多层级特征融合的视网膜血管分割

Retinal Vessel Segmentation with Multi-layer Feature Fusion

为了从视网膜图像中精准分割视网膜血管,实现图像分割任务的性能提升,提出一种嵌套U型网络(NestedNet):通过多层级捕获高级特征,增强网络的表达和特征融合能力;基于U型网络的编码器-解码器结构,NestedNet采用三层嵌套,形成倒金字塔式结构,最外层两个U型结构的编码器输出传递给下一层编码器;解码器与下一编码器的Addition操作构成多条从输入到输出的路径,以丰富特征,促进特征传递和融合,提升图像表达能力;并行残差注意力机制(PRAM)增强网络对局部和全局结构的理解,生成更准确的预测结果.在DRIVE和CHASE_DB1数据集上的实验结果显示,平均准确率分别达到0.9576和0.9691,受试者工作特性曲线下面积分别为0.9819和0.9901,精确率-召回率曲线下面积分别为0.9182和0.9411,在多项测试指标上表现较好.

To accurately segment retinal vessels from retinal images and achieve performance improvement in image segmentation tasks,a nested U-shaped network(NestedNet)was proposed:the expression and feature fusion ability of the network was enhanced by capturing high-level features at multiple levels;based on the encoder-decoder structure of the U-shaped network,the NestedNet adopts a three-layer nesting to form an inverted pyramid structure,and the encoder output of the two outermost U-shaped structure was passed to the encoder of the next layer;the Addition operation of the decoder and the next encoder constitutes multiple paths from the input to the output in order to enrich the features,promoted the feature transfer and fusion,and enhanced the image expression capability;the Parallel Residual Attention Mechanism(PRAM)enhanced the network’s understanding of local and global structures to generate more accurate predictions.Experimental results on the DRIVE and CHASE_DB1 datasets show that the average accuracy reaches 0.9576 and 0.9691,respectively,the area under the curve(AUC)of the subjects’work characteristics is 0.9819 and 0.9901,and the area under the curve(P-R AUC)of the precision-recall ratio is 0.9182 and 0.9411,which performs well on multiple test metrics with better performance.