Joint Boundary-Enhanced and Topology-Preserving Dual-Path Network for Retinal Layer Segmentation in OCT Images with Pigment Epithelial Detachment

Automatic retinal layer segmentation methods are currently successful in normal Optical Coherence Tomography (OCT) images, but they face great challenges for eyes with Pigment Epithelial Detachment (PED), where the morphology and structure of the retina change dramatically. Therefore, we propose a novel dual-path network that uses residual blocks as the encoders, and semantic path and boundary paths for layer segmentation and boundary regression, respectively. Specifically, to capture the shape and boundary information of objects accurately, we design a Boundary-Enhanced Global Attention (BEGA) module for semantic path to provide global guidance for high-level features. Additionally, we use a Gradient-Guided Spatial Attention (GGSA) module in the boundary path to acquire boundary features, alleviating interference in non-boundary regions. Finally, the segmentation probability map and the distance map output by the network are fed into the Topology Correction (TC) module to obtain the topology-guaranteed results. We investigate the proposed method on the OCTA-500 dataset and the experimental results prove that the proposed method outperformed the other four existing methods.