Association of Deep Learning Imaging Algorithm Measures of Microbial Keratitis With Vision Outcomes

To compare associations between 90-day vision and microbial keratitis-related corneal changes when predicted by an artificial intelligence image-based algorithm or when measured by a clinician.

This prospective cohort study followed patients with microbial keratitis in the United States and India for up to 90 days. Participants' clinical history, best-corrected visual acuity (BCVA), slitlamp imaging, and anterior exam findings were collected. A novel, artificial intelligence imaging segmentation algorithm was developed to quantify relevant corneal morphologic features, including presenting stromal infiltrate (SI) area and hypopyon presence using corneal photography. Multivariable linear regression models assessed associations of presenting features, including algorithm-predicted measurements, with 90-day logMAR BCVA. Results are reported with model estimates (β) and 95% confidence intervals.

The study included 479 participants (N = 138 United States, N = 341 India). In the United States, larger algorithm-predicted SI area [β = 0.02 per 1 mm2 area increase for bacterial ulcers (P = 0.002); β = 0.09 per 1 mm2 increase for fungal ulcers (P < 0.05); and β = 0.25 per 1 mm2 increase for viral ulcers (P = 0.0095)] was associated with worse 90-day vision. In India, larger algorithm-predicted SI area (β = 0.05 per 1 mm2 area increase for fungal ulcers, P < 0.001) and algorithm-predicted hypopyon (β = 0.22, 95%, confidence interval 0.02-0.41; P = 0.0290) were associated with worse 90-day vision. Algorithm-based measurements of SI area and hypopyon presence showed a similar association with 90-day logMAR BCVA, in magnitude and significance, to those using clinicians' measurements.

Worse vision outcomes were associated with larger SI area and hypopyon presence when predicted by the imaging segmentation algorithm or measured by clinicians in both cohorts.