Automatic Artery/Vein Classification Using a Vessel-Constraint Network for Multicenter Fundus Images.

Area of Science:

• Medical Image Analysis
• Ophthalmology
• Cardiovascular Disease Research

Background:

• Retinal blood vessel abnormalities are linked to cardiovascular, cerebrovascular, and systemic diseases.
• Accurate artery/vein (A/V) classification is crucial for medical diagnosis.
• Existing A/V classification methods struggle with blurred boundaries and scale variations, leading to errors at vessel edges and ends.

Purpose of the Study:

• To develop a high-precision A/V classification model that addresses limitations of current methods.
• To enhance the accuracy of A/V classification by incorporating vessel distribution and edge information.
• To simultaneously achieve accurate vessel segmentation alongside A/V classification.

Main Methods:

• Proposed a novel vessel-constraint network (VC-Net) for data fusion-based A/V classification.
• Introduced a vessel-constraint (VC) module to combine local and global vessel information, generating a weight map to refine A/V features.
• Incorporated a multiscale feature (MSF) module to improve feature extraction capability and model robustness.

Main Results:

• Achieved a balance accuracy of 0.9554 and F1 scores of 0.7616 (arteries) and 0.7971 (veins) on the DRIVE dataset.
• Demonstrated competitive performance in A/V classification and vessel segmentation compared to state-of-the-art methods.
• Showcased good robustness on newly created multicenter datasets (Tongren, Kailuan) and cross-dataset validation.

Conclusions:

• The proposed VC-Net effectively enhances A/V classification by suppressing background noise and improving focus on vessel edges and ends.
• The model offers a robust and accurate solution for both A/V classification and vessel segmentation in retinal fundus images.
• The public release of the Tongren dataset and source code aims to foster further research in this critical diagnostic area.