AI-MK: artificial intelligence for assessing and monitoring microbial keratitis

To evaluate the performance of an artificial intelligence (AI) model for detecting and monitoring microbial keratitis (MK) using anterior segment optical coherence tomography (AS-OCT).

This is a prospective observational study. Patients with clinically suspected MK and healthy participants were included. In addition to routine assessment and treatment with topical fluoroquinolone therapy, patients underwent AS-OCT at each clinic visit. These images were tested on our DeepLabV3 network-based AI model, which aims to diagnose and record changes to infiltrate sizes of MK lesions over time.

The AI model accurately captured MK lesions in 93% of cases (152/163). MK was not detected in scans from healthy eyes, and there were no cases of artefact being falsely detected. The model had a sensitivity of 93% (95% CI 88% to 97%), specificity of 100% (95% CI 88% to 100%), positive predictive value of 100% (95% CI 98% to 100%) and negative predictive value of 73% (95% CI 61% to 83%). Using only the corneal component with masking of the anterior chamber, the AI model showed agreement on change with both observers in 76% (13/18) cases.

This AI framework reliably identified MK lesions using AS-OCT, with high sensitivity and specificity. The framework was able to identify change in most cases compared with corneal specialists.