Bundleseg: two-stage retinal blood vessel segmentation based on Bundle Modulese

This study presents an advanced two-stage retinal blood vessel segmentation network based on Bundle Modules (Bundle Segmentation Network, referred to as BD-Net), aiming to enhance the accuracy and robustness of retinal blood vessel segmentation. BD-Net evolves from the classic U-Net architecture, incorporating a layered network design concept and an innovative two-stage encoding-decoding strategy. The network introduces Bundle Modules to achieve multi-path feature map transmission, significantly enhancing the network's feature fusion and information flow capabilities. In the coarse segmentat ion stage, BD-Net initially detects retinal blood vessels and generates a probability map; in the fine segmentation stage, the probability map undergoes refined processing to improve segmentation accuracy. Experimental results demonstrate that BD - Net achieves an F1 score of 83.0, an AUC of 98.1, and an accuracy of 95.8% on the public dataset DRIVE, outperforming existing retinal blood vessel segmentation methods.