Artificial intelligence-assisted diagnosis and subtype differentiation of infectious keratitis.

Abstract

Background: Infectious keratitis (IK) is a major cause of corneal blindness world-wide, and prompt identification of IK and its etiologic subtype is essential for appropriate management. We developed deep learning (DL) models to detect IK and differentiate common subtypes from slit-lamp photographs.

Methods: In this retrospective study, slit-lamp photographs were collected from patients presenting to the emergency department of Farabi Eye Hospital (2014-2021) with bacterial keratitis (BK), fungal keratitis (FK), Acanthamoeba keratitis (AK), or herpes simplex keratitis (HSK), along with healthy controls and corneal scars. A total of 13,953 images were included. Three DL classifiers were trained: Model 1 (IK vs. normal), Model 2 (healthy vs. corneal scar vs. IK [pooled subtypes]), and Model 3 (BK vs. FK vs. AK vs. HSK).

Results: Model 1 achieved 99.9% accuracy for IK vs. normal (ROC-AUC 0.999). In five-fold cross-validation, Model 2 achieved mean accuracy 0.975 (95% CI 0.955-0.996), macro-F1 0.970 (95% CI 0.945-0.995), and macro-average AU-ROC 0.998 (95% CI 0.995-1.000). For subtype classification (Model 3), overall accuracy was 81.6% with balanced recall 83.3%; class accuracies were 88% (BK), 71% (FK), 72% (AK), and 93% (HSK) with ROC-AUCs 0.90-0.98. External validation of Model 3 (Ankara City Hospital; 665 images from 96 patients) showed accuracy 92.5%, macro-F1 93%, macro-average AUROC 0.996, and sensitivities of 95.2% (AK), 92.0% (BK), 85.5% (FK), and 99.5% (HSK).

Conclusions: DL models applied to slit-lamp photographs showed high performance for IK detection and clinically relevant differentiation of IK from corneal scars and among major IK subtypes, with external validation supporting generalisability.