Abstract
PURPOSE: Retinopathy of prematurity (ROP) is an eye disease that affects premature infants. Early diagnosis is important to prevent vision loss and enable timely intervention. The work aims to develop a deep learning based automated ROP detection and classification pipeline from a large retinal image dataset. MATERIAL AND METHODS: A dataset of 400 GB containing retinal fundus images was obtained from Narayana Nethralaya, Bengaluru, India. Both blood vessel and ridge feature segmentation were performed using a U-Net model. Gabor enhanced retinal images were used for ridge segmentation, while original retinal images were used for blood vessel segmentation as original images preserve critical vascular features. The segmented ridge and blood vessel masks were superimposed on sigmoid-enhanced retinal images and used as input to a ResNet50 classifier. The aim of this pipeline was to preserve the most prevalent disease features to facilitate in the stage wise classification of ROP. RESULTS: The pipeline achieves an accuracy of 98.40% in detecting ROP and 92.80% in staging ROP. This method improves image-based ROP screening by effectively identifying critical disease features. CONCLUSION: The results of this automated pipeline demonstrate its potential in supporting an early detection of ROP with high accuracy and efficiency. The integration of deep learning models such as U-Net and ResNet50 during the screening process helps support clinical decision making and advances neonatal ophthalmology practice.