LCNet: lightweight segmentation network for blood vessel segmentation in retinal imaging

Precise retinal vessel segmentation techniques are crucial for computer-aided clinical diagnosis. Recent advancements in deep learning have considerably enhanced segmentation accuracy (ACC); however, existing methods often struggle with thin and fuzzy boundaries due to cross-connection redundancy. Most mainstream models rely on complex encoders, resulting in high parameter counts and resource demands. Therefore, we propose LCNet, a lightweight U-shaped network with depth-separable convolution to minimize parameters and computational costs. It incorporates a synergistic coordinate attention module to enhance feature learning and captures multiscale features through the atrous spatial pyramid pooling module. Additionally, it introduces four side-output layers for extra supervision. Evaluating LCNet on four classical datasets (DRIVE, STARE, CHASEDB1, and IOSTAR), LCNet achieves global accuracies of 96.02%, 97.95%, 97.95%, and 97.77%. Our experiments that LCNet delivers enhanced performance with only 2.65 M parameters and 21.2 GFLOPs on DRIVE. We also demonstrate the effectiveness of LCNet on fundus images with lesions and optical coherence tomography angiography images, establishing it as a lightweight model with high efficiency and ACC for retinal vessel segmentation.