A Convolutional Neural Network Using Anterior Segment Photos for Infectious Keratitis Identification.

Abstract

Purpose: To develop a comprehensively deep learning algorithm to differentiate between bacterial keratitis, fungal keratitis, non-infectious corneal lesions, and normal corneas.

Methods: This retrospective study collected slit-lamp photos of patients with bacterial keratitis, fungal keratitis, non-infectious corneal lesions, and normal cornea. Causative organisms of infectious keratitis were identified by either positive culture or clinical response to single treatment. Convolutional neural networks (ResNet50, DenseNet121, VGG19) and Ensemble with probability weighting were used to develop a deep learning algorithm. The performance including accuracy, precision, recall, F1 score, specificity and AUC has been reported.

Results: Total of 6478 photos from 2171 eyes, composed of 2400 bacterial keratitis, 1616 fungal keratitis, 1545 non-infectious corneal lesions, and 917 normal corneas were collected from hospital database. DenseNet121 demonstrated the best performance among three convolutional neural networks with the accuracy of 0.8 (95% CI 0.74-0.86). The ensemble technique showed higher performance than single algorithm with the accuracy of 0.83 (95% 0.78-0.88).

Conclusion: Convolutional neural networks with ensemble techniques provided the best performance in discriminating bacterial keratitis, fungal keratitis, non-infectious corneal lesions, and normal corneas. Our models can be used as a screening tool for non-ophthalmic health care providers and ophthalmologists for rapid provisional diagnosis of infectious keratitis.