Few-Shot Retinal Vessel Semantic Segmentation Under Threshold-Based Co-training Dual Network

Abstract

Retinal vessel segmentation plays an important role in the early diagnosis and monitoring of many ocular and systemic diseases. However, labeled medical imaging data is scarce, costly, and requires pixel-level precision, making few-shot learning a promising solution to this challenge. This paper presents a novel threshold-based co-training framework using dual network for few-shot retinal vessel segmentation. Specifically, the method employs two segmentation models initialized with different parameters, which collaboratively learn by leveraging high-confidence pseudo-labels generated through a thresholding mechanism. The proposed method balances segmentation accuracy and robustness by integrating binary cross-entropy and Dice loss, effectively minimizing noise and uncertainty. Evaluation on the CHASE_DB1 dataset shows superior performance compared to state-of-the-art methods, achieving improvements in accuracy, sensitivity, specificity, and Dice score. These findings highlight the potential of threshold-based co-trained dual network for efficient and accurate retinal vessel segmentation using limited data.