HCM: A Hybrid CNN-Mamba Architecture for Semi-Supervised Retinal Vessel Segmentation

Retinal vessel segmentation plays a critical role in the diagnosis and monitoring of ocular and systemic diseases. However, the task remains challenging due to low contrast between vessels and background, complex morphological structures, and the scarcity of annotated data. To address these challenges, we propose HCM, a novel Hybrid CNN-Mamba Architecture for semi-supervised retinal vessel segmentation. HCM integrates the strengths of Convolutional Neural Networks (CNNs) and the Mamba architecture, leveraging CNN's local feature extraction capabilities and Mamba's global dependency modeling. Furthermore, we propose Multi-Scale Contrastive Learning (MSCL), which combines image-level, region-level, and pixel-level contrastive losses to optimize feature learning across different scales. Extensive experiments on the DRIVE, CHASE_DB1, and STARE datasets demonstrate that HCM outperforms existing semi-supervised methods, achieving superior performance in terms of Dice coefficient, sensitivity, and intersection-over-union (IoU). Ablation studies validate the effectiveness of the hybrid architecture and the MSCL, highlighting their contributions to robust and accurate retinal vessel segmentation. Our work provides a promising solution for leveraging limited annotated data and improving segmentation performance in retinal vessel segmentation.