采用双支路和Transformer的视杯视盘分割方法

视网膜血管复杂且背景与视杯视盘区域相似，是造成视杯视盘分割精度不高的原因。为了更加准确地分割视杯视盘，设计了一种具有双支路特征融合的分割网络。网络主支使用Transformer对特征进行提取，弥补了卷积运算在建立远程关系方面存在的不足。采用多个模块来融合浅层空间特征与高级语义特征：尺度感知-特征融合模块(SCA-FFM)用于从高层次特征中收集视盘和视杯的语义和位置信息；识别模块(IM)利用注意力机制减少低层次特征中存在的错误信息和噪声，增强空间细节特征的提取；使用图卷积域-特征融合模块(GCD-FFM)将高级语义特征和低级特征进行融合，使特征图同时具有全局和局部信息。对比实验表明，本文方法表现出更好的分割效果，且具备良好的泛化能力。

Optic Disk and Cup Segmentation Using Dual Branch and Transformer

The complex retinal vasculature and similar background to the optic cup optic disc region are the reasons for the poor accuracy of the optic cup and optic disc segmentation. We design a segmentation model with dual-branch feature fusion to improve the segmentation performance. The main branch of the model uses a Transformer for feature extraction, which makes up for the deficiency of convolutional operations in establishing remote relationships. We use several modules to fuse shallow spatial features with high-level semantic features: SCA-FFM (Scale Awareness-Feature Fusion Module) is used to collect semantic and positional information about the optic disc and optic cup from high-level features; IM (Identification Module) uses an attention mechanism to reduce the error information and noise present in low-level features and enhance the extraction of spatial detail features; GCD-FFM (Graph Convolution Domain- feature fusion module) to fuse high-level semantic features and low-level features so that the feature map has both global and local information. The comparison experiments show that the method in this paper exhibits a better segmentation effect and good generalization ability.