Generation of Artificial Intelligence Models for Predicting the Etiological Agents of Infectious Keratitis

To develop and validate an artificial intelligence (AI) model for classifying infectious keratitis (IK) etiologies from anterior segment images. Retrospective development and validation study of an artificial intelligence diagnostic system. A total of 708 anterior segment images were collected, comprising Acanthamoeba (n=159), bacterial (n=139), fungal (n=188), and viral (n=222) cases. Of these, 628 images were used to train three convolutional neural network architectures (DenseNet-121, ResNet-50, and EfficientNet-B6) using stratified 5-fold cross-validation. Model performance was then assessed on an independent 80-image test set, with diagnostic accuracy benchmarked against the averaged diagnoses of 33 board-certified ophthalmologists. The validated model was subsequently integrated into a smartphone application. EfficientNet-B6 achieved the highest performance among tested architectures, achieving a mean accuracy of 61.5% (95% CI: 57.3-65.7%) across pathogen categories and significantly outperforming ophthalmologists, whose mean accuracy was 39.1% (95% CI: 36.4-41.8%). Receiver Operating Characteristic analysis was performed using the model from the best-performing cross-validation fold, evaluated on the independent 80-image test set. The analysis showed Area Under the Curve (AUC) values of 0.84 for AK, 0.84 for BK, 0.88 for FK, and 0.87 for VK. The smartphone-based application incorporating our AI model demonstrated a diagnostic accuracy comparable to that of the computer-based system when the same test dataset was re-captured from a monitor using the smartphone camera, suggesting successful translation of the algorithm to mobile platforms. Our smartphone-based AI system demonstrates moderate-to-good diagnostic performance for identifying IK pathogens using diffuser-anterior segment photography, thereby enabling practical implementation via smartphone cameras.