No-code bespoke AI model beneficial for retinal imaging

A custom-made code-free deep learning (CFDL) model can distinguish healthy retinal images from those with pre- or plus disease features indicative of retinopathy of prematurity (ROP), a research paper appeared online April 21. announced in lancet digital health.

Siegfried K. Wagner, M.D., of the NIHR Moorfields Biomedical Research Center in London, and colleagues conducted a retrospective cohort study using retinal imaging of 1,370 newborns admitted to the neonatal ward between 2008 and 2018. . A bespoke model and a CFDL model were developed. Performance was evaluated internally on his 200 images, then his 338 from his four separate datasets in the United States, Brazil, Egypt (Retcam images) and India (images from the 3nethra neo device). It was validated externally on single retinal images.

The researchers found an area under the curve (AUC) of 0.986 for the tailor-made model and an AUC of 0.989 for the CFDL model in the internal test set for distinguishing between healthy and pre-plus or plus disease. Both models generalized well in distinguishing between healthy and pre-plus or plus disease on the external validation test set acquired using Retcam (custom range, 0.975 to 1.000; CFDL range, 0.969 to 0.995). The CFDL model was inferior to the bespoke model when distinguishing between pre-plus disease and healthy or plus disease in the US dataset (CFDL: 0.808, bespoke: 0.942). Performance was also degraded when tested on the 3nethra neo imaging device (CFDL: 0.865, Bespoke: 0.891).

“Although further validation and research of efficacy across different populations is needed before implementation, deep learning could be a tool to reduce the risk of lifelong vision loss in these young patients,” the authors said. writing.

Several authors disclosed ties to the pharmaceutical industry.