Retinal artery/vein classification using genetic-search feature selection.

Area of Science:

• Ophthalmology
• Medical Image Analysis
• Computer Vision

Background:

• Automatic classification of retinal blood vessels into arteries and veins (A/V) remains a significant challenge in retinal image analysis.
• Existing methods often rely on graph analysis of the vasculature, potentially overlooking the contribution of pixel-level information.
• The importance of pixel-wise classification for accurate A/V identification has been underexplored.

Purpose of the Study:

• To demonstrate the efficacy of a comprehensive feature set for classifying vessel centerline pixels.
• To investigate the potential of pixel-wise classification as a key component in A/V classification.
• To develop a robust method for artery and vein classification in retinal images.

Main Methods:

• Extraction of a large number of features from vessel centerline pixels.
• Application of a genetic-search based feature selection technique to identify an optimal feature subset.
• Evaluation of the selected features for A/V classification.

Main Results:

• The proposed method achieved 90.2% accuracy, 89.6% sensitivity, and 91.3% specificity on the INSPIRE dataset.
• The approach, utilizing only centerline pixel information, demonstrated performance comparable to complex graph analysis techniques.
• The framework proved capable of discriminating vessels across different fundus cameras, independent of imaging device characteristics, resolution, and quality.

Conclusions:

• A comprehensive feature set is crucial for accurate A/V classification, particularly for individual vessels where graph-based methods have limitations.
• Pixel-wise classification using detailed features can enhance A/V labeling and improve upon existing graph-analysis approaches.
• The findings suggest a promising direction for improving automated retinal vessel analysis.