Decom-UNet3+: A Retinal Vessel Segmentation Method Optimized With Decomposed Convolutions

The intricate and highly branched structure of retinal blood vessels, along with the fragility of fine vessels, makes segmentation a challenging task. To address this issue, we propose Decom-UNet3+, a model that optimizes the encoders by employing decomposed convolutions. Specifically, the encoders replace standard convolutional layers with asymmetric convolutions and depthwise separable convolutions, reducing the number of parameters while enhancing capability for feature extraction. Additionally, a spatial attention mechanism is integrated to improve focus on vessel regions and suppress background noise. The model is evaluated on high-resolution, expertly annotated datasets including CHASEDB1, DRIVE, STARE and HRF, achieving an average accuracy of 97.2% on CHASEDB1, 96.4% on DRIVE, 94.3% on STARE and 97.6% on HRF, outperforming the original UNet3+ model. The results demonstrates that Decom-UNet3+ effectively improves vascular segmentation performance with lower computational cost and parameter overhead, offering a more efficient and robust solution for automated retinal disease screening.