An overview of deep-learning enabled microbial keratitis classification

Infectious keratitis (IK) is the fifth leading cause of blindness globally. Gold-standard diagnostic techniques of microscopy, cultures and sensitivity from corneal scrapes are limited by low sensitivity, invasiveness, and time lag for identifying causative organisms.
This review explores how AI can enhance IK diagnosis. An English Language literature search on Medline, ISI Web of Knowledge, Science Direct and Google Scholar generated 193 studies from 2004-2024.
32 studies on classification of bacterial, fungal, viral, parasitic, non-infectious keratitis and normal corneas from anterior segment photography were evaluated. Variation was noted in reporting of ground truth definition, sampling techniques, imaging acquisition protocols and outcomes. Only 3 studies reported externally validated outcomes. Deep learning and convolutional neural networks (CNN) achieved or surpassed human experts in IK classification, with area under the receiver operator curve (AUROC) ranging from 0.5 to 0.96 and accuracy from 71% to 98%. Performance drops in areas where human evaluation is poor (e.g., classifying atypical microorganisms).
This review demonstrates how CNNs can improve IK classification. It highlights the need for external validation and standardised reporting of ground truth definition, evaluation metrics and image acquisition techniques. Promising techniques identified include knowledge enhanced multimodal classifiers and incorporating facial recognition techniques into CNN architecture.