多尺度特征融合双U型视网膜分割算法

视网膜血管形态结构是反映人体健康的重要指标 ,针对现有视网膜血管分割存在主 血管模糊、微细血管断裂和视盘误分割等问题,提出多尺度特征融合双U型视网膜分割算 法。首先,利用低层U-Net高效循环残差模块对眼底图像进行粗粒度分割,得到视网膜血 管 初步轮廓。其次,将粗分割图与原始特征图像素相乘送入高层U-Net,利用其缩放宽残差 模 块进行细粒度图像解码,丰富视网膜血管细节信息。同时利用3路径注意力机制复合性连接 双网络的编码层与解码层,实现特征映射跨网络传播,减小上下文语义差异。最后,融合双 层网络输出提取血管区域,双U 型网络能够更深层次提取血管像素,精准分割出视网膜细 节。在DRIVE与STARE数据集上进行实验,其准确率分别为96.45%和97.02%,敏感度分 别为83.35%和81.40%,特异性分别为98.38%和 98.83%,总体性能优于现有算法。

Multi scale feature fusion double U-shaped retinal segmentation algorithm

The morphological structure of retinal vessels is an important index to reflect human health.In order to solve the existing problems in retinal vessel segmentation,su ch as blurre d main vessels,broken microvessels and false segmentation of optic disc,the mult i-scale fea ture fusion double U-shaped retinal segmentation algorithm is proposed.Firstly,the low level U- Net efficient cyclic residual module is used for coarse-grained segmentation o f fundus imag es to obtain the initial contour of retinal vessels.Secondly,the coarse segmenta tion image is multiplied by the pixels of the original feature image into the high level U -Net,and the scal ing wide residual model is used to decode the fine-grained image to enrich the details of retinal vessels.At the sametime, the three pathway attention mechanism is used to connect the encoding layer and decoding layer of the double network in a compound way to realize the cross network propagation of feature mappi ng and reduce the semantic difference of context.Finally,the double U-shaped ne twork can ext ract vascular pixels at a deeper level and accurately segment retinal details.Ex periments we re conducted on DRIVE and STARE datasets,the accuracy was 96.45% and 97.02%,the sensitivity w as 83.35% and 81.40%,and the specificity was 98.38% and 98.83%,respectively.The overall performance is better than existing algorithms.