Edge-Prior Contrastive Transformer for Optic Cup and Optic Disc Segmentation

Optic Cup and Optic Disc segmentation plays a vital role in retinal image analysis, with significant implications for automated diagnosis. In fundus images, due to the difference between intra-class features and the complexity of inter-class features, existing methods often fail to explicitly consider the correlation and discrimination of target edge features. To overcome this limitation, our method aims to capture interdependency and consistency by involving differences in pixels on the edge. To accomplish this, we propose an Edge-Prior Contrastive Transformer (EPCT) architecture to augment the focus on the indistinct edge information. Our method incorporates pixel-to-pixel and pixel-to-region contrastive learning to achieve higher-level semantic information and global contextual feature representations. Furthermore, we incorporate prior information on edges with the Transformer model, which aims to capture the prior knowledge of the location and structure of the target edges. In addition, we propose an anchor sampling strategy tailored to the edge regions to achieve efficient edge features. Experimental results on three publicly available datasets demonstrate the effectiveness of our proposed method, as it achieves excellent segmentation performance.