Deep Learning Model for Image-Based Classification of Four Keratitis Categories

Purpose: CorneAI was developed to classify anterior segment conditions into 9 categories. We aimed to develop a deep learning (DL) model for image-based classification among 4 selected keratitis categories: catarrhal ulcer, herpes stromal keratitis, peripheral ulcerative keratitis (PUK), and phlyctenular keratitis. Methods: We retrospectively collected 973 anterior segment images from 576 patients at 7 institutions in Japan and the Japan Ocular Imaging Registry. A YOLOv5 model was developed and evaluated using five-fold cross-validation. In an exploratory analysis, 17 ophthalmologists classified 192 images from 1 randomly selected validation fold. Results: The positive predictive values (95% confidence interval) were 0.72 (0.59–0.83) for catarrhal ulcer, 0.84 (0.77–0.89) for herpes stromal keratitis, 0.88 (0.85–0.91) for PUK, and 0.72 (0.65–0.77) for phlyctenular keratitis. The areas under the receiver operating characteristic curves were 0.81 (0.74–0.87), 0.90 (0.87–0.93), 0.94 (0.93–0.96), and 0.91 (0.88–0.93), respectively. Sensitivity and specificity were 0.47 (0.37–0.58) and 0.98 (0.97–0.99) for catarrhal ulcer, 0.72 (0.65–0.79) and 0.97 (0.95–0.98) for herpes stromal keratitis, 0.93 (0.91–0.95) and 0.87 (0.84–0.90) for PUK, and 0.81 (0.75–0.86) and 0.91 (0.89–0.93) for phlyctenular keratitis. In an exploratory analysis, the diagnostic accuracies were 76.6% (95% CI, 70.1%–82.0%) for YOLOv5, 72.5 ± 9.3% (range, 53.6%–80.2%; n = 7) for corneal specialists, and 64.5 ± 10.2% (range, 45.3%–78.1%; n = 10) for ophthalmology residents. Conclusions: We developed a DL model for image-based classification among 4 selected keratitis categories. Further validation in broader clinical populations, including other corneal diseases and normal corneas, is required before clinical application.