Deep Learning Approach to Identify Diabetic Retinopathy Severity and Progression Using Ultra-Wide Field Retinal Images

Diabetic retinopathy (DR) is a major progressive microvascular complication of Type 1 or Type 2 diabetes. It is irreversible and can lead to poor vision and blindness. We present a deep-learning approach to classify patient retinal images based on DR severity and progression using a set of 14,524 retinal images from the Joslin Diabetes Center. Unlike most existing approaches, which usually use 30–60 \(^{\circ }\) , centrally focused retinal images, our work utilizes ultra-wide field (UWF) retinal images, capturing 4-times more retinal area. Moreover, we introduce a classification scale with 8 labels, each a combination of disease severity level and progression risk (e.g. mild DR with progression, severe DR without progression, etc.). We achieve an overall classification accuracy of 77.04% and AUC of 98.6% for 8 classes, using a hierarchical model with pre-trained weights from EfficientNetV2-S. Given the progressive nature of DR, our method is better positioned to detect risk for progression than standard approaches relying solely on DR severity.