncapNet-3D has achieved 52 times smaller
capsule can represent numerous features while CNN can only network parameter than U-net and has achieved very close
represent one feature at a time. Preserving features together in accuracy in mean IoU compared to U-net.
form of capsule is very efficient way rather than keeping
features in a separate manner. This property can be seen that We used 3-class segmentation dataset, images of liver cell
many capsule networks succeed to accomplish higher from mice to show that U-EncapNet outperforms U-net. We
accuracy with fewer layer than CNN. Routing-by-agreement also compared EncapNet-3D with resnet-6 which has only
is a pooling layer for capsules. Unlike max pooling, which single resnet block. EncapNet had maximum of 5% increase
loses precise location information every time it has been used, in comparison with resnet-6 in mIoU. We also used
routing-by-agreement will learn to preserve location segmented anisotropic ssTEM dataset [3] which is 5-class
information no matter how many times routing has been used. segmentation dataset.
Dynamic routing was the first routing-by-agreement to be This paper is organized as follows. In section 2, we
introduced. It used huge parameters for pooling. Dynamic introduce related works in capsule network and real time
routing had achieved high accuracy in dataset like MNIST. semantic segmentation. We are going to introduce our
However, because of the huge parameters, it had difficulty proposed method EncapNet-3D in section 3 and U-EncapNet
getting high accuracy in datasets which has bigger input image in section 4. We show the experimental results of EncapNet-
size than MNIST. Recently, EncapNet [2] has been introduced 3D and U-EncapNet on mice liver image dataset and
which achieved very high accuracy in CIFAR10, CIFAR100, anisotropic ssTEM dataset in section 5. Conclusions and
SVHN and ImageNet. EncapNet shared a new routing-by- future works are described in section 6.
agreement algorithm called approximate routing.
Approximate routing has succeeded to significantly improve II. RELATED WORKS
routing-by-agreement, which has thought to be the one of the Capsule network [1] was introduced by S.Sabour et al. and
biggest bottlenecks of capsule network. has been adopted to various tasks. SegCaps [4] has
In this paper, we prove that EncapNet can achieve high successfully adopted it to object segmentation using U-net
accuracy not only in classification but also in semantic architecture and successfully reduced dynamic routing
segmentation. SegCaps[4] has succeeded to achieve high
978-1-7281-3857-2/19/$31.00 ©2019 IEEE

Authorized licensed use limited to: Institut Pasteur. Downloaded on February 17,2023 at 14:13:24 UTC from IEEE Xplore. Restrictions apply.
Figure 1 Network architecture of EncapNet-3D
parameter size by using local capsules instead of using all the structures compared to CNN. Capsule Network tends to
capsules in the layer. J.Gong at el. [5] adopted Capsule achieve higher accuracy in very simple network compared to
network to sequence Encoding. F.Vedperini [6] applied it to CNN. These networks have very few parameters but do not
SED task. have network depths as shallow as capsule network. Our
Due the huge parameters used in dynamic routing, Capsule EncapNet-3D only uses one block which means shallow
network had difficulty achieving high accuracy in some tasks network and has very few parameters at the same time.
that has large input image size. To improve the Capsule
network, improving one of the biggest bottlenecks of capsule III. ENCAPNET-3D
network which was the dynamic routing is one of the main EncapNet has achieved very high accuracy on ImageNet-
goals in research. There are many improvements about 1K. EncapNet uses approximate routing which greatly
dynamic routing. D.Wang et al. [7] has formulated Dynamic reduces parameters used in routing. Approximate routing
routing as an optimization problem. This idea has been used divides input capsule into two branches; the master and aide
in EncapNet which uses OT loss to enforce the network branch as.
regularization. G.Hinton et al. [8] proposed Matrix capsules
which uses matrix to reduce parameters and EM routing for ܵ௝ ൎ ݉௔௜ௗ௘ ‫ݒ‬ො௠௔௦௧௘௥ ൅ ݉௠௔௦௧௘௥ ‫ݒ‬ො௔௜ௗ௘ (1)
routing-by-agreement which is more superior than dynamic
routing but had huge parameter size like dynamic routing . where ܵ௝ is an output of j-th channel and ‫ݒ‬ො is the input
K.Duarte at el. [9] proposed capsule-pooling to reduce capsule for approximate routing. Master branch ‫ݒ‬ො௠௔௦௧௘௥ uses
parameters in matrix capsule and applied it for action group convolution to see only capsules that are directly
detection. A.Mobiny et al. [10] proposed 3D capsule Network related its own capsule. Aide branch ‫ݒ‬ ෝ ௔௜ௗ௘ uses normal
and reduced coefficient parameters for dynamic routing. convolution to gather information from capsules located in
M.Bahadori proposed Spectral Capsule Networks for other channels. Master and aide branch uses co-efficient
leaning-to-diagnose problem which preforms better than ݉௠௔௦௧௘௥ , ݉௔௜ௗ௘ to interact with each other. Co-efficient is
Matrix Capsules. L.Zhang at el. proposed CapProNet [11] generated by group convolution from both master and aide
which learns group of capsule subspaces to represent each branches which are multiplied to each capsule in aide and
class.MS-CapsNet [12] proposed Dropout for capsule and master branch.
achieved higher accuracy in CIFAR10 than original capsule We propose EncapNet-3D, which is based on EncapNet.
network. These routing improvements are likely to be EnCapNet-3D uses single EncapNet-3D module to reduce
specialized to certain tasks and are difficult to apply to parameters. The goal is to achieve close accuracy in just one
another tasks. In semantic segmentation, pre-training with module. EncapNet-3D uses 3D convolution to generate co-
datasets like classification task helps to achieve higher efficient. Which is the red block shown in Figure 1. Fast 3D
accuracy in U-net [13]. In order to do this in capsule network, capsule network [8] has used 3D convolution instead of 2D
routing algorithm which can achieve high accuracy in convolution and succeeded to get higher accuracy than 2D
classification and segmentation is needed. capsule Network.3D convolution can easily preserve capsule
U-net [13] archived high performance in semantic structure than 2D convolution.
segmentation. In cell segmentation, some methods [18-22] Inputs for 3D convolution and 2D group convolution are
had been proposed. A.Paszke [14] proposed ENet, that has (4x6,128,128) and (4,6,128,128). Both are separated by
very high performance considering the small parameters used groups and have no interaction between channels which avoid
in the network. MobileNet [15] uses depthwise separable capsules form corruption. The main difference is that 3D
convolution to achieve high accuracy in low parameters. convolution uses 3x3x3 kernel which helps to get the
ShuffleNet [16] uses group convolution and channel shuffle information of different element of neighbor capsules. 2D
and succeeded to reduce network parameters. A.Lin et al. [17] convolution only has 3x3 kernel which can get only
showed that capsule network can capture global manifold information from same neighbor capsule element which is

Authorized licensed use limited to: Institut Pasteur. Downloaded on February 17,2023 at 14:13:24 UTC from IEEE Xplore. Restrictions apply.
Figure 2 Details of U-EncapNet

placed in same plane. EncapNet uses aide branch to share Considering the skip connection between encoder and
information, but 3D convolution generates additional decoder, we changed the measuring point of OT loss. OT loss
information about neighbor capsules which helps EncapNet- is one way to describe the divergence of two distributions.
3D to achieve higher accuracy. EncapNet only generated co- Dynamic routing can be described as optimization problem [7]
efficient per capsule. In order to make master and aide branch and original EncapNet used OT loss as part of approximate
interact more efficient, it is better to generate co-efficient for routing. Each block contains dropout layer since U-EncapNet
each capsule element. Without 3D convolution, capsule tends to cause overfitting very easily.
structure will likely collapse. In order to supply capsule V. EXPERIMENTS
structure to another branch, we used 3D convolution.
EncapNet-3D also uses dropout to reduce overfitting. This section shows the experimental results of EncapNet-
EncapNet is very similar to normal CNN because of the 3D and U-EncapNet. We explain the dataset used in section
approximate algorithm structured only by the combination of 5.A and evaluation results for EncapNet-3D in section 5.B
convolution layers and squash function. The distance between and U-EncapNet in section 5.C.
them can be easily narrowed down by changing the capsule A. Dataset
dimension to 1 which makes capsule a scalar which makes the
network very similar to normal CNN. Changing the channels We used two datasets to evaluate our methods. One is a
to 1 also makes EncapNet very close to normal group fluorescence image of the transgenic mice that expressed
convolution. We propose the learning sequence of stair fluorescent markers on the cell membrane, cell nucleus and
learning for EncapNet that gently updates the channels from 1 nucleus [18]. There are 50 images total and the size of the
and gently increasing it to aim parameter helps learning image is 256 x 256 pixels. We random cropped the image into
capsules to learn better. We used normal 2D convolution to 128 x 128 pixels and increased the image 2500 total. We
generate feature maps from input image and prediction output divided the dataset as 2000 for train and 250 for validation
and 250 for test.
IV. U-ENCAPNET Another dataset is segmented anisotropic ssTEM dataset
Encapnet is specialized in classification task. Our [12] which has 20 images total and a pixel size of 1024x1024.
proposed method U-EncapNet uses U-net architecture to It includes membrane, mitochondria, synapses,
archive high IoU. U-net has achieved high accuracy on glia/extracellular and intracellular classes. We cropped the
various segmentation tasks. U-net uses encoder-decoder dataset and reduced the image size into 512x512 and raised
architecture and skip-connection to supply feature map to the image number to 80. We divided the dataset as 60 for train
decoder part to support Encoder to get high IoU on and 8 for validation and 12 for test. As an evaluation measure,
segmentation task. CNN loses precise location information we used intersection over union (IoU) which is described as
when feature maps are passed through pooling many times. In ܶܲ
order to supply location information, U-net concatenated ‘ ൌ ሺʹሻ
feature maps of same layers in encoder together where ܶܲ ൅ ‫ ܲܨ‬൅ ‫ܰܨ‬
where it describes the overlap rate between ground-truth area
location information has not been lost so much.
and predicted area.
Figure 2 shows the structure of U-EncapNet. U-Encapnet
uses EncapNet block at each layer. First 2 blocks do not use
routing. This is because original EncapNet has achieved high
accuracy in ImageNet when first 2 blocks had no routing.

Authorized licensed use limited to: Institut Pasteur. Downloaded on February 17,2023 at 14:13:24 UTC from IEEE Xplore. Restrictions apply.
Table 1 comparison in cell dataset on EncapNet-3D in average (channel, dim)
Membrane [%] Nucleus [%] Background [%] mIoU [%] parameters(Mb)
EncapNet-3D+MS-dropout(7,6) 42.66 62.81 69.10 58.19 0.741
EncapNet-3D(7,6) 42.08 63.40 70.52 58.66 0.741
EncapNet-3D(6,7) 41.91 63.01 70.64 58.52 0.787
EncapNet-3D-stair(7,6) 42.55 61.88 70.09 58.17 0.741
EncapNet-3D-stair(6,7) 40.37 59.20 69.40 56.32 0.787
EncapNet(7,6) 38.98 58.85 69.31 55.72 0.720
EncapNet(6,7) 38.46 60.73 69.84 56.34 0.773
Resnet-6 41.02 52.25 66.12 53.13 5.006

Table 2 comparison in cell dataset on EncapNet-3D in max (channel, dim)


Membrane [%] Nucleus [%] Background [%] mIoU [%]
EncapNet-3D+MS-dropout(7,6) 43.49 63.07 69.43 58.25
EncapNet-3D(7,6) 43.26 64.86 70.96 59.66
EncapNet-3D(6,7) 43.29 63.98 70.84 59.37
EncapNet-3D-stair(7,6) 44.28 62.90 70.47 58.61
EncapNet-3D-stair(6,7) 41.59 61.29 69.95 57.61
EncapNet(7,6) 39.09 60.24 69.88 56.40
EncapNet(6,7) 40.78 61.43 70.37 57.16
Resnet-6 41.51 53.08 66.63 53.40

Table 3 comparison in ssTEM dataset on EncapNet-3D in average (channel, dim)


Membrane[%] Mitochondria[%] Synapse[%] Glia/Extracellular[%] Intracellular[%] mIoU[%] parameters(Mb)
EncapNet-3D 57.79 45.64 32.19 36.32 89.64 52.32 0.74
EncapNet-2D 59.82 40.79 20.19 34.64 89.85 49.06 0.77
EncapNet 54.04 44.76 29.42 37.76 89.60 51.11 0.72
ResNet-6 62.38 37.97 30.81 32.45 89.67 50.66 0.88
U-Net 70.58 67.63 0.00 36.03 91.78 53.20 39.23

Figure 3 Segmentation results on EncapNet-3D and ResNet-6

compare to original EncapNet. We used only one EncapNet


B. EncapNet-3D block to match the network depth. Similar comparison has
We compare ResNet-6 which uses only one block. been in original EncapNet paper [2]. We also added Dropout
ResNet-6 was chosen to make network depth same. We also which was proposed in MS-CapNet [10] shows that it had

Authorized licensed use limited to: Institut Pasteur. Downloaded on February 17,2023 at 14:13:24 UTC from IEEE Xplore. Restrictions apply.
improvements. Table 1 shows the IoU for the EncapNet-3D.

Figure 4 Segmentation results on EncapNet-3D in ssTEM dataset

Table 4 comparison on U-EncapNet in average


Membrane [%] Nucleus [%] Background [%] mIoU [%] parameters(Mb)
U-Net 47.29 66.88 67.87 60.68 39.23
U-EncapNet 45.26 68.54 71.63 61.81 238.53
U-EncapNet-OT_loss 44.69 68.73 71.91 61.78 238.53
U-EncapNet-OT_loss- 45.87 67.69 71.22 61.59 244.01
U-EncapNet+KL_loss 45.93 68.42 71.56 61.97 238.53

Table 5 comparison on U-EncapNet in max


Membrane [%] Nucleus [%] Background [%] mIoU [%]
U-Net 47.51 67.31 69.02 61.15
U-EncapNet 46.76 69.84 71.84 62.33
U-EncapNet-OT_loss 45.02 68.85 72.09 61.81
U-EncapNet-OT_loss- 47.38 68.34 71.95 62.14
U-EncapNet+KL_loss 46.45 69.30 71.88 62.25

Table 6 comparison in ssTEM dataset on U-EncapNet in average (channel, dim)


Membrane[%] Mitochondria[%] Synapse[%] Glia/Extracellular[%] Intracellular[%] mIoU[%] parameters(Mb)
U-EncapNet 66.50 69.81 30.63 54.55 90.86 62.47 180.42
U-Net 70.58 67.63 0.00 36.03 91.78 53.20 39.23

Authorized licensed use limited to: Institut Pasteur. Downloaded on February 17,2023 at 14:13:24 UTC from IEEE Xplore. Restrictions apply.
All IoU have been measured 3 times and then being averaged. C. U-EnCapNet
We can clearly see that all EncapNet-3D exceeds the IoU in Table 3 shows the evaluation result of U-EncapNet.
Resnet-6 and original EncapNet. Parameters are very low UEncapNet-OT_loss- uses the same measuring point as
compared to ResNet-6. EncapNet-3D-stair which used stair
original EncapNet. UEncapNet-OT_loss has no OT loss.
leaning has higher mIoU than ResNet-6 and original
UEncapNet+KL_loss uses additional KL_loss at the same
EncapNet.
measuring point. We can see that U-EncapNet has about

Figure 5 segmentation result on U-EncapNet

Figure 6 segmentation result of U-EncapNet in ssTEM dataset


compared to ResNet-6. EncapNet-3D (7,6) has achieved 1.1% increase in mIoU. Changing OT loss has improved
EncapNet-3D-Capsule-dropout which used Capsule-dropout mIoU in 0.1%. U-net has higher membrane IoU but also the
only in co-efficient part of the approximate routing has background IoU is lower than U-EncapNet. Table 4 shows
achieved the highest membrane IoU of all. EncapNet-3D (7,6) the evaluation result in max. We can see that U-EncapNet has
has membrane IoU increase 1.6% and 5.5% in mIoU the small difference between average and max.
highest mIoU of all in this comparison we can clearly see that Table 6 shows the evaluation results in ssTEM dataset. We
using 3D convolution in approximate routings co-efficient shrunk the network filter size in order to fit in one GPU. We
part helps the EncapNet to achieved better performance in can see that U-EncapNet has 2.2% increased in comparison
IoU. Table 2 shows the maximum IoU in EncapNet-3D. This in Glia/Extracellular class. U-EncapNet also has 18.52%
shows that EncapNet-3D-stair is stronger than EncapNet-3D
increased in comparison with U-net in Synapse class which
in membrane IoU. EncapNet-3D has highest IoU in all expect
membrane. We can see the clear difference between ResNet- tends to be small. Figure 4 shows the segmentation results
6 which has nearly 3% increase in membrane IoU and 6.2% from U-EncapNet. We can clearly see that U-net has thick
increase in mIoU. Figure 3 shows the segmentation results of membrane to get higher membrane IoU. U-EncapNet has
EncapNet-3D we can see the clear difference between thinner membrane which is more difficult to get higher
EncapNet-3D and ResNet-6. Membrane IoU because in order to get a thin membrane and
achieve a high membrane IoU at the same time needs precise
Table 3 shows that EncapNet has overcome most of the location information. Capsule network preserves the location
class IoU other than membrane IoU. Mitochondria class had information which makes it possible to achieve this.
7.66% increased in comparison in ResNet-6. EncapNet-2D Figure 7 shows the segmentation result of U-EncapNet. We
which uses 2D group convolution in co-efficient part has
can see that U-EncapNet has thinnest cell membrane of all.
lower mIoU than EncapNet-3D. Capsule Network tends to
By the segmentation results of cell image dataset and ssTEM
struggle with large images because of the huge parameter in
routing, but our proposed method EncapNet-3D successfully dataset, U-EncapNet tends to predict thinner membrane
outperforms ResNet-6 and original EncapNet with very little compared to U-net.
parameter increase compared to original EncapNet. Figure 6 D. Consideration
shows the segmentation results in ssTEM dataset. We can see
that EncapNet-3D has the clearest segmentation image Looking at Table 1 and Table 4, mIoU has about 3.31%
compared to ResNet-6 and EncapNet. difference and membrane IoU has about 4.62% difference,
but comparing Table 2 and Table 4 narrows the difference
between U-EncapNet and EncapNet-3D to 3.2% in
membrane IoU and 2.6% in mIoU. Considering the

Authorized licensed use limited to: Institut Pasteur. Downloaded on February 17,2023 at 14:13:24 UTC from IEEE Xplore. Restrictions apply.
parameters of EncapNet-3D, it has achieved very low [5] J.Gong,X.Qiu,S.Wang,X.Huang. Information aggregation via dynamic
routing for sequence encoding. arXiv preprint arXiv:1806.01501.
performance decrease.
[6] T.Iqbal, Y.Qu, Q.Kong, W.Wang. Capsule routing for sound event
Looking at Table 3 and Table 6, We can see that EncapNet- detection arXiv preprint arXiv:1806.04699
3D and U-net has only 0.88% difference in mIoU and has 52 [7] D.Wang, Q.Liu , An Optimization View on Dynamic Routing Between
times smaller network parameter in comparison in U-net. We Capsules , In International Conference on Learning Representations
successfully got very close accuracy in EncapNet-3D (ICLR),2018
comparted to U-net. This dataset was huge and is challenging [8] G.Hinton , S.Sabour ,N.Frosst, Matrix Capsules with EM routing, In
for capsule network because not so many has succeeded to International Conference on Learning Representations (ICLR),2018
deal with large image input like 512x512 in capsule network [9] K.Darte, Y.S.Rawat, M.Shah. Videocapsulenet :A simplified network
for action detection arXiv preprint arXiv:1805.08162,2018
[10] A.Mobiny,H.V.Nguyen. Fast capsnet for lung cancer screening. arXiv
preprint arXiv:1806.07416, 2018
VI. CONCLUSION
[11] L. Zhang, M. Edraki, G.Qi. Cappronet: Deep feature learning via
In this paper, we proposed EncapNet-3D, which uses 3D orthogonal projections onto capsule subspaces. arXiv preprint
convolution for co-efficient parameter in master, aide branch arXiv:1805.07621, 2018
of approximate routing. EncapNet-3D had successfully [12] C.Xiang, L.Zhang, Y.Tang, W.Zou, C.Xu. MS-CapsNet: A novel
multi-scale capsule network. IEEE Signal Processing Letters,
achieved very close accuracy with only one block in some 25(12):1850–1854, 2018.
class IoU in semantic segmentation compared to U-net in cell [13] P.Fischer, and T.Brox. U-net: Convolutional Networks for Biomedical
image dataset and ssTEM dataset. We reduced the network Image Segmentation. Medical Image Computing and Computer-
parameters to 52 times smaller than U-net. Assisted Intervention, pp.234-241, 2015.
We also proposed U-EncapNet, which is EncapNet that [14] A. Paszke, A. Chaurasia, S. Kim, and E. Culurciello, “Enet: A deep
uses U-net architecture. U-EncapNet had successfully neural network architecture for real-time semantic segmentation,”
arXiv preprint arXiv:1606.02147, 2016
outperformed U-net in most class IoU on cell image dataset
[15] A. G. Howard, M.Zhu, B.Chen, D.Kalenichenko, W.Wang, T.Weyand,
and ssTEM dataset. M.Andreetto, H.Adam, Mobilenets: Efficient convolutional neural
However, both networks had difficulty achieving high networks for mobile vision applications, arXiv preprint
accuracy in membrane IoU in both datasets. U-EncapNet had arXiv:1704.04861, 2017.
thinner membrane which is more difficult to achieve high [16] X.Zhang, X.Zhou, M.Lin, J.Sun, Shufflenet: An extremely efficient
accuracy than U-net which had thick membrane. EncapNet- convolutional neural network for mobile devices, arXiv preprint
arXiv:1707.01083, 2017.
3D on other had higher cell membrane IoU than ResNet-6 in
[17] A.Lin, J.Li2,Z.Ma On Learning and Learned Representation with
cell image dataset but it had difficulty in ssTEM dataset. In Dynamic Routing in Capsule Networks arXiv preprint
future, we would like to improve the accuracy of cell arXiv:1810.04041, 2018.
membrane by improving the approximate routing of an [18] A.Imanishi, T.Murata, M.Sato, K.Hotta, I.Imayoshi, M.Matsuda,
EncapNet-3D. K.Terai, A novel morphological marker for the analysis of molecular
activities at the single-cell level, Cell Structure and function
43(2),pp.129-140,2018
ACKNOWLEDGMENT
[19] M. Havaei, A. Davy, D.W. Farley, A. Biard, A. Courville, Y.Begio, C.
This work is partially supported by MEXT/JSPS Pal, P.M. Jodoin, and H. Larochelle, “Brain tumor segmentation with
KAKENHI Grant Number 18H04746 "Resonance Bio". Deep Neural Networks,” Medical Image Analysis, vol. 35 pp. 18-31,
2017.
REFERENCES [20] O.X. Ji, Y. Li, J. Cheng, Y. Yu, and M. Wang, “Cell image
[1] S.Sabour, N.Frosst, G.E.Hinton Dynamic Routing Between Capsules, segmentation based on an improved watershed algorithm,”
Advances in Neural Information Processing Systems, pp.3859-3869, International Congress on Image and Signal Processing, pp. 433-437,
2017 2015.
[2] H.Li, X.Guo, B.Dai, W. Ouyang, X. Wang. Neural Network [21] Y. Hiramatsu, K. Hotta, A. Imanishi, M. Matsuda, and K. Terai, “Cell
Encapsulation, European Conference on Computer Vision, pp.266- Image Segmentation by Integrating Multiple CNNs,” IEEE Conference
282, 2018. on Computer Vision and Pattern Recognition Workshops, pp. 2205-
2211, 2018.
[3] S.Gerhard, J.Funke, J.Martel, A.Cardona, R.Fetter. Segmented
anisotropic ssTEM dataset of neural tissue, [22] T.Hiroki,K.Hotta,”Cell Image Segmentation by Integrating Pix2pixs
http://dx.doi.org/10.6084/m9.figshare.856713 ,2013 for Each Class”The IEEE Conference on Computer Vision and Pattern
Recognition (CVPR) Workshops, 2019
[4] R.Lalonde and U.Bagci. Capsules for objects segmentation. arXiv
preprint arXiv:1804.04241,2018.

Authorized licensed use limited to: Institut Pasteur. Downloaded on February 17,2023 at 14:13:24 UTC from IEEE Xplore. Restrictions apply.

You might also like