Bespoke AI Models Differentiate Healthy Retinas from ROP Features

Bespoke and code-free deep learning (CFDL) models demonstrate the ability to differentiate between healthy retinal images and those displaying features of pre-plus or plus disease, thus providing an indication of retinopathy of prematurity (ROP), according to a study published online in the Lancet Digital Health.

In the retrospective cohort study led by Siegfried K. Wagner, M.D., of the NIHR Moorfields Biomedical Research Center in London, a total of 1,370 neonates admitted to a neonatal unit between 2008 and 2018 were included. The study involved the development of bespoke and code-free deep learning (CFDL) models.

The internal performance of these models was evaluated using 200 retinal images, followed by external validation on 338 retinal images obtained from four distinct datasets originating from the United States, Brazil, Egypt (Retcam images), and India (images from the 3nethra neo device).

The results of the study revealed that the bespoke model achieved an impressive area under the curve (AUC) of 0.986, while the CFDL model obtained an AUC of 0.989 when tested internally to distinguish between healthy retinas and those with pre-plus or plus disease.

When subjected to external validation tests utilizing Retcam images, both models exhibited strong generalization capabilities in discriminating healthy retinas from those with pre-plus or plus disease. The bespoke model demonstrated a range of 0.975 to 1.000, while the CFDL model exhibited a range of 0.969 to 0.995.

In the U.S. dataset, the performance of the CFDL model was found to be inferior to that of the bespoke model for discriminating pre-plus disease from healthy retinas or plus disease (CFDL: 0.808; bespoke: 0.942). Similarly, when tested on the 3nethra neo imaging device, both models exhibited reduced performance compared to the internal test sets (CFDL: 0.865; bespoke: 0.891).

"Although further validation and studies of effectiveness across different populations are needed before deployment, deep learning might provide a tool for mitigating the risk of lifelong sight impairment in these young patients," the authors concluded.

Reference

Siegfried K Wagner et al, Development and international validation of custom-engineered and code-free deep-learning models for detection of plus disease in retinopathy of prematurity: a retrospective study, The Lancet Digital Health (2023). DOI: 10.1016/S2589-7500(23)00050-X