ACL-Net: Semi-supervised Polyp Segmentation via Affinity Contrastive Learning

Automatic polyp segmentation from colonoscopy images is an essential prerequisite for the development of computer-assisted therapy. However, complex semantic information and blurred edges polyps make extremely difficult. In this paper, we propose a novel semi-supervised framework using affinity contrastive learning (ACL-Net), which implemented between student teacher networks to consistently refine pseudo-labels segmentation. By aligning maps two branches, better region activation can be obtained fully exploit appearance-level context encoded in feature maps, thereby improving capability capturing not only global localization shape context, but also local textural boundary details. utilizing rich inter-image establishing based on memory bank, cross-image aggregation (CAA) module further branches. continuously adaptively refining with optimized affinity, improve mutually reinforced knowledge interaction among consistency iterations. Extensive experiments five benchmark datasets, including Kvasir-SEG, CVC-ClinicDB, CVC-300, CVC-ColonDB ETIS, demonstrate effectiveness superiority our method. Codes are available at https://github.com/xiewende/ACL-Net.