跨级融合门控自适应网络用于视网膜血管分割

梁礼明; 余洁; 陈鑫; 雷坤; 周珑颂

doi:10.13700/j.bh.1001-5965.2022.0410

跨级融合门控自适应网络用于视网膜血管分割

doi: 10.13700/j.bh.1001-5965.2022.0410

江西理工大学电气工程与自动化学院，赣州 341000

基金项目: 国家自然科学基金(51365017,61463018)；江西省自然科学基金面上项目(20192BAB205084)；江西省研究生创新专项资金(YC2021-S585)；江西省教育厅科学技术项目(GJJ170491,GJJ2200848)

详细信息

通讯作者:
E-mail：9119890012@ jxust.edu.cn

中图分类号: TP391
计量
- 文章访问数: 357
- HTML全文浏览量: 88
- PDF下载量: 23
- 被引次数: 0
出版历程
- 收稿日期: 2022-05-21
- 录用日期: 2022-08-05
- 网络出版日期: 2022-08-23
- 整期出版日期: 2024-04-29

Cross-level fusion gated adaptive network for retinal vessel segmentation

School of Electrical Engineering and Automation，Jiangxi University of Science and Technology，Ganzhou 341000，China

Funds: National Natural Science Foundation of China (51365017,61463018);General Project of Natural Science Foundation of Jiangxi, China (20192BAB205084); Jiangxi Province Graduate Innovation Special Fund (YC2021-S585); Science and Technology Project of Jiangxi Provincial Department of Education (GJJ170491,GJJ2200848)

More Information

Corresponding author: E-mail：9119890012@ jxust.edu.cn

摘要

摘要:
针对现有多数算法对浅层特征提取不足，导致分割结果中血管边界模糊、毛细血管欠分割且包含噪声等问题，提出一种跨级融合门控自适应网络。该网络中的密集门控通道变换模块，通过促进通道之间的竞争或协同关系充分提取浅层特征信息，避免浅层粗粒度特征信息丢失；通过跨层次融合模块捕获各层跨维度交互信息，有效聚合多尺度上下文特征；采用双自适应特征融合方法有效引导相邻层次特征融合，抑制噪声。在公共数据集DRIVE、CHASEDB1和STARE上进行验证，结果表明：所提网络准确率分别为0.9652、0.9668和0.9695，F₁值分别为0.8544、0.8152和0.8412，在多个指标上均处于较高水平，优于现有先进算法。
- 视网膜血管分割 /
- 密集门控通道变换 /
- 跨层次融合模块 /
- 双自适应特征融合 /
- 三重注意力模块
Abstract:
To address the insufficient shallow feature extraction of most existing algorithms, which results in noise, blurred vascular boundary and capillary under segmentation, a cross-level fusion gated adaptive network is proposed. Firstly, shallow feature information is fully extracted by the dense gated channel transformation module in the network with promotion of competition or cooperation of channels to avoid the loss of shallow coarse-grained feature information. Secondly, cross-dimensional interaction information of each layer is captured by cross-level fusion module to effectively aggregate multi-scale context features. Thirdly, dual adaptive feature fusion method is used to guide the feature fusion of adjacent layers effectively and suppress noise. The validation was performed on public data sets DRIVE, CHASEDB1 and STARE, and the accuracy rates were 0.9652, 0.9668 and 0.9695 respectively; the F₁ values were 0.8544, 0.8152 and 0.8412 respectively. The results show that the proposed network is at a high level in many indexes, and is superior to the existing advanced algorithms.
- retinal vessel segmentation /
- dense gated channel transformation /
- cross-level fusion module /
- dual adaptive feature fusion /
- triple attention module

HTML全文

图 1 GCT模块

Figure 1. GCT module

下载: 全尺寸图片幻灯片

图 2 DGCT模块

Figure 2. DGCT module

下载: 全尺寸图片幻灯片

图 3 三重注意力机制

Figure 3. Triple attention mechanism

下载: 全尺寸图片幻灯片

图 4 跨层次融合模块

Figure 4. Cross-level fusion module

下载: 全尺寸图片幻灯片

图 5 双自适应特征融合模块

Figure 5. Dual adaptive feature fusion module

下载: 全尺寸图片幻灯片

图 6 跨级融合门控自适应网络

Figure 6. Cross-level fusion gated adaptive network

下载: 全尺寸图片幻灯片

图 7 预处理各阶段效果

Figure 7. Effect of each stage of preprocessing

下载: 全尺寸图片幻灯片

图 8 不同算法在DRIVE数据集上ROC曲线与PR曲线对比

Figure 8. Comparison of ROC curves and PR curves of different algorithms on DRIVE dataset

下载: 全尺寸图片幻灯片

图 9 不同算法在CHASSEDB1数据集上ROC曲线与PR曲线对比

Figure 9. Comparison of ROC curves and PR curves of different algorithms on CHASSEDB1 dataset

下载: 全尺寸图片幻灯片

图 10 不同算法分割结果

Figure 10. Segmentation results of different algorithms

下载: 全尺寸图片幻灯片

图 11 分割结果细节对比

Figure 11. Details comparison of segmentation results

下载: 全尺寸图片幻灯片

图 12 模型改进分割图像结果对比

Figure 12. Comparison of image segmentation results of improved model

下载: 全尺寸图片幻灯片

表 1 不同注意力机制的参数复杂度和开销比较

Table 1. comparison of parameter complexity and overhead of various attentional mechanisms

注意力机制	参数量	存储开销/10⁶
SE	2C²/r	2.514
CBAM	2C²/r+2k²	2.532
全局上下文	2C²/r+C	2.548
TAM(本文)	6k²	0.004 8

下载: 导出CSV

表 2 不同数据集上的平均性能指标评估结果

Table 2. Average performance index evaluation results on different datasets

数据集	算法	F₁	A_cc	S_e	S_p	AUC(ROC)	AUC(PR)
DRIVE	U-Net^[13]	0.8474	0.9631	0.8302	0.9812	0.9855	0.9310
	Attention U-Net^[14]	0.8449	0.9627	0.8287	0.9814	0.9852	0.9299
	ASR-UNet^[15]	0.8507	0.9641	0.8351	0.9821	0.9862	0.9336
	CFGA-Net(本文)	0.8544	0.9652	0.8326	0.9837	0.9873	0.9379
CHASEDB1	U-Net^[13]	0.8002	0.9641	0.7934	0.9811	0.9834	0.8848
	Attention U-Net^[14]	0.8109	0.9658	0.8090	0.9813	0.9845	0.8942
	ASR-UNet^[15]	0.8061	0.9650	0.8047	0.9809	0.9848	0.8916
	CFGA-Net(本文)	0.8152	0.9668	0.8097	0.9824	0.9866	0.9026
STARE	U-Net^[13]	0.8192	0.9644	0.7591	0.9888	0.9843	0.9158
	Attention U-Net^[14]	0.8184	0.9646	0.7522	0.9898	0.9836	0.9161
	ASR-UNet^[15]	0.8239	0.9653	0.7649	0.9891	0.9853	0.9201
	CFGA-Net(本文)	0.8412	0.9695	0.8023	0.9886	0.9899	0.9347

下载: 导出CSV

表 3 不同算法在数据集DRIVE、CHASEDB1和STARE上的性能指标对比

Table 3. Comparison of performance indicators of different algorithms in DRIVE, CHASEDB1 and STARE datasets

算法	S_e			S_p			A_cc
算法	DRIVE	CHASEDB1	STARE	DRIVE	CHASEDB1	STARE	DRIVE	CHASEDB1	STARE
文献[16]	0.7653	0.7633	0.7581	0.9818	0.9809	0.9846	0.9542	0.9610	0.9612
文献[17]	0.7632	0.7815	0.7423				0.9536	0.9587	0.9603
文献[18]	0.7631	0.7641	0.7735	0.9820	0.9806	0.9857	0.9538	0.9607	0.9638
文献[19]	0.7918	0.6457	0.8021	0.9708	0.9653	0.9561	0.9577	0.9340	0.9445
文献[20]	0.7941	0.8176	0.7598	0.9798	0.9704	0.9878	0.9558	0.9608	0.9640
文献[21]	0.8213	0.8035		0.9807	0.9787		0.9615	0.9639
文献[22]	0.7352	0.7279	0.7265	0.9775	0.9658	0.9759	0.9480	0.9452	0.9548
文献[23]	0.8353	0.8176	0.7946	0.9751	0.9776	0.9821	0.9579	0.9632	0.9626
文献[24]	0.8125	0.8012	0.8078	0.9763	0.9730	0.9721	0.9610	0.9578	0.9586
CFGA-Net(本文)	0.8326	0.8097	0.8023	0.9837	0.9824	0.9886	0.9652	0.9668	0.9695

下载: 导出CSV

表 4 算法改进前后结果对比

Table 4. Comparison of results before and after improvement

模型	F₁			A_cc			S_e			S_p			AUC
模型	DRIVE	CHASEDB1	STARE	DRIVE	CHASEDB1	STARE	DRIVE	CHASEDB1	STARE	DRIVE	CHASEDB1	STARE	DRIVE	CHASEDB1	STARE
1	0.8474	0.8002	0.8192	0.9631	0.9641	0.9644	0.8302	0.7934	0.7591	0.9812	0.9811	0.9888	0.9855	0.9834	0.9843
2	0.8527	0.8139	0.8243	0.9648	0.9661	0.9659	0.8323	0.8193	0.7810	0.9833	0.9807	0.9873	0.9867	0.9861	0.9870
3	0.8530	0.8122	0.8241	0.9647	0.9659	0.9652	0.8296	0.8134	0.7770	0.9826	0.9811	0.9879	0.9822	0.9854	0.9860
4	0.8528	0.8136	0.8276	0.9649	0.9660	0.9665	0.8297	0.8223	0.7865	0.9839	0.9803	0.9874	0.9870	0.9860	0.9875
5	0.8530	0.8132	0.8331	0.9650	0.9661	0.9672	0.8290	0.8138	0.7839	0.9840	0.9813	0.9886	0.9855	0.9859	0.9888
6	0.8544	0.8152	0.8412	0.9652	0.9668	0.9695	0.8326	0.8097	0.8023	0.9837	0.9824	0.9886	0.9873	0.9866	0.9899

下载: 导出CSV

参考文献(24)

[1]	贾洪, 郑楚君, 李灿标, 等. 基于局部线结构约束的FCM聚类视网膜血管分割[J]. 光学学报, 2020, 40(9): 0910001. doi: 10.3788/AOS202040.0910001 JIA H, ZHENG C J, LI C B, et al. Retinal blood vessel segmentation based on fuzzy C-means clustering according to the local line structural constraints[J]. Acta Optica Sinica, 2020, 40(9): 0910001 (in Chinese). doi: 10.3788/AOS202040.0910001
[2]	周阳. 基于匹配滤波引导局部特征空间仿射传播聚类的视网膜血管分割[J]. 信息通信, 2019, 32(7): 35-39. ZHOU Y. Retinal vascular segmentation based on affine propagation clustering in local feature space guided by matched filtering[J]. Information & Communications, 2019, 32(7): 35-39(in Chinese).
[3]	DASH J, BHOI N. Retinal blood vessel extraction using morphological operators and Kirsch’s template[C]//Proceedings of the Soft Computing and Signal Processing. Berlin: Springer, 2019: 603-611.
[4]	汪维华, 张景中, 吴文渊. 改进的形态学与Otsu相结合的视网膜血管分割[J]. 计算机应用研究, 2019, 36(7): 2228-2231. WANG W H, ZHANG J Z, WU W Y. New approach to segment retinal vessel using morphology and Otsu[J]. Application Research of Computers, 2019, 36(7): 2228-2231(in Chinese).
[5]	LI X, JIANG Y C, LI M L, et al. Lightweight attention convolutional neural network for retinal vessel image segmentation[J]. IEEE Transactions on Industrial Informatics, 2021, 17(3): 1958-1967.
[6]	WANG C, ZHAO Z Y, REN Q Q, et al. Dense U-Net based on patch-based learning for retinal vessel segmentation[J]. Entropy, 2019, 21(2): 168. doi: 10.3390/e21020168
[7]	YUE K J, ZOU B J, CHEN Z L, et al. Retinal vessel segmentation using dense U-Net with multiscale inputs[J]. Journal of Medical Imaging, 2019, 6(3): 034004.
[8]	胡扬涛, 裴洋, 林川, 等. 空洞残差U型网络用于视网膜血管分割[J]. 计算机工程与应用, 2021, 57(7): 185-191. doi: 10.3778/j.issn.1002-8331.2009-0215 HU Y T, PEI Y, LIN C, et al. Atrous residual U-Net for retinal vessel segmentation[J]. Computer Engineering and Applications, 2021, 57(7): 185-191(in Chinese). doi: 10.3778/j.issn.1002-8331.2009-0215
[9]	易三莉, 陈建亭, 贺建峰. 基于多路径输入和多尺度特征融合的视网膜血管分割[J]. 光电子·激光, 2021, 32(7): 735-741. YI S L, CHEN J T, HE J F. Retinal vascular segmentation based on multi-scale input and multi-scale feature fusion[J]. Journal of Optoelectronics·Laser, 2021, 32(7): 735-741(in Chinese).
[10]	HUANG G, LIU Z, VAN DER MAATEN L, et al. Densely connected convolutional networks[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2017: 2261-2269.
[11]	WOO S, PARK J, LEE J Y, et al. CBAM: Convolutional block attention module[C]//Proceedings of the European Conference on Computer Vision. Berlin: Springer, 2018: 3-19.
[12]	姜大光, 李明鸣, 陈羽中, 等. 骨架图引导的级联视网膜血管分割网络[J]. 工程科学学报, 2021, 43(9): 1244-1252. JIANG D G, LI M M, CHEN Y Z, et al. Cascaded retinal vessel segmentation network guided by a skeleton map[J]. Chinese Journal of Engineering, 2021, 43(9): 1244-1252(in Chinese).
[13]	RONNEBERGER O, FISCHER P, BROX T. U-Net: Convolutional networks for biomedical image segmentation[C]//Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention. Berlin: Springer, 2015: 234-241.
[14]	OKTAY O, SCHLEMPER J, LE FOLGOC L, et al. Attention U-Net: Learning where to look for the pancreas[EB/OL]. (2018-05-20)[2022-05-01]. http://arxiv.org/abs/1804.03999.
[15]	易三莉, 陈建亭, 贺建峰. ASR-UNet: 一种基于注意力机制改进的视网膜血管分割算法[J]. 山东大学学报(理学版), 2021, 56(9): 13-20. YI S L, CHEN J T, HE J F. ASR-UNet: An improved retinal vessels segmentation algorithm based on attention mechanism[J]. Journal of Shandong University (Natural Science), 2021, 56(9): 13-20 (in Chinese).
[16]	YAN Z Q, YANG X, CHENG K T. Joint segment-level and pixel-wise losses for deep learning based retinal vessel segmentation[J]. IEEE Transactions on Biomedical Engineering, 2018, 65(9): 1912-1923. doi: 10.1109/TBME.2018.2828137
[17]	LIN Y, ZHANG H G, HU G. Automatic retinal vessel segmentation via deeply supervised and smoothly regularized network[J]. IEEE Access, 2018, 7: 57717-57724.
[18]	YAN Z Q, YANG X, CHENG K T. A three-stage deep learning model for accurate retinal vessel segmentation[J]. IEEE Journal of Biomedical and Health Informatics, 2019, 23(4): 1427-1436. doi: 10.1109/JBHI.2018.2872813
[19]	SANTOS J C M D, CARRIJO G A, CARDOSO C D F D S, et al. Fundus image quality enhancement for blood vessel detection via a neural network using CLAHE and Wiener filter[J]. Research on Biomedical Engineering, 2020, 36(2): 107-119. doi: 10.1007/s42600-020-00046-y
[20]	LV Y, MA H, LI J N, et al. Attention guided U-Net with atrous convolution for accurate retinal vessels segmentation[J]. IEEE Access, 2020, 8: 32826-32839. doi: 10.1109/ACCESS.2020.2974027
[21]	ZHANG J W, ZHANG Y C, XU X W. Pyramid U-Net for retinal vessel segmentation[C]//Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing. Piscataway: IEEE Press, 2021: 1125-1129.
[22]	TCHINDA B S, TCHIOTSOP D, NOUBOM M, et al. Retinal blood vessels segmentation using classical edge detection filters and the neural network[J]. Informatics in Medicine Unlocked, 2021, 23: 100521. doi: 10.1016/j.imu.2021.100521
[23]	YANG L, WANG H X, ZENG Q S, et al. A hybrid deep segmentation network for fundus vessels via deep-learning framework[J]. Neurocomputing, 2021, 448: 168-178. doi: 10.1016/j.neucom.2021.03.085
[24]	KHAN T M, KHAN M A U, REHMAN N U, et al. Width-wise vessel bifurcation for improved retinal vessel segmentation[J]. Biomedical Signal Processing and Control, 2022, 71: 103169. doi: 10.1016/j.bspc.2021.103169