Advancements In The Classification Of Retinopathy Of Prematurity: An Overview

Abstract

Retinopathy of Prematurity (ROP) is a major cause of preventable childhood blindness worldwide, particularly 
affecting premature and low birth weight infants. Timely and accurate classification of ROP is critical for effective 
clinical intervention. This paper presents a comprehensive overview of recent technological advancements in the 
classification of ROP, focusing on the integration of artificial intelligence (AI), deep learning (DL), and image 
processing techniques. Convolutional Neural Networks (CNNs), transfer learning models such as ResNet, 
InceptionV3, and VGG16, and hybrid architectures combining CNN with LSTM have shown promising 
performance in detecting disease stages, zones, and plus disease from retinal fundus images. Additionally, attention
based models and ensemble methods have been explored to enhance classification accuracy and model interpretability. 
Despite these advancements, several limitations persist. The lack of large, annotated, and standardized ROP datasets 
restricts model generalizability across diverse populations. Variations in image quality, illumination, and field-of-view 
introduce noise and hinder consistent classification. Moreover, most current systems operate as black boxes, offering 
limited transparency to clinicians regarding decision-making rationale. The future scope lies in developing explainable 
AI frameworks, federated learning models for cross-institutional collaboration without data sharing, and multi-modal 
approaches integrating clinical and demographic data. Integration of telemedicine-based diagnostic platforms and real
time screening tools can further expand the reach of ROP care in resource-limited settings. This paper underscores the 
need for interdisciplinary collaboration to translate these technologies into scalable, ethical, and clinically acceptable 
tools for ROP management.