An Improved Semi-Supervised Segmentation Framework for Optic Cup Segmentation

Glaucoma is a serious and irreversible blinding disease characterized by an abnormally large cup-to-disc ratio (CDR). Accurate segmentation of the optic cup can help calculate CDR, thereby helping with the early detection and diagnosis of glaucoma. However, existing deep learning-based optic cup segmentation methods require a large number of pixel-level labeling, and the process of labeling consumes a large amount of the clinician's time and suffers from subjectivity bias. Therefore, in this paper, we propose an improved semi-supervised segmentation algorithm to reduce the dependence of deep learning networks on labeling data. Specifically, we construct a new semi-supervised optic cup segmentation dataset, REFUGE-Semi. In addition, we improve UniMatch in terms of both network structure and loss function, which enhances its performance in the optic cup segmentation task. The results of comparison and ablation experiments show that the improved UniMatch achieves the best optic cup segmentation performance.