Deep Learning Method for Semi-supervised Vessel Segmentation

Retinal fundus vessel segmentation technology is of great significance for the diagnosis and prevention of diabetes and retinal diseases. However, pixel-level annotation of blood vessels is very time-consuming and costly. In the case of limited labeled data, blood vessel segmentation remains a challenging task. To address the above problems, this paper proposes a Dual Uncertainty guided Dynamic-competitive Collaborative Mean Teacher framework ( DU-DCMT) for semi-supervised vessel segmentation. The framework utilizes a Dual Uncertainty Region-processing Module (DURM) to expand data distribution, enhance the confidence of pseudo-labels, and guide network attention to correct potential error-prone regions. Additionally, a dynamic competitive teacher model is designed to select high-quality pseudo-label generators, jointly train networks, and preserve network diversity for better segmentation performance. Experimental results show that when the labeling ratio is 0. 2, the sensitivity and accuracy of this method on the DRIVE dataset are 0. 817 3 and 0. 967 3 respectively, and on the STARE dataset are 0. 768 2 and 0. 978 3 respectively. Therefore, DU-DCMT effectively alleviates the problem of data label scarcity, and its segmentation performance outperforms current state-of-the-art methods.