Abstract
Purpose :
Out of various causes of corneal opacity, microbial keratitis (MK) stands out as the foremost reason for corneal blindness. The current diagnostic methods have their drawbacks, including experience-dependency, tissue damage, and time consumption. We performed a retrospective artificial intelligence (AI)-based slit-lamp data analysis to detect MK, differentiate among bacterial, fungal, and Acanthamoeba keratitis (AK), and distinguish between subtypes of fungal keratitis (FK).
Methods :
The dataset includes 10,739 slit-lamp images, categorized into four primary classes: 2,505 images of healthy corneas, 2,008 of FK, 4,816 of bacterial keratitis, and 1,410 of AK, sourced from 1,428 subjects. Three AI-based models employing Convolutional Neural Networks (CNN) were developed for the following purposes: first, to diagnose MK from healthy corneas; second, to differentiate among bacterial, fungal, and Acanthamoeba infections; and third, to distinguish between filamentous and yeast subtypes of FK cases (Figures 1,2). We calculated accuracy, sensitivity, specificity, precision, F1 score, R2 score, as well as area under the receiver operating characteristic curve (AUC-ROC).
Results :
Model 1 for diagnosing healthy corneas and MK achieved 99.01% and 99.23% of accuracy rate, respectively with a 0.99 AUC-ROC. Model 2 for differentiating bacterial, fungal, and Acanthamoeba infections reached accuracies of 91.91%, 79.77%, and 81.27% with AUC-ROCs of 0.88, 0.87, and 0.90, respectively. Model 3 for discriminating fungal subtypes attained an accuracy of 76.19% and 77.84% for diagnosing filamentous and yeast subtypes, respectively with an AUC-ROC of 0.78. Findings were validated using 5-fold cross-validation.
Conclusions :
The developed AI-based models are promising in early and precise diagnosis of MK, enhancing the differentiation of bacterial, fungal, and Acanthamoeba keratitis, and identifying fungal subtypes. These models may significantly aid timely and appropriate treatment interventions, especially in resource-privileged populations, thus mitigate health disparities. It will also reduce costs related to complications arising from late diagnosis.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.