SSL-ENSAM: Utilizing Sam for Semi-Supervised Retinal Vessel Segmentation with Quality-Aware Enhancement

Mengxian He, Minqing Zhang, Hing Lam Chang, Xiaoyi Feng, Mengdi Gao, Shuai Jiang, Wu Yuan

Abstract: Retinal vessel segmentation (RVS) in fundus images based on deep neural networks (DNNs) is beneficial to the quantitative diagnosis of ophthalmic diseases, in particular retinal vascular diseases. However, the scarcity of labeled datasets and the inconsistent quality of fundus images both pose a challenge in RVS. In this work, we present a plug-and-play quality-aware module and embed it with SAM and specialist model into a semi-supervised learning (SSL) pipeline, named SSL-ENSAM. Specifically, we first classify the fundus images into low- and high-quality subsets in an unsupervised manner, then implement different image enhancement or augmentation accordingly, which enables SAM to serve as a label engine to generate more reliable pseudo-labels from unlabeled data to supervise the specialist model. Comparison experiments and ablation studies are conducted to demonstrate that our proposed method achieves the SOTA performance.