Image Segmentation of Retinal Blood Vessels Based on Dual-Attention Multiscale Feature Fusion

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Current retinal blood vessel segmentation methods lack sufficient detail.
• Accurate segmentation is crucial for diagnosing various eye diseases.

Purpose of the Study:

• To propose a novel multiscale feature fusion residual network with dual attention for enhanced retinal blood vessel segmentation.
• To improve the extraction of fine details in retinal images.

Main Methods:

• Designed a feature fusion residual module with adaptive calibration weights to prevent gradient dispersion and network degradation.
• Employed spatial attention (SA) and efficient channel attention (ECA) modules in the backbone network for adaptive feature selection.
• Fused multi-level network information, integrating long-range and short-range features.

Main Results:

• Achieved high classification accuracy: 0.9795 on the STARE dataset and 0.9785 on the DRIVE dataset.
• Demonstrated superior segmentation performance compared to current mainstream methods.
• Effectively aggregated low-level and high-level feature information for improved segmentation.

Conclusions:

• The proposed dual-attention-based multiscale feature fusion residual network significantly enhances retinal blood vessel segmentation.
• The method effectively extracts image details and improves overall segmentation performance.
• This approach offers a promising tool for clinical applications in ophthalmology.