Clinically Deployable ResNet50 AI Model for Diabetic Retinopathy Screening: A Robust Multicenter Validation

Abstract

Diabetic retinopathy (DR) is still the most common preventable cause of blindness in the world, but screening programs often can't do their jobs because they don't have enough skilled staff and specialized tools. This study thoroughly tests a ResNet50-based deep learning model for separating referable DR from retinal fundus images, with a focus on how it can be used in a variety of healthcare situations. The model was made with 5436 pictures from the Messidor, APTOS, and EyePACS datasets, and reference marks were set up by ophthalmologists who are board-certified. The model got 93.14% accuracy, 90.24% sensitivity, and 95.35% precision on a separate set of 815 pictures. It also had an area under the receiver operating characteristic curve of 0.963. Gradient-weighted Class Activation Mapping showed that 89.3% of model attention maps matched abnormal traits that are important for clinical practice. With inference times of 1.87 s on central processing units and 0.12 s on graphics processing units, the model showed strong computing performance. In this study, an open-source, fully confirmed artificial intelligence model for DR screening is created with real-world usefulness and diagnostic accuracy in mind. This fills in important gaps in the field of medical artificial intelligence.

Received: 28 November 2025 | Revised: 26 March 2026 | Accepted: 31 March 2026

Conflicts of Interest

The author declares that they have no conflicts of interest to this work.

Data Availability Statement

The data used in this study are derived from three publicly available repositories: the Messidor dataset (available at https://www.adcis.net/en/third-party/messidor/), the Kaggle APTOS 2019 Blindness Detection dataset (available at https://www.kaggle.com/c/aptos2019-blindness-detection/data), and the EyePACS dataset (available at https://www.kaggle.com/c/diabetic-retinopathy-detection/data and also as a curated composite at https://www.kaggle.com/datasets/ascanipek/eyepacs-aptos-messidor-diabetic-retinopathy). No new primary data were generated or collected by the authors for this study; all data processing, model training, and validation were performed using these publicly available sources. The code, trained model weights, and detailed implementation instructions are available from the corresponding author upon reasonable request and are intended for noncommercial research purposes.

Author Contribution Statement

Gabriel Silva-Atencio: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Resources, Data curation, Writing - original draft, Writing - review & editing, Visualization, Supervision, Project administration.