基于改进U-Net网络的腺体细胞图像分割算法

doi:10.16180/j.cnki.issn1007-7820.2019.11.004

摘要/Abstract

摘要：

针对腺体图像在自动分割过程中由于多尺度目标和信息丢失影响导致准确率降低的问题,文中采用了一种引入注意力模块的全卷积神经网络模型。该模型遵循编码器-解码器结构,在编码网络中用空洞残差卷积层代替原有的普通卷积层,并添加空洞金字塔池;再在解码网络中加入注意力模块,使模型输出高分辨率特征图,提高对多尺度目标的分割精度。实验结果表明,提出的网络模型参数少分割精度高,对腺体图像的平均分割精度高达89.7%,具有较好的鲁棒性。

关键词: 全卷积神经网络, 编码器-解码器结构, 空洞金字塔池, 注意力模块, 高分辨率特征图, 分割精度高

Abstract:

This paper proposed a full convolutional neural network model with attention module to solve the problem that multi-scale targets and information loss affect the segmentation accuracy of gland images in the automatic segmentation process. This model followed the encoder-decoder structure. Firstly, the atrous spatial pyramid pooling was added to the encoder path, and the original residual convolution layer was replaced by the atrous residual convolution layer in the encoder path. Secondly, the attention module was added to the decoder path to make the model output high-resolution feature map and improve the segmentation accuracy of the multi-scale object. The experimental results showed that the proposed network model had fewer parameters, high segmentation precision and good robustness, besides, the average segmentation accuracy of gland images was as high as 89.7%.

Key words: full convolutional neural network, encoder-decoder structure, atrous spatial pyramid pooling, attention module, high-resolution feature map, high segmentation precision

中图分类号:

TP391

贝琛圆,于海滨,潘勉,蒋洁,吕炳赟. 基于改进U-Net网络的腺体细胞图像分割算法[J]. 电子科技, 2019, 32(11): 18-22.

BEI Chenyuan,YU Haibin,PAN Mian,JIANG Jie,LÜ Bingyun. Gland Cell Image Segmentation Algorithm Based on Improved U-Net Network[J]. Electronic Science and Technology, 2019, 32(11): 18-22.

图/表 7

图1

图2

图3

表1

图4

表2

表3

参考文献 27

[1]	Winawer S J, Fletcher R H, Miller L , et al. Colorectal cancer screening: clinical guidelines and rationale[J]. Gastroenterology, 1997,113(4):594-642.
[2]	陈爱斌, 江霞 . 细胞分割算法研究方法综述[J].电子世界 2011(15):76-79.
	Chen Aibin, Jiang Xia . Review of research methods for cell segmentation algorithms[J].Electronic World, 2011(15):76-79.
[3]	Elston C W, Ellis I O . Pathological prognostic factors in breast cancer. I. The value of histological grade in breast cancer: experience from a large study with long-term follow-up[J]. Histopathology, 2010,19(5):403-410.
[4]	Gleason D F . Histologic grading of prostate cancer:A perspective[J]. Human Pathology, 1992,23(3):273-279.
[5]	Fleming M, Ravula S, Tatishchev S F , et al. Colorectal carcinoma: Pathologic aspects[J]. Journal of Gastrointestinal Oncology, 2012,3(3):153-153.
[6]	章毓晋 . 图像处理和分析基础[M]. 北京: 高等教育出版社, 2002.
	Zhang Yujin. The basis of image processing and analysis [M]. Beijing: Higher Education Press, 2002.
[7]	Gurcan M N, Boucheron L E, Can A , et al. Histopathological image analysis: a review[J].IEEE Reviews in Biomedical Engineering 2009(2):147-147.
[8]	Altunbay D, Cigir C, Sokmensuer C , et al. Color graphs for automated cancer diagnosis and grading[J]. IEEE Transactions on Biomedical Engineering, 2010,57(3):665-665.
[9]	Gunduz Demir C, Kandemir M, Tosun A B , et al. Automatic segmentation of colon glands using object-graphs[J]. Medical Image Analysis, 2010,14(1):1-12. doi: 10.1016/j.media.2009.09.001
[10]	Fu H, Qiu G, Shu J , et al. A novel polar space random field model for the detection of glandular structures[J]. IEEE Transactions on Medical Imaging, 2014,33(3):764-764. doi: 10.1109/TMI.2013.2296572
[11]	Khasawneh S, Al Wahadni A, Lloyd C H . A stochastic polygons model for glandular structures in colon histology images[J]. IEEE Transactions on Medical Imaging, 2015,34(11):2366-2378.
[12]	Ronneberger O, Fischer P, Brox T. U-Net:convolutional networks for biomedical image segmentation[M]. Berlin: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015,Springer International Publishing, 2015.
[13]	Chen H, Qi X, Yu L , et al. DCAN:Deep contour-aware networks for object instance segmentation from histology images[J]. Medical Image Analysis, 2017,36(9):135-146.
[14]	Xu Y, Li Y, Liu M, et al. Gland instance segmentation by deep multichannel side supervision [C].Athens: International Conference on Medical Image Computing and Computer-Assisted Intervention,Springer,Cham, 2016.
[15]	Xu Y, Li Y, Wang Y , et al. Gland instance segmentation using deep multichannel neural networks[J]. IEEE Transactions on Biomedical Engineering, 2017,64(12):2901-2912.
[16]	Raza S E A, Cheung L, Epstein D, et al. MIMONet:gland segmentation using multi-input-multi-output convolutional neural network [C].Edinburgh:Annual Conference on Medical Image Understanding and Analysis,Springer,Cham, 2017.
[17]	Chen L C, Papandreou G, Kokkinos I , et al. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018,40(4):834-848.
[18]	Drelie Gelasca E, Obara B, Fedorov D , et al. A biosegmentation benchmark for evaluation of bioimage analysis methods[J]. BMC Bioinformatics, 2009,10(1):368-368.
[19]	Sirinukunwattana K, Pluim J P, Chen H , et al. Gland segmentation in colon histology images: The glas challenge contest[J]. Medical Image Analysis, 2017,35(8):489-502.
[20]	Lecun Y, Bengio Y, Hinton G . Deep learning[J]. Nature, 2015,521(7553):436-436.
[21]	Yu F, Koltun V, Funkhouser T A. Dilated residual networks [C].Hawaii:Conference on Computer Vision and Pattern Recognition, 2017.
[22]	Wang P, Chen P, Yuan Y , et al. Understanding convolution for semantic segmentation[C].California:IEEE Winter Conference on Applications of Computer Vision ( WACV), 2018.
[23]	Hu J, Shen L, Sun G . Squeeze-and-excitation networks[J]. IEEE Transaction on Network, 2017,28(4):523-539.
[24]	Liu Z, Li X, Luo P, et al. Semantic image segmentation via deep parsing network [C].Santiago:Proceedings of the IEEE International Conference on Computer Vision, 2015.
[25]	Oktay O, Schlemper J, Folgoc L L , et al. Attention U-Net: learning where to look for the pancreas[J]. IEEE Transaction on Network, 2018,29(2):117-129.
[26]	Abadi , Martín , Barham P, et al.TensorFlow:A system for large-scale machine learning[J]. IEEE Transaction on Network, 2016,27(9):1024-1037.
[27]	何菁, 陈胜 . 一种全新的两步自动化医学图像分割方案[J]. 电子科技, 2016,29(7):85-87.
	He Jing, Chen Sheng . A new two-step automated medical image segmentation scheme[J]. Electronic Science and Technology, 2016,29(7):85-87.

方法	F₁score
U-Net	0.788
U-Net \dilated CNN	0.836
U-Net \attention	0.832
U-Net \attention \dilated CNN	0.871
U-Net \ASPP	0.847
U-Net \ASPP \dilated CNN	0.885
U-Net \ASPP \attention \dilated CNN	0.897

	F₁score				Object Dice				Object Hausdorff				Sum
	A	Rank	B	Rank	A	Rank	B	Rank	A	Rank	B	Rank	Sum
Proposed DeepLab-v3 FCN-8 SegNet U-Net Freidburg1 CUMedVision1 Freidburg2 CVML ExB3 ExB2 ExB1 LIB vision4GlaS	0.897 0.858 0.783 0.862 0.788 0.834 0.868 0.870 0.652 0.896 0.892 0.891 0.777 0.635	1 8 11 7 10 9 6 5 13 2 3 4 12 14	0.832 0.753 0.692 0.764 0.697 0.605 0.769 0.695 0.541 0.719 0.686 0.703 0.306 0.527	1 4 9 3 7 11 2 8 12 5 10 6 14 13	0.885 0.864 0.795 0.859 0.781 0.875 0.867 0.876 0.644 0.886 0.884 0.882 0.781 0.737	2 8 10 9 11 6 7 5 14 1 3 4 12 13	0.825 0.807 0.767 0.804 0.781 0.783 0.800 0.786 0.654 0.765 0.754 0.786 0.617 0.610	1 2 9 3 8 7 4 5 12 10 11 5 13 14	54.20 62.62 105.04 65.72 102.47 57.19 74.60 57.09 155.43 57.36 54.79 57.41 112.71 107.49	1 7 11 8 10 4 9 3 14 5 2 6 13 12	119.93 118.51 147.28 124.97 143.75 146.61 153.65 148.47 176.24 159.87 187.44 145.58 190.45 210.10	2 1 7 3 4 6 9 8 11 10 12 5 13 14	8 30 57 33 50 43 37 34 76 33 41 30 77 80

	F1score						Sum
	score	rank	score	rank	score	rank	Sum
Proposed	0.725	1	0.775	1	240.14	1	3
Deeplab-v3	0.648	2	0.745	2	281.45	3	7
SegNet	0.622	3	0.739	3	247.84	2	7
U-Net	0.600	4	0.654	4	354.09	4	12
FCN-8	0.558	5	0.640	5	436.43	5	15