A High-Resolution Network with Strip Attention for Retinal Vessel Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Accurate retinal vessel segmentation is crucial for fundus image analysis.
• Deep learning methods, particularly U-Net variants, show promise but struggle with tiny/low-contrast vessels due to spatial detail loss and inadequate feature fusion.

Purpose of the Study:

• To propose a novel high-resolution network with strip attention for enhanced retinal vessel segmentation.
• To address limitations of existing methods in detecting fine and subtle vascular structures.

Main Methods:

• Utilized an HRNet-shaped architecture to maintain high-resolution representations throughout training.
• Developed a strip attention module with horizontal and vertical attention mechanisms to capture long-range dependencies.
• Integrated strip attention for dynamic, multi-layer feature fusion.

Main Results:

• Achieved high accuracy (96.16% on DRIVE, 97.08% on STARE) and sensitivity (82.68% on DRIVE, 89.36% on STARE).
• Demonstrated superior extraction of tiny and low-contrast vessels compared to mainstream methods.

Conclusions:

• The proposed high-resolution network with strip attention effectively improves retinal vessel segmentation.
• This method shows potential for aiding in the clinical analysis of fundus images.