RCNet: A Hybrid Network Adapted for OCTA Retinal Capillary Segmentation

Abstract

Optical Coherence Tomography Angiography (OCTA) provides clear visualization of ocular microvascular details, serving as a critical tool for assessing retinal and choroidal vascular systems. However, the complex capillary structure and low capillary-to-background contrast in high-resolution OCTA images pose significant challenges for segmentation. To address the low segmentation accuracy of traditional methods, this paper proposes RCNet, a hybrid CNN-Transformer network. RCNet adopts an encoder–decoder architecture based on Bi-Level Routing Attention (BRA), where bi-level routing attention focuses on densely vascularized regions, optimizing global capillary feature extraction. A Convolutional Bottleneck Module (CBM) splits bottleneck features into two channel groups using a fixed ratio α. A high-capacity branch enhances local details, while a lightweight branch preserves complementary information. The branches are then fused. Additionally, a Dense Skip Connection Module (DSCM) enhances feature information flow in low-contrast regions, compensating for spatial information loss. By integrating CNN’s local feature extraction with Transformer’s global dependency modeling, RCNet achieves accurate segmentation of OCTA capillary images. Experiments on the OCTA-500 and ROSE datasets demonstrate that RCNet achieved a higher point estimate under the current protocol, facilitating and assisting in the clinical assessment of ocular diseases.