Semi-supervised Choroidal Segmentation with Attention-based Refinement Contrastive Learning

A comprehensive measurement of choroidal thickness is essential for diagnosing various ocular conditions. Automated measurement encounters difficulties due to the high proportion of fuzzy boundaries in choroidal segmentation. In case of limited annotated data, large amounts of unlabeled data hinder the effective learning of boundary features makes it a challenge to extract features from fuzzy boundaries. To address this issue, we propose a semi-supervised segmentation framework based on contrastive learning, called Attention-based Refinement Contrastive Semi-Supervised Learning (ARC-SSL). This framework is designed to utilize labeled data to effectively extract boundary features from unlabeled data. Furthermore, we introduce an innovative attention-based strategy for refining the assignment of samples in contrastive learning. Experimental results demonstrate that our ARC-SSL outperforms state-of-the-art models on both public and private datasets.