Implementing AI-supported Screening for Retinopathy of Prematurity in Low-resource setting on a Relatively Small Set of Data: A Retrospective Cohort Study

Background: Context-adapted screening programs supported by AI-enabled teleretinopathy of prematurity (ROP) systems are essential to improve access, efficiency, and early diagnosis in underserved settings, particularly for often overlooked diseases such as ROP. This study aims to evaluate the diagnostic performance of an AI-based model for detecting and classifying ROP using wide-field retinal images and assess its relevance for neonatal screening systems in limited and middle-income countries (LMIC). Methods: This retrospective cohort study analyzed 5531 RetCam II images from 950 preemies (2018–2023) at Eye World Hospital, Giza, Egypt. Images were graded using international classification of ROP (ICROP) 3rd edition by a trainee and confirmed by a vitreoretinal consultant. A CNN-based model was developed to classify ROP stage, zone, and severity after augmentation up to 8500 images. Performance was assessed using standard metrics. Results: A total of 5531 RetCam II images from 950 preemies were analyzed. The AI model demonstrated good performance in detecting advanced ROP, particularly treatment-requiring stages (stages 4 and 5 and type 1 ROP), with high precision and recall. Overall classification accuracy was 66%. Performance was lower in early-stage disease and in differentiating closely related subtypes, particularly aggressive type ROP (A-ROP) and early stages. Conclusion: The CNN model based on 5531 RetCam II images, augmented up to 8500 images, shows good performance in detecting advanced ROP but limited sensitivity for early stages. Constraints include small, imbalanced data and reduced generalization. The system is intended for triage and decision support in tele-ROP screening, not autonomous diagnosis. Future improvements require larger multicenter datasets, better class balance, and advanced architectures with transfer learning to enhance early detection accuracy and clinical reliability.