Diagnostic Performance of a Deep Learning Algorithm for Multi-Lesion Detection in Diabetic Retinopathy Screening from Fundus Photographs with Ophthalmologist Benchmarking

Abstract

Background: Diabetic retinopathy is among the leading preventable causes of blindness worldwide, but the shortage of trained ophthalmologists and the growing burden of screening still remain a considerable challenge, particularly in resource-constrained settings. Purpose: The aim of this study was to assess the diagnostic accuracy of a deep learning algorithm for detecting multiple retinal lesions associated with diabetic retinopathy in fundus photographs and to compare its performance with that of ophthalmologists. Methods: This retrospective analytical study used a combined dataset of approximately 38,788 annotated colour fundus photographs drawn from the publicly available EyePACS and APTOS 2019 datasets. A deep learning model based on EfficientNet-B4 was developed using supervised learning. The dataset was divided into training, validation, and testing sets. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and F1-score. A subset of test images was independently assessed by ophthalmologists for benchmarking. Findings: The deep learning model achieved a mean AUC of 0.92 across five categories of retinal lesions. The highest performance was observed for hard exudate detection (AUC = 0.96) and retinal haemorrhage detection (AUC = 0.94), with relatively lower performance for neovascularization (AUC = 0.87). The model performed similarly to ophthalmologists and demonstrated higher sensitivity in detecting hard exudates. Conclusion: Deep learning algorithms are capable of high diagnostic accuracy in multi-lesion detection of diabetic retinopathy in fundus photographs and represent a promising adjunct to ophthalmologist-led screening; however, prospective real-world validation and further work on model interpretability are needed before routine clinical deployment.

J.Natl.Inst.Ophthalmol.2025;8(1): 56-62