全尺度密集卷积U型网络的视网膜血管分割算法  

Retinal vascular segmentation algorithm based on full scale dense convolutional U-shaped networks

作　　者：夏平[1,2] 何志豪 雷帮军[1,2] 彭程[1,2] 王雨蝶 XIA Ping;HE Zhi-hao;LEI Bang-jun;PENG Cheng;WANG Yu-die(Hubei Key Laboratory of Intelligent Vision based Monitoring for Hydroelectric Engineering,Three Gorges University,Yichang 443002,China;College of Computer and Information Technology,Three Gorges University,Yichang 443002,China)

机构地区：[1]三峡大学水电工程智能视觉监测湖北省重点实验室,湖北宜昌443002 [2]三峡大学计算机与信息学院,湖北宜昌443002

出　　处：《计算机工程与设计》2024年第3期866-873,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(U1401252);湖北省重点实验室开放基金项目(2018SDSJ07)。

摘　　要：针对视网膜图像中血管尺度跨度大、细小血管与背景高度相似导致误分割和未分割等问题,提出一种全尺度密集卷积U型网络的视网膜血管分割方法。为提取更复杂的特征信息,构建级联卷积融合密集块(cascade convolutional fusion dense blocks, CCF-DB)作为U型网络的编解码器用以提取视网膜血管的特征信息;在网络最底端嵌入混合注意力级联卷积密集块(mixed attention cascaded convolutional dense block, MACC-DB),进一步提升感受野,获取更高维的语义特征信息;在模型的解码部分采用全尺度的跳跃连接,捕获不同尺度下的血管特征信息,提升模型的分割精度。实验结果表明,在DRIVE数据集上,相比于U-Net、U-Net3+、SA-Unet、FR-Unet等算法,此算法的AUC值达到了98.26%,准确率为95.82%;在CHASE-DB1数据集上,此算法的AUC值达98.84%,准确率达96.66%。采用此算法进行视网膜血管分割,分割的精度和鲁棒性均有不同程度的提升,对细小血管分割达到了优良的效果。A full scale dense convolutional U-shaped network based retinal blood vessel segmentation method was proposed to address the issues of large blood vessel scale span and similarity between small blood vessels and background height in retinal images,resulting in mis-segmentation and non-segmentation.To extract more complex feature information,cascade convolutional fusion dense blocks(CCF-DB)were constructed as the codec of the U-shaped network to extract the feature information of retinal blood vessels.The mixed attention cascaded convolutional density block(MACC-DB)was embedded at the bottom of the network to further enhance the receptive field and obtain higher dimensional semantic feature information.In the decoding part of the model,a full scale skip connection was used to capture vascular feature information at different scales,improving the segmentation accuracy of the model.Experimental results show that on the DRIVE dataset,compared to algorithms such as U-Net,U-Net3+,SA-Unet,FR-Unet,etc.,the AUC value of this algorithm reaches 98.26%,with an accuracy of 95.82%.On the CHASE-DB1 dataset,the AUC value of this algorithm reaches 98.84%,with an accuracy rate of 96.66%.Therefore,using this algorithm for retinal vessel segmentation improves the accuracy and robustness to varying degrees,achieving excellent results for small vessel segmentation.

关 键 词：医学图像分割 深度学习 视网膜血管分割 全尺度密集卷积 编解码结构 混合注意力 级联卷积 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]