Comparative Evaluation of Multiple Deep Learning Models for Six-Class Keratitis Classification

Keratitis can be classified as infectious or noninfectious, with infectious keratitis being a major cause of blindness. In this study, multiple deep learning (DL) models were applied to perform six-class classification of corneal lesions, and their diagnostic performance was systematically compared.

We analyzed 3797 slit-lamp images from six categories: normal cornea, bacterial keratitis, fungal keratitis, viral keratitis, Acanthamoeba keratitis, and non-infectious keratitis. Convolutional neural network (CNN), transformer, self-supervised, and vision-language models were evaluated at image and patient levels using patient-level splits, with clinician comparison on a 240-image test subset.

RegNet had the best image-level performance (accuracy, 76.8%; macro-F1 score, 75.7%; macro-AUC, 0.946), and EfficientNet_B0 had the highest patient-level accuracy (77.5%; macro-F1 score, 76.6%; macro-AUC, 0.954). On the same 240-image subset, RegNet accuracy was 75.4% versus 72.7%, 56.3%, and 44.4% for senior corneal specialists, junior corneal specialists, and junior general ophthalmologists.

CNN-based models showed stable performance under the present image-based evaluation pipeline. RegNet achieved the most balanced image-level results, whereas EfficientNet_B0 showed the strongest patient-level performance. On the same 240-image subset, RegNet showed numerically higher accuracy than clinician-group means. However, performance remains insufficient for autonomous diagnosis and requires external and prospective validation.

This comparative DL framework could serve as an adjunctive image-based decision-support tool for six-class keratitis classification after further external and prospective validation, which may assist clinicians in early diagnosis and improve decision making in clinical management.