A deep learning-based ADRPPA algorithm for the prediction of diabetic retinopathy progression.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Diabetic retinopathy (DR) monitoring often relies on human experts.
• Current AI models offer diagnoses but lack predictive insights for DR prognosis and treatment.
• Longitudinal retinal imaging data holds potential for predicting DR progression.

Purpose of the Study:

• To develop and validate an AI-driven algorithm for predicting DR progression using longitudinal retinal imaging data.
• To assess the efficacy of deep learning models in forecasting the transition from nonreferable DR (NRDR) to referable DR (RDR).

Main Methods:

• Retrospective analysis of paired retinal fundus images from the EyePACS dataset (≥1-year intervals).
• Training a ResNeXt neural network for DR severity grading and a Mask R-CNN for microaneurysm quantification.
• Utilizing DR and microaneurysm scores to predict NRDR-to-RDR progression.

Main Results:

• ResNeXt models achieved high performance in diagnosing RDR (AUCs ranging from 0.963 to 0.971).
• Mask R-CNN demonstrated effective microaneurysm detection (recall: 0.786, precision: 0.615, F1-score: 0.690).
• The combined ResNeXt (768-pixel) and Mask R-CNN (1600-pixel) model predicted NRDR-to-RDR progression with an F1-score of 0.422.

Conclusions:

• Deep learning models trained on longitudinal data can effectively predict DR progression.
• AI-generated DR and microaneurysm scores aid in identifying patients at high risk for NRDR.
• This facilitates early intervention and timely treatment for diabetic retinopathy.