A Diabetic Retinopathy Classification Framework Based on Deep-Learning Analysis of OCT Angiography

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) classification is crucial for preventing blindness in diabetic patients.
• Optical coherence tomography (OCT) and OCT angiography (OCTA) offer advantages over traditional fundus photography for retinal imaging.

Purpose of the Study:

• To evaluate a deep-learning-aided framework for classifying referable (rDR) and vision-threatening diabetic retinopathy (vtDR) using volumetric OCT and OCTA data.
• To assess the framework's ability to provide specialist-level DR classification from a single imaging modality.

Main Methods:

• Trained a deep learning framework on 456 OCT and OCTA volumes from healthy participants and diabetic patients.
• Utilized retina specialist labels (non-referable, referable, vision-threatening DR) for training and validation.
• Generated 3D class activation maps to identify regions crucial for DR classification.

Main Results:

• The framework achieved an AUC of 0.96 ± 0.01 for rDR classification and 0.92 ± 0.02 for vtDR classification.
• Quadratic-weighted kappa values were 0.83 ± 0.04 for rDR and 0.73 ± 0.04 for vtDR.
• Multi-class DR classification achieved a quadratic-weighted kappa of 0.83 ± 0.03.

Conclusions:

• A deep learning framework utilizing only OCT and OCTA can achieve specialist-level DR classification.
• The framework demonstrates potential for developing clinically valuable automated DR diagnosis systems.