并行双分支融合网络用于视网膜血管分割

Parallel dual-branch fusion network for retinal vessel segmentation

针对视网膜图像中拓扑结构复杂、对比度差的毛细血管,已有的深度学习模型存在分割精度低、分割结果不连续问题,提出一种基于深度学习的并行双分支融合网络,该网络包含并行双分支模块和融合模块两部分.并行双分支模块使用两条U形支路:像素级分割支路和中心线级分割支路,旨在分别获取视网膜图像中粗血管和细血管.基于空间注意力的融合模块旨在进一步优化并行双分支模块输出的初步分割图,获取毛细血管的更多细节特征.在ROSE-1的SVC数据集上的实验结果表明:所提出的模型在所有评价指标中表现优异,其中,ROC曲线下的面积占比(AUC)达到了93.63%,精确度(ACC)达到了92.25%.

It existed the weakness of low segmentation accuracy for capillaries in retinal images with complex topology and low contrast with current deep learning models and the segmentation results were discontinuous.To address these problems,a parallel dual-branch fusion network based on deep learning was proposed,which was consisted of two parts:parallel dual-branch module and fusion module.The parallel dual-branch module was designed to use two U-shaped branches:pixel-level segmentation branch and centerline-level segmentation branch,aiming to obtain thick and thin vessels in retinal images respectively.The fusion module was built based on spatial attention which was used to further optimize the preliminary segmentation map output by the parallel dual-branch module and obtain more detail features of capillaries.The experimental results on SVC dataset of ROSE-1 showed that the proposed model performed excellently in all evaluation indicators,among which the area under the ROC curve(AUC)reached 93.63%,and the accuracy(ACC)reached 92.25%.