AI-Assisted Microbial Keratitis Diagnosis

An artificial intelligence (AI)-assisted tool has been shown to reliably diagnose microbial keratitis (MK) and accurately distinguish among bacterial, fungal, and parasitic infections based on slit-lamp photographs. The tool, which is in development, could improve outcomes by enabling preliminary diagnoses sooner, avoiding the complications and extra costs associated with delayed or incorrect treatment. 

The technology also could help mitigate health disparities by reducing the expense and high expertise currently needed to diagnose MK and distinguish among infection subtypes, which is critical to guide effective treatment.

“Treatment of bacterial keratitis is very different from fungal keratitis,” as are treatment of acanthamoeba, and of filamentous yeast fungal subtypes, said Mohammad Soleimani MD.

Accurate MK diagnosis now requires specialised cornea experience and expensive equipment, such as confocal scanning microscopes costing about US$75,000 each, which limits access, Dr Soleimani said. Culturing infected tissue samples is also time consuming, expensive, and not always available or reliable. “We need something more affordable and user-friendly, especially in resource-limited settings.”

Deep-learning models

With that goal in mind, Dr Soleimani developed 3 diagnostic models using convolutional neural networks to analyse more than 10,000 slit-lamp images from about 1,400 patients. These included about 2,000 healthy cornea images, 2,000 fungal keratitis images, 4,800 bacterial keratitis images, and 1,400 acanthamoeba keratitis images. About 80% of the images went towards training the deep-learning models and 20% for validation.

The first model, which distinguished between healthy and MK corneas, proved the most accurate, correctly diagnosing more than 99% of cases. The second model distinguished among bacterial, fungal, and acanthamoeba in about 80% of cases, reaching accuracies of 91%, 80%, and 81%, respectively. The third model successfully distinguished filamentous and yeast fungal subtypes in about 77% of cases, reaching 76% and 78%, respectively.

By comparison, experienced cornea subspecialists correctly differentiate such cases about 50% of the time. Acanthamoeba is particularly hard to diagnose, with accuracy sometimes running around 10%.

Making an app

To strengthen the models’ performance, Dr Soleimani—with collaborators in Germany, the United Kingdom, Canada, the United States, China, South America, India, and the Middle East—are adding images from diverse patient populations. Validation tests in large external data sets are also underway. External validation is critical to ensure reliability before clinical use.

The ultimate goal is to create an AI-powered cell phone app that will allow diagnosis using images from a variety of cameras, including any kind of digital or cell phone camera, Dr Soleimani said. That way, the tool can be used not just by ophthalmologists, but other types of eye care and health professionals.

“It needs to be available and user-friendly so it can be used anywhere in the world” without specialised cornea training or equipment, Dr Soleimani said. He anticipates the app will be clinically available in two to three years.

Dr Soleimani spoke at ARVO 2024 in Seattle, US.

Mohammad Soleimani MD, FICO is a cornea and ocular surface sub-specialist and professor of ophthalmology at the University of North Carolina, US. msolei2@uic.edu