Deep Learning Models May Be Useful in Identifying Retinopathy of Prematurity Risk

New research suggests deep learning approaches may be useful to detect retinopathy of prematurity (ROP) development in high-risk infants and reduce blindness in the population.

Findings from the retrospective study show the deep learning system provided accurate prediction of occurrence and severity of ROP before 45 weeks postmenstrual age based on retinal photographs and clinical characteristics before or at the first screening.

“With the help of our deep learning system, ophthalmologists and parents could be alerted to ensure that regular screening is performed before the onset of ROP,” wrote study authors Honghua Yu, PhD, Guangdong Eye Institute and Songfu Feng, PhD, Department of Ophthalmology, Zhujiang Hospital of Southern Medical University. “This deep learning system may also be applied to reduce the workload for pediatric ophthalmologists and to increase the treatment rate of severe ROP.”

Being the leading cause of childhood blindness, early screening for ROP can reduce rates and healthcare costs and result in multigenerational benefits for individuals, families, and society itself. As a result, early screening, regular monitoring, and timely treatment is crucial, according to the investigators.

However, due to the extensive effort required for ROP screening and low rate of treatment-requiring ROP, cost-effective programs are needed to identify infants with high risk of severe ROP.

Investigators in the current study collected data on 988 infants delivered between June 2017 and August 2019 who underwent ROP examination at 2 sites in China to develop the deep learning system. Ultimately, the study included 7033 retinal photographs of 725 infants in the training set and 763 retinal photographs of 90 infants in the external validation set, alongside 46 characteristics for each infant.

All images of both eyes from each infant taken at first screening were labeled according to the final diagnosis between screening and 45 weeks’ postmenstrual age, investigators said. Moreover, they used 2 models specifically designed for predictions of the occurrence (occurrence network [OC-Net]) and severity (severity network [SE-Net]) of ROP.

Main outcomes were considered the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity to evaluate the performance in ROP prediction.

Of the 815 infants, 450 (55.2%) were boys, the mean gestational age was 33.1 weeks (95% confidence interval [CI], 32.9 - 33.3 weeks; median, 33 weeks; range, 25 - 40 weeks) and the mean birth weight was 1.91 kg (95% CI, 1.87 - 1.95 kg).

In internal validation, using the threshold of 0.10, investigators found the mean AUC, accuracy, sensitivity, and specificity of OCT-NET to predict ROP occurrence were 0.90 (95% CI, 0.88 - 0.92), 52.8% (95% CI, 49.2% - 56.4%), 100% (95% CI, 97.4% - 100%), and 37.8% (95% CI, 33.7% - 42.1%), respectively.

Then, applying the threshold of 0.26, the mean AUC, accuracy, sensitivity, and specificity of SE-Net to predict severe ROP were 0.87 (95% CI, 0.82 - 0.91), 68.0% (95% CI, 61.2% - 74.8%), 100% (95% CI, 93.2% - 100%), and 46.6% (95% CI, 37.3% - 56.0%).

Using external validation, data show the AUC, accuracy, sensitivity, and specificity were 0.94, 33.3%, 100%, and 75%, respectively for OC-Net and 0.88, 56.0%, 100%, and 35.3%, respectively, for SE-Net.

The study, “Development and Validation of a Deep Learning Model to Predict the Occurrence and Severity of Retinopathy of Prematurity," was published in JAMA Network Open.