Deep learning-assisted interpretable retinopathy of prematurity (ROP) diagnosis

Abstract

Retinopathy of Prematurity (ROP) is a sight-threatening retinal eye disease primarily affecting premature babies due to the growth of abnormal blood vessels in the retina. Early detection and timely treatment is very important to prevent irreversible vision loss; however, effective screening for ROP is limited by the lack of resources and trained ophthalmologists, especially in underserved and rural areas. In recent years, several studies have been conducted on the development of reliable AI-based screening systems. However, due to the lack of well-annotated public datasets most of them gives partial diagnostic solutions and are limited to experimental research, using single-central datasets. Our work presents a deep learning-assisted diagnostic framework for automated and interpretable ROP screening. The proposed system is composed of three key modules: (1) zone separation, (2) ridge (or demarcation line) detection, and (3) blood vessel segmentation—each targeting clinically relevant retinal features defined by the International Classification of ROP (ICROP) guidelines. To support and evaluate the framework, we created a comprehensive and expert-annotated dataset - Macretina, consisting of 1,432 retinal fundus images from 112 premature infants. Finally, we integrated our novel diagnostic framework into a lightweight mobile application, designed for real-time deployment in neonatal care units. Our proposed solution demonstrates high diagnostic accuracy, interpretability, and scalability, offering a clinically viable tool for early ROP screening, especially in low-resource and telemedicine settings.