A deep learning approach for the segmentation and classification of Plus disease in retinal fundus images of preterm infants

Abstract

Retinopathy of Prematurity (ROP) is a severe retinal vascular disorder affecting premature infants, characterized by abnormal vessel proliferation that can lead to vision impairment or blindness if untreated. Timely and accurate diagnosis is crucial. However, current diagnostic methods often rely on subjective manual assessment, which may introduce errors. This project aims to overcome these challenges by applying artificial intelligence (AI) to enhance ROP diagnosis. The objectives within this study include developing and evaluating deep learning models for image segmentation and classification of ROP, with a particular focus on Plus disease phases. For segmentation tasks, the nnU-Net architecture was employed to create five models trained on three publicly available datasets (RETA, HVDROPDB, FIVES). These models achieved robust segmentation metrics on the training dataset (Dice coefficient similarity of 0.9525 and clDice of 0.9363). In terms of classification, models were trained on a dataset sourced from the Ophthalmic Telemedicine Network in Catalonia (RTOC), focusing on distinguishing between three labels: Plus, No-Plus, and Pre-Plus. The models achieved an impressive accuracy of 81.04%, surpassing previous studies, highlighting their ability to identify unique features associated with Pre-Plus disease. Additionally, a mockup interface was developed to visualize and interact with the diagnostic outputs of the AI models. This interface facilitates clinical integration and ongoing evaluation and refinement of the diagnostic approach for ROP.