An Autonomous Deep-Learning System Shows Potential for Detecting Severe ROP

This study evaluated the performance of a deep learning (DL)–based autonomous artificial intelligence algorithm in identifying more-than-mild retinopathy of prematurity (mtm ROP) and type 1 ROP.

Study Design

Retinal fundus imaging datasets from the US-based Stanford University Network for Diagnosis of ROP (SUNDROP) telemedicine cohort (6245 examinations from 1545 infants) and the India-based Aravind Eye Care Systems (ACES) program (5635 examinations from 2699 infants) were included in the i-ROP DL algorithm. Sensitivity, specificity, and area under the receiver operating curves were assessed. Investigators also evaluated the diagnostic performance of the 2 fundus cameras used, RetCam and 3nethra. All SUNDROP images were taken with the RetCam, while ACES images were taken with either the RetCam or the 3nethra.

Outcomes

The iROP DL system showed good sensitivity and specificity in detecting mtm ROP and type 1 ROP in both the SUNDROP and ACES cohorts. Of note, all 155 infants who developed type 1 ROP were screened positive by the system. The algorithm had higher sensitivity but lower specificity in detecting mtm ROP and type 1 ROP from the ACES RetCam images than from the ACES 3nethra images.

Limitations

While these results suggest that an AI algorithm could help ROP prevention globally, especially in low- and middle-income countries, none of the images were from infants with stage 4 or 5 ROP (retinal detachment). Therefore, it is unclear as to how the algorithm would perform in those rare cases where a premature infant initially presents with a retinal detachment. However, the authors suggest that future autonomous ROP screening studies have a safeguard where all first examinations are reviewed manually to ensure that all pathologies are captured.

Clinical Significance

At this time, these findings do not have a direct clinical application for the pediatric ophthalmologist who diagnoses and treats ROP. However, they provide substantial evidence that autonomous ROP screening can be effective in future ROP telemedicine programs to reduce the incidence of ROP-related blindness in children worldwide. While data from the RetCam were used to develop the algorithm, the 3nethra is a more affordable option and has a sufficient field of view for ROP screening. The fundus camera issue is relevant because of the challenge to implement an affordable and precise autonomous ROP screening technology to detect severe ROP while minimizing the risk of visual loss.

Financial Disclosures: Dr. Jennifer Galvin discloses no financial relationships.