融合Resnet50和U-Net的眼底彩色血管图像分割

doi:10.16180/j.cnki.issn1007-7820.2021.08.004

摘要/Abstract

摘要：

眼底血管图像特征信息复杂度较高且现有算法对视网膜微血管分割不足,对病灶区域会产生误分割。针对以上问题,文中提出了一种融合Resnet50和U-Net的眼底彩色血管图像分割模型。通过数据增强提高数据集的数量以增加算法的鲁棒性。融合高斯双边滤波和限制对比度直方图均衡化来增强视网膜微细管的特征信息。利用自适应 Gamma 矫正提升图像亮度信息并降低图像伪影的干扰。在编码部分使用残差网路对彩色图像进行深层次的特征提取,将得到的特征图进行4次上采样和跳跃连接,实现像素级的分割。在每个卷积层使用ReLU激活函数和批量归一化来提高模型的性能。最后,利用公开数据集DRIVE进行实验,结果显示文中算法的特异性为0.989 2,准确率为0.967 6,灵敏度为0.812 0。

关键词: 眼底血管, 神经网络, 数据增强, 图像预处理, 双边滤波, 特征提取, 图像分割, 残差网络, U-net

Abstract:

The characteristic information of the fundus blood vessel image has high complexity, and the existing algorithms have the problems of insufficient retinal microvascular segmentation and wrong segmentation of the focus area. In view of these problems,a fundus color blood vessel image segmentation model combined RESNET 50 and U-Net is proposed. Through data enhancement, the number of datasets is increased, so as to improve the robustness of the algorithm. The adaptive Gamma correction is used to improve the brightness information and reduce the interference of the image artifacts. In the coding part, the residual network is used to extract the deep-seated features of the color image. After four times up-sampling and skip-connection of the feature map, pixel level segmentation is achieved. In each convolution layer, the ReLU activation function and batch normalization are used to improve the performance of the model. A large number of experiments in the open dataset DRIVE show that the specificity of the proposed algorithm is 0.989 2, the accuracy is 0.967 6, and the sensitivity is 0.812 0.

Key words: retinal blood vessel, neural network, data enhancement, image preprocessing, bilateral filtering, feature extraction, image segmentation, the residual network, U-net

中图分类号:

TP391.4

司明明,陈玮,胡春燕,尹钟. 融合Resnet50和U-Net的眼底彩色血管图像分割[J]. 电子科技, 2021, 34(8): 19-24.

SI Mingming,CHEN Wei,HU Chunyan,YIN Zhong. Fundus Blood Vessel Image Segmentation Combining Resnet50 and U-Net[J]. Electronic Science and Technology, 2021, 34(8): 19-24.

图/表 7

图1

图2

图3

图4

图5

图6

表1

参考文献 19

[1]	张亮军, 承垠林, 马梦楠, 等. 基于Dense U-Net方法的眼底彩色照片图像血管分割研究[J]. 生物医学工程研究, 2019,38(4):397-402.
	Zhang Liangjun, Cheng Yinlin, Ma Mengnan, et al. Research oncolor fundus image blood vessel segmentation based on Dense U-Net[J]. Journal of Biomedical Engineering Research, 2019,38(4):397-402.
[2]	梁礼明, 黄朝林, 石霏, 等. 融合形状先验的水平集眼底图像血管分割[J]. 计算机学报, 2018,41(7):1678-1692.
	Liang Liming, Huang Chaolin, Shi Fei, et al. Fusion of shape priori level set for fundus image vascular segmentation[J]. Journal of Computer Science, 2018,41(7):1678-1692.
[3]	Li Q, You J, Zhang D. Vessel segmentation and width estimation in retinal images ussing multi- scale production of matched filter responses[J]. Expert Systems with Appl- ications, 2012,39(9):7600-7610.
[4]	Fraz M M, Barman S A, Remagnino P, et al. An approach to localize the retinal blood vessels using bit planes and centerline detection[J]. Computer Methods and Programs in Biomedicine, 2012,108(2):600-616.
[5]	贝琛圆, 于海滨, 潘勉, 等. 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, et al. Gland cell image segmentation algorithm based on improved U-Net network[J]. Electronic Science and Technology, 2019,32(11):18-22.
[6]	Orlando J I, Prokofyeva E, Blaschko M B. A discri-minatively trained fully connected conditional rand field model for blood vessel Segmentation in fundus images[J]. IEEE Trans on Biomedical Engineering, 2016,64(1):16-27.
[7]	Staal J, Abramoff M D, Niemeijer M, et al. Ridgebased vessel segmentation in color images of the retina[J]. IEEE Transactions on Medical Imaging, 2004,23(4):501-509.
[8]	朱承璋, 邹北骥, 向遥, 等. 彩色眼底图像视网膜血管分割方法研究进展[J]. 计算机辅助设计与图形学学报, 2015,27(11):2046-2057.
	Zhu Chengzhang, Zou Beiji, Xiang Yao, et al. A survey of retinal vessel segmentation in fundus images[J]. Journal of Computer-Aided Design & Computer Graphics, 2015,27(11):2046-2057.
[9]	储清翠, 王华彬, 陶亮. 图像的局部自适应 Gamma校正[J]. 计算机工程与应用, 2015,51(7):189-193.
	Chu Qingcui, Wang Huabin, Tao Liang. Local adaptive Gamma correction of images[J]. Computer Engineering and Applications, 2015,51(7):189-193.
[10]	张利欣, 车世界, 徐正光, 等. 基于残差网络的高温合金微观组织图像分割方法[J]. 科学技术与工程, 2020,20(1):246-251.
	Zhang Lixin, Che Shijie, Xu Zhengguang, et al. Microstructure image segmentation of superalloy based on resnunet[J]. Science Technology and Engineering , 2020,20(1):246-251.
[11]	梁礼明, 盛校棋, 郭凯, 等. 基于改进的U-Net眼底视网膜血管分割[J]. 计算机应用研究, 2020,37(4):1247-1251.
	Liang Liming, Sheng Xiaoqi, Guo Kai, et al. Improved U-Net fundus retinal vessels segmentation[J]. Application Research of Computers, 2020,37(4):1247-1251.
[12]	Roychowdhury S, Koozekanani D D, Parhi K K. Blood vessel segmentation of fundus images by major vessel extraction and subimage classification[J]. IEEE Journal of Biomed Health Information, 2017,19(3):1118-1128
[13]	王婷, 李航, 胡智. 一种VGG Net的图像风格迁移算法设计与实现[J]. 计算机应用与软件, 2019,36(11):224-228.
	Wang Ting, Li Hang, Hu Zhi. Design and implementatio n of image style migration algorithm based on VGG Net[J]. Computer Applications and Softwar, 2019,36(11):224-228.
[14]	Cheng E, Du L, Wu Y, et al. Discriminative vessel segmentation in retinal images by fusing contextaware hybrid features[J]. Machine Vision an Applications, 2014,25(7):1779-1792.
[15]	Li Q L, Feng B, Xie L P, et al. A cross-modality learning approach for vessel segmentation in retinal images[J]. IEEE Transactions on Medical Imaging, 2016,35(1):109-118.
[16]	Strisciuglio N, Azzopardi G, Vento M, et al. Supervised vessel delineation in retinal fundus images with the automatic selection of B-COSFIRE filters[J]. Machine Vision & Applications, 2016,27(8):1137-1149.
[17]	Liskowski P, Krzysztof K. Segmenting retinal blood vesselswith deep neural networks[J]. IEEE Transactions on Medical Imaging, 2016,35(11):2369-2380.
[18]	吴晨玥, 易本顺, 章云港, 等. 基于改进卷积神经网络的视网膜血管图像分割[J]. 光学学报, 2018,38(11):133-139.
	Wu Chenyu, Yi Benshun, Zhang Yungang, et al. Retinal vascular image segmentation based on improved convolution neural network[J]. Journal of Optics, 2018,38(11):133-139.
[19]	Gao X R, Cai Y H, Qiu C Y, et al. Retinal blood vessel segmentation based on the Gaussian matched filter and U-net[C]. Shanghai:The Tenth International Congress on Image and Signal Processing,Biomedical Engineering and Informatics, 2017.

算法来源	年份	灵敏度	特异性	准确率
文献[14]	2014	0.725 2	0.979 8	0.947 4
文献[15]	2016	0.756 9	0.981 6	0.952 7
文献[16]	2016	0.773 1	0.972 4	0.946 7
文献[17]	2016	0.766 3	0.976 8	0.949 5
文献[6]	2017	0.789 7	0.968 4	-
文献[18]	2018	0.806 3	0.977 8	0.955 6
文献[19]	2018	0.780 2	0.987 6	0.963 6
本文算法	2020	0.812 0	0.989 2	0.967 6