Adaptive multi-scale feature extraction and fusion network with deep supervision for retinal vessel segmentation

Accurate retinal vessel segmentation is crucial for early clinical diagnosis and effective disease treatment guidance. Due to the large scale variation and complex structure of retinal vessels, common U-shaped networks fail to capture distinct and representative features. Furthermore, the continuous downsampling leads to loss of spatial features. To address these challenges, an adaptive multi-scale feature extraction and fusion network with deep supervision (AMFEF-Net) is proposed for retinal vessel segmentation. First, a structured residual module that integrates local and global information to preserve spatial features is built via residual connection. A multi-scale features aggregated attention module is then designed to obtain high-level feature representations of multi-scale vessels. A feature fusion module is utilized to guide the fusion of features at different levels, which exploits the complementary of high-level and low-level features. Additionally, multi-scale deep supervisionis used to learn hierarchical representations from multi-scale aggregated feature maps. Ablation and comparison study on three public datasets (DRIVE, CHASE_DB1, and STARE) are performed. Results demonstrate AMFEF-Net's superior segmentation performance, particularly in the segmentation of tiny vessels and the extraction of the whole vascular network.