Retinal vessel segmentation method based on RSP-SA Unet network.

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Retinal vessel segmentation is crucial for diagnosing fundus disorders.
• Existing methods struggle with extracting fine-grained vessel features and preserving edge details.

Purpose of the Study:

• To introduce a novel Unet-based model, RSP-SA Unet, to address limitations in retinal vessel segmentation.
• To enhance the extraction of fine vessel features and the accuracy of vessel edge detection.

Main Methods:

• The proposed RSP-SA Unet incorporates a Residual SimAM Pyramid-Spatial Attention (RSP) structure in encoding/decoding layers for multi-scale feature extraction.
• Spatial Attention (SA) is integrated into upsampling layers to improve segmentation of low-contrast, small blood vessels.
• The model was evaluated on the CHASE_DB1, DRIVE, and STARE datasets.

Main Results:

• RSP-SA Unet achieved high segmentation accuracy (ACC) of 0.9763, 0.9704, and 0.9724 on the tested datasets.
• The model demonstrated strong performance with Area Under the ROC Curve (AUC) reaching 0.9896, 0.9858, and 0.9906.
• Overall performance surpassed existing comparative methods.

Conclusions:

• The RSP-SA Unet model significantly improves retinal vessel segmentation compared to previous methods.
• The enhanced feature extraction and attention mechanisms contribute to superior accuracy in identifying fine vessels and their edges.
• This model shows promise for improved automated diagnosis of fundus-related diseases.