FMCA-Net: A context-aware network for precise retinal vessel parsing via dual-path attention and fusion

Retinal vessel segmentation is vital for diagnosing eye diseases and planning treatment. To address challenges such as vessel scale variation, limited receptive fields, and small vessel segmentation, FMCA-Net, a context-aware network with dual-path attention and fusion was proposed in this paper. The network is structured in three stages: the first stage preprocesses data to enhance the features of fundus images and reduce noise; In the second stage, the Adaptive Detail Enhancement Module (ADEM) and the Hyper-Resolution Fusion Network (HRF-Net) are adopted to extract the features of shallow texture and high-level semantic information of fundus images respectively; In the third stage, the Spatial Enhancement Fusion Module (SEFM) is used to enhance the spatial structure of the blood vessels and fuse the extracted features, making the spatial structure features of the blood vessels clearer. Experimental findings reveal that FMCA-Net attains accuracy scores of 97.02%, 96.73%, and 97.46%, sensitivity values of 84.94%, 73.44%, and 86.97%, along with AUC metrics of 98.88%, 97.64%, and 99.06% on the DRIVE, STARE, and CHASEDB1 datasets, respectively. Compared to other cutting-edge segmentation networks, the method shows enhanced performance, exceptional generalization ability, and reliable robustness.