Upgraded W-Net with Attention Gates and Its Application in

Unsupervised 3D Liver Segmentation

Dhanunjaya Mitta

, Soumick Chatterjee

1,2,3 a

, Oliver Speck

3,4,5,6 b

and Andreas N

urnberger

1,2 c

Faculty of Computer Science, Otto von Guericke University, Magdeburg, Germany

Data and Knowledge Engineering Group, Otto von Guericke University, Magdeburg, Germany

Biomedical Magnetic Resonance, Otto von Guericke University Magdeburg, Germany

German Center for Neurodegenerative Disease, Magdeburg, Germany

Center for Behavioral Brain Sciences, Magdeburg, Germany

Leibniz Institute for Neurobiology, Magdeburg, Germany

Keywords:

Unsupervised Learning, Deep Learning, MRI Segmentation, Liver Segmentation.

Abstract:

Segmentation of biomedical images can assist radiologists to make a better diagnosis and take decisions faster

by helping in the detection of abnormalities, such as tumors. Manual or semi-automated segmentation, how-

ever, can be a time-consuming task. Most deep learning based automated segmentation methods are supervised

and rely on manually segmented ground-truth. A possible solution for the problem would be an unsupervised

deep learning based approach for automated segmentation, which this research work tries to address. We use a

W-Net architecture and modiﬁed it, such that it can be applied to 3D volumes. In addition, to suppress noise in

the segmentation we added attention gates to the skip connections. The loss for the segmentation output was

calculated using soft N-Cuts and for the reconstruction output using SSIM. Conditional Random Fields were

used as a post-processing step to ﬁne-tune the results. The proposed method has shown promising results,

with a dice coefﬁcient of 0.88 for the liver segmentation compared against manual segmentation.

1 INTRODUCTION

Image Segmentation is the process of dividing an im-

age into multiple segments, where the pixels in each

segment are connected with respect to their intensities

or by Regions of Interest (Anjna and Er, 2017). Seg-

mentation of biomedical images is a major advance-

ment in the ﬁeld of medical imaging, as it helps ra-

diologists and doctors to make better and faster deci-

sions. Many approaches to medical image segmen-

tation using various deep learning techniques have

been proposed. These methods, however, require a

large amount of training data with their respective

segmentation masks also known as ground truth im-

ages (Badrinarayanan et al., 2017; Chaurasia and Cu-

lurciello, 2017; Kr

ahenb

uhl and Koltun, 2011; Paszke

et al., 2016; Zheng et al., 2015). Abdominal MR im-

age segmentation is an interesting and challenging re-

search area (Gotra et al., 2017), but not yet very much

https://orcid.org/0000-0001-7594-1188

https://orcid.org/0000-0002-6019-5597

https://orcid.org/0000-0003-4311-0624

explored, until recently (Kavur et al., 2020). While

performing abdominal segmentation, liver segmenta-

tion is one of the most challenging task due to the high

variability of its shape and its proximity to various

other organs (Gotra et al., 2017). This research ad-

dresses the challenge of segmenting the liver from 3D

MR images without using any manual ground truth

for training the deep neural network model.

Our state of the art model is based on W-Net (Xia

and Kulis, 2017) with both the U-Nets replaced by

Attention U-Nets (Oktay et al., 2018). The original

W-Net works with 2D images, but as we want to work

with volumetric 3D MR images, the network architec-

ture was adapted for 3D images by using 3D convo-

lution layers (Sect. 2.3) and by modifying the calcu-

lation of pixel weights to voxel weights (Sect. 2.2).

We show the applicability of our approach for liver

segmentation using the CHAOS (Kavur et al., 2020)

challenge dataset (Sect. 2.1).

488

Mitta, D., Chatterjee, S., Speck, O. and Nürnberger, A.

Upgraded W-Net with Attention Gates and Its Application in Unsupervised 3D Liver Segmentation.

DOI: 10.5220/0010221504880494

In Proceedings of the 10th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2021), pages 488-494

ISBN: 978-989-758-486-2

1.1 Related Work

Many approaches to image segmentation have been

proposed by different researchers. A variety of atlas-

based segmentation methods have been described

(Gee et al., 1993; Crum et al., 2001; Baillard et al.,

2001). Aganj et.al introduced an approach by com-

puting the local center of mass of the putative region

of each pixel, to perform unsupervised medical im-

age segmentation (Aganj et al., 2018). Dong Nie et.al

proposed an approach for brain image segmentation

of infants by using deep neural networks (Nie et al.,

2016). Christ et. al designed an approach by joining

two fully cascaded neural networks for automatic seg-

mentation of the liver and its lesions in low-contrast

heterogeneous medical volumes (Christ et al., 2017).

Oktay et.al introduced a novel attention gate, which

will implicitly learn to suppress regions that are not

relevant (Oktay et al., 2018). These gates are ap-

plied to the standard U-Net architecture to highlight

the important features, which are passed through skip

connections. Noise and irrelevant information in skip

connections are eliminated by extracting coarse-scale

information in gating. This is performed right before

the concatenation operation to merge only relevant ac-

tivations. Xide Xia et.al proposed an approach for a

W-Net model by stacking two U-Nets one after an-

other, for unsupervised image segmentation, but for

non-medical RGB images (Xia and Kulis, 2017). By

using this model, segmentation maps can also be pre-

dicted even for applications, which do not have any

labeling information available.

1.2 Contribution

Most of the research on biomedical image segmen-

tation using deep learning by now has been focused

on supervised learning. This research is a proof-of-

concept for biomedical image segmentation using un-

supervised learning. The current results are not per-

fect, but there are many scopes for improvements -

that will be discussed later. In this research, a novel

3D Attention W-Net architecture has been proposed,

which has been built by replacing the 2D U-Nets of

the original W-Net (Xia and Kulis, 2017), by the 3D

Attention U-Nets (Oktay et al., 2018), and for the re-

construction loss, SSIM (Larkin, 2015) has been used.

Furthermore, some minor changes were introduced

to the Attention U-Net architecture before incorporat-

ing them to the W-Net, which are discussed in a later

chapter.

2 METHODOLOGY

2.1 Dataset

The dataset that has been used in this study has been

provided by the CHAOS Challange (Kavur et al.,

2019; Kavur et al., 2020). The dataset consists of

a CT Dataset of 40 subjects, and an MRI Dataset

of 40 subjects, with two different sequences - T1-

DUAL and T2-SPIR. T1-DUAL contains in-phase

and opposed-phase images. For our work, we choose

the available 40 volumes of T1-DUAL in-phase. The

dataset came with a manually labeled ground-truth.

For the purpose of this research they were intention-

ally ignored during training. Those ground-truths

were used only during the evaluation of the algo-

rithm’s performance.

2.2 Pre-processing

The images were normalized to have pixel values be-

tween (0,1) before supplying them to the network, to

bring them to a common scale for faster convergence

while training. Simultaneously, the weights between

the voxels were calculated using Eq. 1, where w

i j

is the weight between the pixel i and j, which is re-

quired in calculating Normalized-Cuts using Eq. 3

(loss function). The architecture is based on auto-

encoders in which the encoder part maps the input

to the pixel-wise segmentation layer without losing

its original spatial size and the decoder part recon-

structs the original input image from the dense pre-

diction layer.

i, j

= e

−

−F

∗



−

−X

(1)

2.3 Model Construction: 3D Attention

W-Net

The base W-Net architecture proposed by (Xia and

Kulis, 2017) was modiﬁed by replacing both the U-

Nets with 3D Attention U-Nets (Oktay et al., 2018).

The original W-Net was proposed for 2D Images,

both weight calculation and soft ncuts loss calculation

have been adopted for 3D.

The network is illustrated in 1. The network con-

sists of two parts.

• AU-Encoder, which is on the left side of the net-

work and

• AU-Decoder, which is on the right.

Upgraded W-Net with Attention Gates and Its Application in Unsupervised 3D Liver Segmentation

489

Figure 1: 3D Attention W-Net.

The network consists of 18 modules (marked with

dotted lines in Figure 1), each module consists of two

3D convolutional layers with kernel size three. Each

layer is followed by a non-linear activation function

and an instance normalization layer. In total, we are

using 46 3D convolutional layers. The ﬁrst nine mod-

ules represent the encoder network which predicts the

segmentation maps and the next nine modules recon-

struct the original input image from the segmentation

output coming from the encoder part.

The most frequently used non-linear activation

function is the Rectiﬁed Linear Unit (ReLU). How-

ever, there is a chance of dying neurons (Lu et al.,

2019). Therefore, we used the Parametric Rectiﬁed

Linear Unit or PReLU (He et al., 2015), which is

similar to LeakyReLU with the difference of using

the hyper-parameter α for negative results, which is

adaptively learnt during the training, instead of using

a ﬁxed value (such as 0.01) as in LeakyReLU. The

data used in this research contain different patient data

and the number of slices differs between subjects. To

construct batches, data padding to an equal number

of pixels would be required. Instead, we used a batch

size of one while training the network.

As mentioned in the literature (Oktay et al., 2018),

the encoder part consists of a contracting path that

captures context and an expansion path that enables

precise localization. As shown in Figure 1, an input

image is given to the ﬁrst module of the encoder part.

Then it undergoes convolutional operations followed

by PReLU and instance normalization twice before

moving forward to the next module. The modules are

connected through 3D max pooling layers, which de-

crease the image size by two. We also store the orig-

inal image size before performing the pooling opera-

tion recover the image size during the expansion path

of the U-Nets. The initial module produces 64 feature

Figure 2: Attention gate (Oktay et al., 2018).

maps as output and after every module, the number of

features is increased by two.

In the contraction path, modules are connected via

max pool, which is indicated in brown color; in the

expansion path, modules are connected through the

upsample layer followed by modules similar to the

contraction path and are denoted with green color ar-

rows. Upsampling is performed using trilinear inter-

polation, and the output size of the interpolation is set

to the image sizes saved ptior to each of the max pool

operations. Skip connections are passed through at-

tention gates to suppress irrelevant regions and noisy

responses. The attention gate architecture is from

(Oktay et al., 2018) and shown in Figure 2.

The output of the encoder is passed to a fully con-

nected 3D convolution layer with a kernel size of one

and a stride of one, followed by a softmax layer. This

convolution layer helps to map the 64 feature maps of

the output to the required number of K classes, and the

softmax function rescales them to (0,1) with a sum-

mation of all the K feature maps as one. During the

inference stage, the output of the softmax layer is the

ﬁnal output of the model. During training, the output

of the softmax is given as the input of the ﬁrst module

of the second U-Net. The second U-Net is similar to

the ﬁrst one, with the only differences being the ﬁnal

fully connected convolution layer and the ﬁnal activa-

tion function. The fully connected layer provides one

ﬁnal output instead of K outputs. For the ﬁnal activa-

ICPRAM 2021 - 10th International Conference on Pattern Recognition Applications and Methods

490

tion function, the sigmoid function is used instead of

the softmax, which also rescales the output between

(0,1) but doesn’t make the sum equal to one.

2.4 Loss Functions

We used two loss functions in this research. The ﬁrst

one is directly after the encoder U-Net, to optimize

the encoder U-Net only; the other one is at the end of

the decoder U-Net to optimize both the U-Nets.

2.4.1 N-Cuts Loss

The ﬁrst loss function applied to the output of the en-

coder U-Net is the N-Cuts loss (Xia and Kulis, 2017).

The output from the softmax layer of the encoder U-

Net is a K-class prediction for each voxel. Normal-

ized cuts from (Shi and Malik, 2000) as a global cri-

terion for image segmentation, as shown in Eq. 2 are

applied, where A

is the number of voxels in segment

k, V is the total number of voxels, and w calculates

the weight between two pixels.

Since the argmax function is non-differentiable, it

is not possible to get the corresponding gradients dur-

ing back-propagation. Therefore, the soft n-cuts loss

(Ghosh et al., 2019) is used as shown in Eq. 3, where

p(u = A

) measures the probability of node u belong-

ing to class A

. The output of the encoder U-Net is

forwarded to this soft N-Cuts loss function along with

the voxel-weight calculated during the pre-processing

stage, following Eq. 1. The network is trained to min-

imize the N-Cuts loss, by optimizing the parameters

of the encoder U-Net.

Ncut

(V ) =

∑

k=1

∑

uεA

,vε(V −A

)

ω(u, v)

∑

uεA

,tε(V )

ω(u, t)

(2)

so f t−Ncut

(V, K) =

K −

∑

k=1

∑

uεV

p(u = A

)

∑

uεV

ω(u, v)P(v = A

)

∑

uεV

p(u = A

)

∑

tεV

ω(u, t)

(3)

2.4.2 Reconstruction Loss

Reconstruction loss is used to calculate the loss be-

tween the output of the decoder U-Net and the in-

put image. The network was trained to minimize the

reconstruction loss similar the auto-encoder architec-

ture. Structural Similarity Index (SSIM) is used to

calculate the reconstruction loss. A higher SSIM,

however, is better and thus the negative of the SSIM

value has been used.The network was trained to min-

imize the reconstruction loss, by optimizing the pa-

rameters of both the U-Nets.

SSIM is used to measure the similarities within

the pixels i.e., whether the pixels in the images those

are being compared have similar pixel density values.

SSIM values lie between (0,1), where 1 indicates that

both images are identical.

We calculate SSIM by using the following for-

mula:

SSIM(x, y) =

(2µ

) + (2σ

)

(µ

+ µ

)(σ

+ σ

)

(4)

where µ

, µ

are the mean values of x and y, σ

are the variance of x and y, σ

is the co-variance

of x and y, C

, C

are two variables to stabilize the

division with weak denominator.

The network was trained with the MR abdominal

dataset provided by (Kavur et al., 2020), which con-

tains 40 volumes. The given set was divided into a

training (25 volumes), a validation (5 volumes), and a

test set (10 volumes). Both N-cuts and reconstruction

loss were minimized, given equal priority (weights)

to both loss functions.

2.5 Post-processing using Conditional

Random Fields

The use of many max-pooling layers may result in in-

creased invariance, which can cause localization ac-

curacy reduction. To obtain ﬁne boundaries in the

output segments, conditional random ﬁelds or CRF

(Chen et al., 2017) were applied as a post-processing

step in a 3D CRF variant (Chatterjee et al., 2020).

E(X) =

∑

φ(u) +

∑

ψ(u, v) (5)

Where u and v are the voxels, φ(u) is the unary

potential and ψ(u, v) is the pair-wise potential.

After the CRF, the cluster values corresponding

to the liver were identiﬁed manually for one volume.

The selected clusters were merged to obtain the liver

segmentation for the remaining volumes.

3 RESULTS

We used two U-Nets to form a W-Net for training the

model on a given training dataset; during testing, only

the ﬁrst U-Net was used, as the output of the ﬁrst U-

Net corresponds to the automatic segmentation. This

predicted segmentation was then passed through the

CRF post-processing to recover the boundaries. The

network was trained to predict 15 different clusters to

segment various parts of the image. Then, the clusters

with the liver segment were identiﬁed as the ﬁnal re-

sult. The relevant cluster numbers were chosen from

Upgraded W-Net with Attention Gates and Its Application in Unsupervised 3D Liver Segmentation

491

only one test volume and applied to all other test vol-

umes.

The results were compared to the available

ground-truth. Only the liver as our region of interest

was considered from both output and ground-truth.

Two representative slices, corresponding to the man-

ually segmented liver and the predicted segmentation

of the liver are shown in Figure 3 and 4. The pro-

posed liver segmentation is compared to the ground

truth liver segmentation quantitatively using intersec-

tion over union and dice coefﬁcient. The quantitative

evaluation results are shown in Table 1. Task 3 of the

CHAOS challenge (Kavur et al., 2019) was for MRI

liver segmentation and the best model reported a dice

coefﬁcient of 0.95 and the average of all the models

was 0.86 (Kavur et al., 2020). While the proposed

model achieved a dice coefﬁcient (0.88) higher than

the average, it failed to perform better than the best

result. All models in the challenge, however, used the

available ground truth segmentation and were trained

in a supervised manner. Also all models were trained

using all three different types of MRIs available in

the CHAOS dataset (types are discussed in Sect. 2.1),

whereas the proposed non-supervised model was only

trained and tested on T1-DUAL in-phase images. It

can be observed that the vessels, which were consid-

ered part of the liver by the rater during manual seg-

mentation, except for one (as can be seen in Figure 4),

were not included by the proposed network.

Table 1: Quantitative analysis of the performance (only for

the ROI).

Metric Values

Intersection over Union (IoU) 0.7885

Dice Coefﬁcient 0.8812

4 FUTURE WORK

This paper stands as a proof of concept for unsu-

pervised biomedical image segmentation using the

proposed 3D Attention W-Net. Further tests will be

performed to evaluate the robustness of the approach

as well as the clinical applicability. This approach

will also be compared against other unsupervised seg-

mentation methods. Only the T1-DUAL in-phase vol-

umes were used, even though the CHAOS dataset also

contains T1-DUAL opposed-phase and T2-SPIR vol-

umes. Evaluation of the performance with the other

available contrasts may further improve the results.

A mixed training approach combining T1-DUAL

in-phase, opposed-phase and T2-SPIR is a fur-

ther option.

In the presented approach, CRF was applied to

post-process the results. A direct inclusion of CRF

within the model before N-Cuts during training may

be beneﬁcial. A further option is a semi-supervised

version of the algorithm, applying pre-training (both

U-Nets separately) with a manually labeled small

dataset followed by unsupervised training as de-

scribed in this contribution.

5 CONCLUSION

In this work, we propose an extension of current deep

learning approaches (W-Net) for unsupervised seg-

mentation of non-Medical RGB to volumetric med-

ical image segmentation. The model was enhanced

by using attention gates and extended to a 3D atten-

tion W-Net. The results demonstrate that the proposed

model can be used for unsupervised segmentation of

medical images. However, further experiments are

needed to judge the robustness and generalizability of

the approach. One reason for the remaining devia-

tion from the manual segmentation may be that the

ground truth images supplied in the dataset provide

liver segmentation including liver vessels. These were

not included by our unsupervised approach but were

naturally included as part of the liver by the super-

vised network. Our proposed network correctly seg-

mented the liver without inclusion of these vessels.

Thus unsupervised learning may be used to enrich or

guide manual expert annotation. Future research on

the learning approach itself will include end-to-end

training by incorporating conditional random ﬁelds in

the training pipeline. We expect that pre-training both

U-Nets of the W-Net separately on a small ground

truth set in a supervised manner may also further im-

prove the results.

ACKNOWLEDGEMENTS

This work was in part conducted within the context

of the International Graduate School MEMoRIAL at

the Otto von Guericke University (OVGU) Magde-

burg, Germany, kindly supported by the European

Structural and Investment Funds (ESF) under the pro-

gramme ”Sachsen-Anhalt WISSENSCHAFT Inter-

nationalisierung” (project no. ZS/2016/08/80646).

ICPRAM 2021 - 10th International Conference on Pattern Recognition Applications and Methods

492

Figure 3: Example slice of a test volume: From left to right - Original slice, Ground Truth, only the liver segment, Output of

the network - Clusters containing the liver segmentation were considered as our region of interest.

Figure 4: Example slice of a test volume: From left to right - Original slice, Ground Truth, only the liver segment, Output of

the network - Clusters containing the liver segmentation were considered as our region of interest.

REFERENCES

Aganj, I., Harisinghani, M. G., Weissleder, R., and Fischl,

B. (2018). Unsupervised medical image segmentation

based on the local center of mass. Scientiﬁc reports,

8(1):13012.

Anjna, E. A. and Er, R. K. (2017). Review of image seg-

mentation technique. International Journal of Ad-

vanced Research in Computer Science, 8(4).

Badrinarayanan, V., Kendall, A., and Cipolla, R. (2017).

Segnet: A deep convolutional encoder-decoder ar-

chitecture for image segmentation. IEEE transac-

tions on pattern analysis and machine intelligence,

39(12):2481–2495.

Baillard, C., Hellier, P., and Barillot, C. (2001). Segmenta-

tion of brain 3d mr images using level sets and dense

registration. Medical image analysis, 5(3):185–194.

Chatterjee, S., Bickel, M., and Krishna, S. (2020).

soumickmj/denseinferencewrapper: Initial release.

Chaurasia, A. and Culurciello, E. (2017). Linknet: Exploit-

ing encoder representations for efﬁcient semantic seg-

mentation. In 2017 IEEE Visual Communications and

Image Processing (VCIP), pages 1–4. IEEE.

Chen, L.-C., Papandreou, G., Kokkinos, I., Murphy, K., and

Yuille, A. L. (2017). Deeplab: Semantic image seg-

mentation with deep convolutional nets, atrous convo-

lution, and fully connected crfs. IEEE transactions on

pattern analysis and machine intelligence, 40(4):834–

848.

Christ, P. F., Ettlinger, F., Gr

un, F., Elshaera, M. E. A.,

Lipkova, J., Schlecht, S., Ahmaddy, F., Tatavarty, S.,

Bickel, M., Bilic, P., et al. (2017). Automatic liver

and tumor segmentation of ct and mri volumes using

cascaded fully convolutional neural networks. arXiv

preprint arXiv:1702.05970.

Crum, W. R., Scahill, R. I., and Fox, N. C. (2001). Au-

tomated hippocampal segmentation by regional ﬂuid

registration of serial mri: validation and application in

alzheimer’s disease. Neuroimage, 13(5):847–855.

Gee, J. C., Reivich, M., and Bajcsy, R. (1993). Elastically

deforming a three-dimensional atlas to match anatom-

ical brain images. University of Pennsylvania Institute

for Research in Cognitive Science Technical Report

No. IRCS-93-37.

Ghosh, S., Das, N., Das, I., and Maulik, U. (2019). Under-

standing deep learning techniques for image segmen-

tation. ACM Computing Surveys (CSUR), 52(4):73.

Gotra, A., Sivakumaran, L., Chartrand, G., Vu, K.-N.,

Vandenbroucke-Menu, F., Kauffmann, C., Kadoury,

S., Gallix, B., de Guise, J. A., and Tang, A. (2017).

Upgraded W-Net with Attention Gates and Its Application in Unsupervised 3D Liver Segmentation

493

Liver segmentation: indications, techniques and fu-

ture directions. Insights into imaging, 8(4):377–392.

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.

Kavur, A. E., Gezer, N. S., Barıs¸, M., Conze, P.-H., Groza,

V., Pham, D. D., Chatterjee, S., Ernst, P.,

Ozkan, S.,

Baydar, B., Lachinov, D., Han, S., Pauli, J., Isensee,

F., Perkonigg, M., Sathish, R., Rajan, R., Aslan, S.,

Sheet, D., Dovletov, G., Speck, O., N

urnberger, A.,

Maier-Hein, K. H., Akar, G. B.,

Unal, G., Dicle, O.,

and Selver, M. A. (2020). Chaos challenge – com-

bined (ct-mr) healthy abdominal organ segmentation.

Kavur, A. E., Selver, M. A., Dicle, O., Barıs¸, M., and Gezer,

N. S. (2019). CHAOS - Combined (CT-MR) Healthy

Abdominal Organ Segmentation Challenge Data.

ahenb

uhl, P. and Koltun, V. (2011). Efﬁcient inference in

fully connected crfs with gaussian edge potentials. In

Advances in neural information processing systems,

pages 109–117.

Larkin, K. G. (2015). Structural similarity index ssimpli-

ﬁed: Is there really a simpler concept at the heart

of image quality measurement? arXiv preprint

arXiv:1503.06680.

Lu, L., Shin, Y., Su, Y., and Karniadakis, G. E. (2019). Dy-

ing relu and initialization: Theory and numerical ex-

amples. arXiv preprint arXiv:1903.06733.

Nie, D., Wang, L., Gao, Y., and Shen, D. (2016). Fully con-

volutional networks for multi-modality isointense in-

fant brain image segmentation. In 2016 IEEE 13Th in-

ternational symposium on biomedical imaging (ISBI),

pages 1342–1345. IEEE.

Oktay, O., Schlemper, J., Folgoc, L. L., Lee, M., Heinrich,

M., Misawa, K., Mori, K., McDonagh, S., Hammerla,

N. Y., Kainz, B., et al. (2018). Attention u-net: Learn-

ing where to look for the pancreas. arXiv preprint

arXiv:1804.03999.

Paszke, A., Chaurasia, A., Kim, S., and Culurciello, E.

(2016). Enet: A deep neural network architecture

for real-time semantic segmentation. arXiv preprint

arXiv:1606.02147.

Shi, J. and Malik, J. (2000). Normalized cuts and image

segmentation. Departmental Papers (CIS), page 107.

Xia, X. and Kulis, B. (2017). W-net: A deep model for fully

unsupervised image segmentation. arXiv preprint

arXiv:1711.08506.

Zheng, S., Jayasumana, S., Romera-Paredes, B., Vineet, V.,

Su, Z., Du, D., Huang, C., and Torr, P. H. (2015). Con-

ditional random ﬁelds as recurrent neural networks. In

Proceedings of the IEEE international conference on

computer vision, pages 1529–1537.

ICPRAM 2021 - 10th International Conference on Pattern Recognition Applications and Methods

494