Journal of Physics: Conference Series

PAPER • OPEN ACCESS

Image Segmentation Based on Improved Unet

To cite this article: Xiaojin Li et al 2021 J. Phys.: Conf. Ser. 1815 012018

View the article online for updates and enhancements.

This content was downloaded from IP address 45.13.249.58 on 25/02/2021 at 03:35

CCME 2020 IOP Publishing
Journal of Physics: Conference Series 1815 (2021) 012018 doi:10.1088/1742-6596/1815/1/012018

Image Segmentation Based on Improved Unet

Xiaojin Li, Wenhua Qian*, Dan Xu and Chunyu Liu

School of Information, Yunnan University, Kunming, 650504, China

*Corresponding author email: whqian@ynu.edu.cn

Abstract. In order to help doctors diagnose and treat liver lesions and accurately segment liver
images, this paper proposes an improved Unet network, which adds compression extraction
modules and full-scale connection blocks, extracts input image features, and achieves accurate
segmentation of liver images. The compression extraction module distributes weights to
convolutional layers of different sizes, which is conducive to the extraction of image spatial
information and context information. Full-scale blocks are connected by skipping,combining the
higher semantic information from the decoder and corresponding the lowwer semantic
information from the encoder to strengthen the ability to extract tumor edge information. This
article includes 25 cases from the Lits liver dataset. The dataset is classified as the training dataset
and the test dataset, and the image blocks are extracted after gray-scale normalization and input
to the network to acquire the final segmentation results. The segmentation result is evaluated by
F1 score. Comparing multiple sets of experiments, compared with general network structures
such as Unet and AttenUnet, it shows that the network architecture proposed in the Dissertation
improves the accuracy and efficiency of liver image segmentations.
Keywords: Gray level normalization; Dice score; Segmentation accuracy; The liver.

1. Introduction

1.1. Background
Liver cancer is the deadliest cancers, its mortality rate ranks second in the world, and it is an important
factor leading to human death [1]. Generally, early diagnosis of liver cancer can significantly reduce
patient mortality. Medical image segmentation of liver is a hot research topic at present. In modern
pathology, liver images rely on MRI and CT scans, and different equipment or technologies are used to
collect images of the liver.Traditional automatic liver segmentation methods are based on thresholds,
regions, edges, etc. [2-4]. These methods have a great influence on subjective factors, which make the
traditional segmentation algorithm unable to accurately locate the liver position and cannot fully
describe the liver image features, resulting in poor segmentation accuracy, deviations and errors. The
segmentation method based on convolutional neural network [5] (CNN) can provide the latest accuracy
for different computer vision problems, and has achieved success in different modes in the field of the
imaging segmentation, including lung segmentation, diabetes recognition, retina Blood vessel
segmentation, brain tumor detection, etc. [6-7]

1.2. Research Work

In 2016, Ben-Cohen et al.[8]used FCN to solve the problem of liver segmentation. Unet uses FCN as
the backbone to fully excavate the multi-scale features of the image. Seo et al.added residual paths on
the basis of Unet to improve the accuracy of liver segmentation. The above-mentioned neural network
realizes the automatic segmentation of liver by independently classifying pixels in the segmentation
Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
CCME 2020 IOP Publishing
Journal of Physics: Conference Series 1815 (2021) 012018 doi:10.1088/1742-6596/1815/1/012018

process, but there are certain defects, which are mainly manifested in the poor segmentation accuracy
of target boundaries and small-volume targets.

1.3. Organizational Structure

The structure of this paper is as follows: Section two mainly introduces the Network structure, and
Section three introduces the experiment, including Data set and Evaluation index. Section four
introduces the conclusion of this article.

2. Network Structure
The network model proposed in the paper is shown in Figure 1. The model is mainly composed of
compressed extraction block, scale connection block, and down-sampling block. The compression
extraction module focuses on the relationship between channels, the model can automatically learn the
importance of different channel features, and the full-scale connection block connects the high-level
compression extraction block and the low-level full-scale connection block through a complete jump,
and merges The small-scale feature map in the encoder, the same-scale feature map, and the largest scale
feature map from the decoder capture the low-level and the higher level semantic information at the full
scale, and restore the feature map. The model encoder path consists of 5 cascaded compressed extraction
blocks x1, x2, x3, x4, x5 and 4 down-sampling layers, each down-sampling layer is composed of a 3x3
convolutional layer with a step size of 2. The decoder consists of 4 full-scale connection blocks.

concat connection The sampling layer

Figure 1. Architecture of the fully connected attention Unet network.
Because Unet relies on fixed weights in the process of image feature extraction, a compressed extraction
block[9] is introduced to distribute the weights of feature extraction to all stages of the path. The
compressed extraction block is mainly composed of the compressed part and the extracted part. The
importance of the channel is predicted, and the importance of different channels is obtained, and then it
is fused to the corresponding channel of the previous feature map. Figure 2 shows that the input image
U (h, w, c) is first made the global average pooling Fsq, and the output data is passed through two-level
fully connected Fex, and finally limited to the range of [0, 1] with the sigmoid activation function. The
value as Fscale is multiplied by the channel c of the input image as the input data of the next stage.

Figure 2. Compressed extraction module diagram.

Figure 3 depicts the x1 full-scale connection[10] process; among them, x1, x2, and x3 pool down the
encoder layer through different maximum pooling operations, so as to transfer the low-level semantic
information of the bottom layer and unify the size of the feature map. x5 is up-sampling by bilinear
interpolation to enlarge the resolution of the feature map. In order to unify the size of the feature map,
x1 must be reduced by 8 times, x2 must be reduced by 4 times, x3 must be reduced by 2 times, and x5
must be enlarged by 2 times. Then, the cascaded feature maps of the five scales of x1, x2, x3, x4, and
x5 are convolved through 64 3*3 convolution kernels in order to merge the shallow semantic
informations and the deep information, and execute the feature Aggregation mechanism. Finally,

2
CCME 2020 IOP Publishing
Journal of Physics: Conference Series 1815 (2021) 012018 doi:10.1088/1742-6596/1815/1/012018

through the channel feature fusion of Unet, and through 3*3 convolution and Relu operations, a new
feature map x1 is obtained.

The sampling layer pooling and convolution

Figure 3. Full scale connection block diagram.
The model in this paper uses binary cross entropy (BCE) as the loss function. The binary cross entropy
formula is:
lb c e    G T ( a , b ) lo g ( S E G ( a , b ) ) 
( a ,b )

(1  G T ( a , b ) ) l o g (1  S E G ( a , b ) ) 
(1)
Among them, is the expert label of pixel (a, b) and is the predicted probability of the segmented object.
The BCE loss function is pixel by pixel. It does not consider the labels of the neighborhood, and weights
both the segmented pixels and the background pixels, which helps the convergence of the loss function.

3. Experiment

3.1. Data Set

This paper selects 25 case images from the LITS2017 [11] data set for experiments. Each patient image
contains a liver image and a real segmentation label image. Among them, 20 cases are used as the
training set for training, and 5 cases are used as the test set, and the image size is 512×512. The expert's
annotation results are used as the gold standard for training and testing.

Figure 4. LITS2017 dataset.

The resolution of the image is uniformly cut to 256X256.Because the contrast of the image is not the
same, the Z-Score method is used to standardize each image, and the data of different magnitudes are
unified into the same magnitude, and the calculated Z -Score value measurement to ensure the
comparability between data. Use formula (2) to calculate:
x
x (2)
2a

3
CCME 2020 IOP Publishing
Journal of Physics: Conference Series 1815 (2021) 012018 doi:10.1088/1742-6596/1815/1/012018

Among them, μ is the average value of the data, a is the deviation of the overall data, and x is the
individual observation value.

3.2. Evaluation Index

For the sake of better experiment, verify the segmentation effect, enhance the data, rotate the image,
interchange left and right, zoom in and out, and expand the image.
F1 score is an important indicator to measure the accuracy of two classification problems, and it is the
harmonic average of recall and precision. When using F1 score to evaluate model accuracy, the
classification result of each pixel will affect the score. The higher the F1 score, the better the accuracy
of the resulting change map. Calculate using formula (3)(4)(5).
TP
Precision  (3)
TP  FP
TP
Recall  (4)
TP  FN
p recisio n  reca ll
F1  2  (5)
p recisio n  reca ll

3.3. The Experimental Results

In order to prove that the network has better results for liver segmentation, Unet, AttenUnet[12], the
network of this paper are used to simultaneously segment the Lits data set. Figures 5 show the
comparison between the test results of different networks for liver segmentation and the manual
annotation results of experts. From top to bottom, each row is the real segmentation label, the
segmentation result of this article, the Unet segmentation result, and the Attention Unet segmentations
result. It can be seen from Figures five that Unet leads to the lack of liver boundary details and obvious
over-segmentation. The liver boundary of AttentionUnet segmentation is more obvious, but the
segmented liver boundary edge area still has unsmooth links. The segmentation method in this paper
strengthens the image boundary to obtain a segmentation result similar to the segmentation label, and
the result at the liver image point is more substantial , With high accuracy, effectively refine the
boundaries of the liver. Solve the over-segmentation and under-segmentation problems of other
algorithms.

Ground This paper Unet AttenUnet

Ground This paper Unet AttenUnet

Figure 5. Segmented image.
Table 1 compares the Unet, Attention Unet and the network structure of this article for liver
segmentation experiments. It can be known that under the evaluation of relevant indicators, the method
proposed in the paper performs better than other network models in liver image segmentation. The lits
data set achieves an F1 score of 94.4306%. This article will improve The attention block is embedded
in the framework of Unet network, and the high-order consistency between pixel categories is learned

4
CCME 2020 IOP Publishing
Journal of Physics: Conference Series 1815 (2021) 012018 doi:10.1088/1742-6596/1815/1/012018

through training, which effectively improves the segmentation accuracy of small scale objective and
target boundaries.
Table 1. The evaluation of different network structures on the Lits test data set.

Method F1 score

Unet 0.943854
0.944306
AttenUnettt[16]

This paper
0.945211
patarticle

4. Conclusion
To solve the problems of missing information and blurring edge in liver segmentation, this paper
proposes an improved Unet network architecture for attention mechanism. The feature maps of different
levels are fused to guide the feature maps to learn feature expression. The classification task and the
segmentation task are combined through the mixed loss function to retain high-level and low-level
semantic information and improve the segmentation accuracy.

Acknowledgement
This research was funded by the grants (No. 61662087, 61761046) from the Research Natural Science
Foundation of China, the Research Foundation of Yunnan Province (No. 2019FA044), Provincial
Foundation for Leaders of Disciplines in Science and Technology (2019HB121), the Postdoctoral fund
of the Ministry of education of China (No. 2017M621591), Jiangsu Planned Projects for Postdoctoral
Research Funds in 2017.

References
[1] Bosch F X, Ribes J, Mireia Díaz, et al. Primary liver cancer: worldwide incidence and
trends.[J].Gastroenterology, 2004, 127(5-supp-S1):S5-S16.
[2] Zheng Zhou, Zhang Xuechang, Zheng Siming, et al. CT liver image segmentation based on region
growth and unified level set[J]. Journal of Zhejiang University (Engineering Science Edition),
2018, 52(12):145-159.
[3] Liu J , Wang S , Linguraru M G , et al. Tumor sensitive matching flow: A variational method to
detecting and segmenting perihepatic and perisplenic ovarian cancer metastases on contrast-
enhanced abdominal CT.[J]. Medical Image Analysis, 2014, 18(5):725-739.
[4] Linguraru M G , Richbourg W J , Liu J , et al. Tumor Burden Analysis on Computed Tomography
by Automated Liver and Tumor Segmentation[J].IEEE Transactions on Medical Imaging, 2012,
31(10):1965-1976.
[5] Lecun Y, Boser B, Denker J, et al. Backpropagation Applied to Handwritten Zip Code
Recognition[J].Neural Computation, 2014, 1(4):541-551.
[6] Pang S , Du A , Orgun M A , et al. CTumorGAN: a unified framework for automatic computed
tomography tumor segmentation[J]. European Journal of Nuclear Medicine and Molecular
Imaging, 2020, 47(10):2248-2268.
[7] Kumar Singh V , Rashwan H A , Romani S , et al. Breast tumor segmentation and shape
classification in mammograms using generative adversarial and convolutional neural network[J].
Expert Systems with Application, 2020, 139(Jan.):112855.1-112855.14.
[8] Long J,Shelhamer E,Darrell T.Fully Convolutional Networks for Semantic
Segmentation[J].IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015,
39(4):640-651.
[9] Jie, Hu, Li, et al. Squeeze-and-Excitation Networks.[J]. IEEE transactions on pattern analysis and
machine intelligence, 2019.

5
CCME 2020 IOP Publishing
Journal of Physics: Conference Series 1815 (2021) 012018 doi:10.1088/1742-6596/1815/1/012018

[10] Huang H , Lin L , Tong R , et al. UNet 3+: A Full-Scale Connected UNet for Medical Image
Segmentation[C]// ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech
and Signal Processing (ICASSP). IEEE, 2020.
[11] Bilic, Patrick , et al. "The Liver Tumor Segmentation Benchmark (LiTS)." (2017).
[12] Oktay O , Schlemper J , Folgoc L L , et al. Attention U-Net: Learning Where to Look for the
Pancreas[J]. 2018.