Deep Learning Model to Detect Diabetic Retinopathy in 45° Images Using Ground Truth from Ultra-Widefield Imaging.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss, necessitating effective screening methods.
• Current screening often relies on wide-field imaging, which may not always be accessible.
• Assessing DR severity and its progression requires accurate and efficient diagnostic tools.

Purpose of the Study:

• To evaluate a deep learning (DL) model's performance in classifying diabetic retinopathy (DR) severity using fundus images with varying fields of view.
• To determine if central retinal features alone can reflect the overall disease burden of DR.
• To investigate vascular biomarkers in regions identified by the DL model to understand the biological basis of its inferences and enhance clinical interpretability.

Main Methods:

• An observational, cross-sectional study involving 2610 participants aged ≥40 years from South India.
• Diabetic retinopathy severity was graded using 200° ultra-widefield (UWF) and centrally masked 45° fundus images.
• A convolutional neural network was trained and evaluated, with performance assessed using accuracy, precision, recall, and ROC curves. Gradient-weighted class activation mapping (Grad-CAM) and vascular biomarker analysis (tortuosity, fractal dimension, vessel density) were performed.

Main Results:

• The DL model achieved high classification accuracy for DR severity (97.12% for UWF, 97.24% for 45° masked images).
• Peripheral DR pathology did not significantly impact classification accuracy. DL assessment of 45° images reduced DR underestimation from 16.5% to 6.2%.
• Grad-CAM highlighted central retinal regions, and vascular analysis revealed differences in vessel density and tortuosity associated with peripheral lesions, suggesting detection of subclinical changes.

Conclusions:

• Deep learning applied to 45° fundus images accurately classifies DR and detects subtle vascular biomarkers predictive of peripheral disease.
• AI-enhanced 45° imaging shows potential as a scalable tool for DR screening, particularly in resource-limited settings.
• AI-powered approaches using affordable fundus cameras can facilitate cost-effective detection and triage of high-risk DR cases in primary care.