Interpretable Machine Learning-Based Concentric Regional Analysis of OCTA Images for Enhanced Diabetic Retinopathy...

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss.
• Early detection of DR is crucial for risk stratification and patient management.
• Optical coherence tomography angiography (OCTA) offers noninvasive visualization of retinal microvasculature.

Purpose of the Study:

• To investigate if dividing OCTA images into anatomical regions enhances DR classification.
• To identify which retinal regions provide the most discriminative information for DR detection.
• To compare regional analysis performance against whole-image machine learning models.

Main Methods:

• 188 OCTA images (normal, mild DR, moderate DR) were analyzed.
• Images were segmented into seven concentric regions.
• Vessel density features were extracted per region.
• Machine learning classifiers were trained and ensembled for classification.
• Local Interpretable Model-Agnostic Explanations (LIME) were used for interpretability.

Main Results:

• The regional analysis ensemble model achieved 97% accuracy, 98% precision, 97% recall, and 97% F1-score.
• Performance surpassed whole-image transfer-learning models.
• Parafoveal regions were identified as most informative.
• Region 3 showed the highest contribution, followed by Regions 2 and 5.

Conclusions:

• Regional analysis of OCTA vessel density improves DR classification performance and interpretability.
• Parafoveal vascular alterations are highly discriminative for DR stages.
• Further validation in larger, multi-center cohorts is warranted.