From Local to Global: A Graph Framework for Retinal Artery/Vein Classification.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Fundus photography is crucial for diagnosing eye disorders and detecting early diabetes through retinal vascular biomarkers.
• Accurate classification of retinal arteries and veins is essential for quantifying vascular changes.

Purpose of the Study:

• To develop a new framework that enhances local vessel classification by incorporating global vascular network information.
• To improve the accuracy of artery and vein classification in retinal images.

Main Methods:

• A novel framework combining local vessel classification with a global vascular network model using graph convolution.
• Comparison with traditional methods like Convolutional Neural Networks (CNN) and Handcrafted Pixel Feature Classification (HPFC).

Main Results:

• The proposed graph convolution model achieved the highest accuracy (86.45%) compared to CNN (85.5%) and HPFC (82.9%).
• The model also obtained the best Area Under the Receiver Operating Characteristic Curve (AUC) of 0.95, outperforming CNN (0.93) and HPFC (0.90).

Conclusions:

• The new classification framework effectively corrects local classification errors by minimizing global errors.
• This approach offers an easy-to-deploy method for improving classification performance with minimal additional computational cost.