Fine-grained vascular segmentation model driven by self-supervised learning and its application in fundus image analysis

Fundus images are very important in the diagnosis of ophthalmic diseases, and blood vessel segmentation is its core task. Traditional methods are limited by high labeling cost, poor generalization of models and low accuracy of micro-vessel segmentation. In this paper, a fine-grained vascular segmentation model driven by self-supervised learning (SSL) is proposed to reduce the dependence on labeled data and improve the segmentation performance. The model adopts encoder-decoder architecture, and the encoder is based on Vision Transformer and learns the vascular feature representation through the self-supervised pre-training strategy of contrast learning and image reconstruction. The decoder integrates multi-scale feature fusion module, combines low-level details and high-level semantic information, and introduces boundary perception module to strengthen blood vessel boundary segmentation. Experiments are conducted on the open DRIVE data set. The results show that the model is superior to many traditional and supervised learning methods in Dice coefficient, sensitivity, specificity and AUC value. The Dice coefficient is 89.2%, the sensitivity is 86.7%, the specificity is 98.9% and the AUC value is 96.1%. Ablation experiments further verify the effectiveness of each component, and prove that self-supervised pre-training can make full use of unlabeled data, multi-scale fusion module can significantly improve capillary segmentation, and boundary sensing branch can effectively reduce the problem of edge blur. This model provides an efficient, accurate and robust new method for blood vessel segmentation in fundus image analysis, which is helpful to improve the accuracy of early diagnosis of ophthalmic diseases.