A retinal vessel segmentation network with multi-scale feature extraction and cross-layer attention fusion (MAF-Net)

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Accurate retinal vessel segmentation is vital for diagnosing ophthalmic and systemic diseases.
• Current methods face challenges with scale variation, contextual information, and directional structures, leading to missed thin vessels and discontinuities.

Purpose of the Study:

• To develop an advanced retinal vessel segmentation network, MAF-Net, addressing limitations of existing methods.
• To improve the accuracy and robustness of retinal vessel segmentation for enhanced diagnostic capabilities.

Main Methods:

• Proposed MAF-Net incorporates an adaptive multi-scale dilated residual (AMDR) module for dynamic scale response adjustment.
• Introduced a windowed hierarchical cross-scale attention fusion (WHCAF) module for efficient contextual information aggregation.
• Developed a semantically guided tri-axis fusion (SGTAF) module in skip connections to preserve vessel connectivity and structural consistency.

Main Results:

• MAF-Net achieved state-of-the-art accuracy on public datasets: 97.14% (DRIVE), 97.99% (STARE), and 97.82% (CHASE-DB1).
• The model demonstrated superior performance in stably extracting and restoring vessel information.
• MAF-Net effectively reduced missed detections of thin vessels and structural discontinuities.

Conclusions:

• MAF-Net significantly advances retinal vessel segmentation accuracy and reliability.
• The proposed modules effectively handle scale variation, contextual fusion, and directional structure recovery.
• MAF-Net shows great potential for improving early screening and diagnosis of eye diseases through fundus image analysis.