Automated diagnosis of plus form and early stages of ROP using deep learning models

TL;DR

This paper presents an AI system that can accurately detect early signs of a serious eye disease in premature infants, potentially helping prevent vision loss.

Contribution

The novel contribution is a deep learning model achieving high accuracy for automated diagnosis of Plus disease and ROP staging using retinal fundus images.

Findings

The model achieved 0.996 accuracy for detecting Plus disease.

It also achieved 0.98 accuracy for ROP stage classification.

The system could support timely diagnosis and intervention for preterm infants.

Abstract

Retinopathy of Prematurity (ROP) represents a critical ophthalmological pathology affecting premature infants, with established associations to low birth weight (BW) and early gestational age (GA). Elevated risk of severe ROP, which can result in irreversible vision loss, is observed in infants exhibiting lower BW and GA. This research investigates the development of an automated diagnostic system designed to classify Plus disease, a marker of abnormal retinal vascularity, and ROP staging, a determinant of disease progression. Specifically, the model facilitates binary classification of Plus disease (Plus/Normal) and multi-class classification of ROP stage (Stage 0, 1, 2, 3) using a meticulously curated dataset of retinal fundus images. The proposed model demonstrates high diagnostic accuracy, achieving 0.996 for Plus disease detection and 0.98 for ROP stage classification. These results suggest potential clinical utility for automated ROP screening methodologies in supporting timely diagnosis and intervention in similar settings, pending multi-center validation, which could help reduce the incidence of vision impairment in preterm populations.

Figures