Differentiation of Active Corneal Infections from Healed Scars Using Deep Learning.

ABSTRACT

PURPOSE: To develop and evaluate an automated, portable algorithm to differentiate active corneal ulcers from healed scars using only external photographs. DESIGN: A convolutional neural network was trained and tested using photographs of corneal ulcers and scars. PARTICIPANTS: De-identified photographs of corneal ulcers were obtained from the Steroids for Corneal Ulcers Trial (SCUT), Mycotic Ulcer Treatment Trial (MUTT), and Byers Eye Institute at Stanford University. METHODS: Photographs of corneal ulcers (n = 1313) and scars (n = 1132) from the SCUT and MUTT were used to train a convolutional neural network (CNN). The CNN was tested on 2 different patient populations from eye clinics in India (n = 200) and the Byers Eye Institute at Stanford University (n = 101). Accuracy was evaluated against gold standard clinical classifications. Feature importances for the trained model were visualized using gradient-weighted class activation mapping. MAIN OUTCOME MEASURES: Accuracy of the CNN was assessed via F1 score. The area under the receiver operating characteristic (ROC) curve (AUC) was used to measure the precision-recall trade-off. RESULTS: The CNN correctly classified 115 of 123 active ulcers and 65 of 77 scars in patients with corneal ulcer from India (F1 score, 92.0% [95% confidence interval (CI), 88.2%-95.8%]; sensitivity, 93.5% [95% CI, 89.1%-97.9%]; specificity, 84.42% [95% CI, 79.42%-89.42%]; ROC AUC, 0.9731). The CNN correctly classified 43 of 55 active ulcers and 42 of 46 scars in patients with corneal ulcers from Northern California (F1 score, 84.3% [95% CI, 77.2%-91.4%]; sensitivity, 78.2% [95% CI, 67.3%-89.1%]; specificity, 91.3% [95% CI, 85.8%-96.8%]; ROC AUC, 0.9474). The CNN visualizations correlated with clinically relevant features such as corneal infiltrate, hypopyon, and conjunctival injection. CONCLUSIONS: The CNN classified corneal ulcers and scars with high accuracy and generalized to patient populations outside of its training data. The CNN focused on clinically relevant features when it made a diagnosis. The CNN demonstrated potential as an inexpensive diagnostic approach that may aid triage in communities with limited access to eye care.