Developing an Artificial Intelligence-Assisted Diabetic Retinopathy Screening Model Calibrated to Nigerian Retinal Image Characteristics

Artificial intelligence (AI)-based diabetic retinopathy (DR) screening systems have demonstrated high diagnostic accuracy in clinical trials, but existing models are predominantly trained on retinal image datasets from high-income countries, limiting their applicability to Nigerian retinal image characteristics, which differ in optic disc-to-cup ratios, retinal pigmentation, and image quality constraints from lower-resource cameras. This study develops and validates an AI-assisted DR screening model calibrated to Nigerian retinal images, using deep convolutional neural network (CNN) architecture. Thirty-five thousand retinal fundus images will be collected from diabetic patients at 10 tertiary and secondary hospitals across all six geopolitical zones, spanning diverse demographic and image quality profiles. Images will be graded by trained graders using the International Clinical DR Severity Scale. The CNN will be trained using transfer learning from the EfficientNet architecture, with domain adaptation for Nigerian image characteristics. Model performance will be validated on an independent test set using AUC-ROC, sensitivity, specificity, and calibration metrics. The AI model will be integrated into a mobile-compatible telemedicine platform for deployment in primary healthcare settings. Findings will be submitted to the Federal Ministry of Health's digital health unit and the Nigeria eHealth program. This research provides an original technological contribution to ophthalmological AI for sub-Saharan Africa. Keywords: artificial intelligence, diabetic retinopathy screening, deep learning, Nigerian retinal images, telemedicine