Deep Learning for the Detection of Corneal Perforation on Anterior-Segment Optical Coherence Tomography in Microbial Keratitis

Purpose: The purpose of this study was to develop and evaluate deep learning models for automated detection of corneal perforation in microbial keratitis using anterior segment optical coherence tomography (ASOCT) images. Methods: We enrolled 150 patients with microbiologically confirmed keratitis. Contralateral healthy eyes served as controls. Ground-truth labels for perforation were established following consensus grading by two masked ophthalmologist graders. A ResNet-34 backbone was used to encode six radial ASOCT scans of an eye independently and mean-pooled into a single eye-level prediction for classification of the presence or absence of corneal perforation. Four model variants were trained. Models differed in the inclusion of healthy controls and stochastic masking of non-corneal anterior segment anatomy during training. All four model variants were evaluated with 5-fold patient-level cross-validation, and the recommended model was chosen on pooled out-of-fold (OOF) test performance. Results: All four model variants achieved high discrimination, with pooled OOF test receiver operating characteristic area under the curve (ROC AUC) between 0.924 and 0.971. The best-performing model (Model 3), which did not include healthy controls or stochastic masking of the inferior image portion during training, achieved an ROC AUC of 0.971 (95% CI, 0.943–0.993), average precision (AP) of 0.863 (95% CI, 0.713–0.963), sensitivity of 0.875 (95% CI, 0.727–1.000), specificity of 0.913 (95% CI, 0.858–0.959), and F1 of 0.750 (95% CI, 0.609–0.870) at the validation-derived Youden threshold. The addition of healthy contralateral eyes to the training set did not improve pooled OOF test metrics, and stochastic inferior blackout produced opposing effects in the two training cohort settings. In the infected-only cohort, it reduced both ROC AUC and AP, whereas in the +healthy cohort, it increased ROC AUC and substantially increased AP. Conclusions: Deep learning models achieved high diagnostic accuracy for detecting corneal perforation on ASOCT imaging in eyes with microbial keratitis. These findings support the potential role of automated ASOCT analysis as a clinical decision-support tool for identifying this vision-threatening complication.