Automated detection of early-stage ROP using a deep convolutional neural network

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinopathy of prematurity (ROP) is a leading cause of visual impairment in premature infants.
• Early detection and classification of ROP are crucial for timely intervention and preventing vision loss.
• Current diagnostic methods rely on manual fundus image interpretation, which can be time-consuming and subjective.

Purpose of the Study:

• To develop and evaluate a deep convolutional neural network (CNN) for automated detection and classification of early-stage retinopathy of prematurity (ROP).
• To assess the performance of the CNN model in differentiating between no ROP, stage 1 ROP, and stage 2 ROP using retinal fundus images.

Main Methods:

• A retrospective cross-sectional study utilizing 11,372 retinal fundus images from premature infants in Taiwan.
• A deep convolutional neural network (CNN) was implemented and trained on 90% of the images, with validation on 10% and testing on 244 images.
• Performance was evaluated using sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) compared to expert diagnosis.

Main Results:

• The CNN model achieved high accuracy, with an average training accuracy of 99.93% and testing accuracy of 92.23%.
• Exceptional sensitivity and specificity scores were reported across different classification tasks, including differentiating no ROP vs. ROP (96.14% sensitivity, 95.95% specificity).
• The model demonstrated robust performance in distinguishing between various early ROP stages.

Conclusions:

• The proposed deep learning system accurately differentiates early stages of retinopathy of prematurity.
• This AI-powered tool has significant potential to assist ophthalmologists in the early and efficient classification of ROP.
• Automated ROP detection can improve diagnostic workflows and patient outcomes.