Generative Neural Framework for Micro-Vessels Classification.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal blood vessel abnormalities correlate with systemic, cardiovascular, and cerebrovascular diseases.
• Accurate retinal artery/vein (A/V) classification is crucial for clinical decision-making.
• Current methods struggle with non-uniform brightness, vessel thickness variations, bifurcations, and endpoints.

Purpose of the Study:

• To propose AV-Net, a novel deep learning model for precise retinal artery/vein classification.
• To address limitations of existing methods in handling complex vessel features.

Main Methods:

• AV-Net utilizes context information and data fusion for improved A/V classification.
• A dedicated module fuses local and global vessel information to generate a weight map.
• This weight map constrains artery/vein features, suppressing background noise and enhancing region extraction at bifurcations and endpoints.
• A multiscale-feature module captures both coarse and fine vessel details to boost model robustness.

Main Results:

• The proposed AV-Net demonstrates improved precision in retinal artery/vein classification.
• Enhanced feature extraction at vessel bifurcations and endpoints was achieved.
• Model robustness was improved through multiscale feature capture.

Conclusions:

• AV-Net offers a robust solution for accurate retinal artery/vein classification.
• The model's ability to handle complex vessel morphologies enhances its clinical utility for diagnosing associated systemic diseases.