Impact of Mydriasis on Image Gradability and Automated Diabetic Retinopathy Screening with a Handheld Camera in Real-World Settings

Download PDF
Open Webpage

Iago Diogenes, David Restrepo, Lucas Zago Ribeiro, Andre Kenzo Aragaki, Fernando Korn Malerbi, Caio Saito Regatieri, Luis Filipe Nakayama

Published: 02 Jan 2025, Last Modified: 08 Jan 2026CrossrefEveryoneRevisionsCC BY-SA 4.0
Abstract: h3>Abstract</h3> <h3>Purpose</h3> <p>Diabetic retinopathy (DR) screening in low- and middle-income countries (LMICs) faces challenges due to limited access to specialized care. Portable retinal cameras provide a practical alternative, but image quality, influenced by mydriasis, affects artificial intelligence (AI) model performance. This study examines the role of mydriasis in improving image quality and AI-based DR detection in resource-limited settings.</p><h3>Methods</h3> <p>We compared the proportion of gradable images between mydriatic and non-mydriatic groups and used logistic regression to identify factors influencing image gradability, including age, gender, race, diabetes duration, and systemic hypertension. A ResNet-200d algorithm was trained on the mBRSET dataset and validated on mydriatic and non-mydriatic images. Performance metrics, such as accuracy, F1 score, and AUC, were evaluated.</p><h3>Results</h3> <p>The mydriatic group had a higher proportion of gradable images (82.1% vs. 55.6%, <i>P</i>&lt; 0.001). Factors such as systemic hypertension, older age, male gender, and longer diabetes duration were associated with lower image gradability in non-mydriatic images. Mydriatic images achieved better AI performance, with accuracy (82.91% vs. 79.23%), F1 score (0.83 vs. 0.79), and AUC (0.94 vs. 0.93). Among gradable images, the performance difference was not statistically significant.</p><h3>Conclusion</h3> <p>Mydriasis improves image gradability and enhances AI model performance in DR screening. However, optimizing AI for non-mydriatic imaging is critical for LMICs where mydriatic agents may be unavailable. Refining AI models for consistent performance across imaging conditions is essential to support the broader implementation of AI-driven DR screening in resource-constrained settings.</p>