Development and Validation of a Multilabel Computer Vision Model for Identification of Bacterial and Fungal Keratitis in a Prospectively Collected Dataset
Travis K Redd1, N. Venkatesh Prajna2, Prajna Lalitha2, Jeremy D Keenan3, J Peter Campbell1, Xubo Song1
1Casey Eye Institute, Oregon Health & Science University, Portland, Oregon, USA; 2Aravind Eye Hospital, Madurai, India;
3Francis I. Proctor Foundation, University of California San Francisco, San Francisco, California, USA
Purpose: Prior studies have demonstrated that corneal photographs collected from a database of randomized clinical trials can be used to train deep learning models to differentiate bacterial and fungal keratitis with moderately high reliability. In this study we assess the generalizability of this technology using a prospectively collected dataset and attempt to develop a multilabel model capable of recognizing polymicrobial infection.
Methods: Subjects with a clinical diagnosis of infectious keratitis were recruited from the Aravind Eye Hospital in Madurai, India. Standardized corneal photographs from cases with positive cultures and smears (Gram stain or KOH) were used to train and evaluate a MobileNet deep convolutional neural network to identify bacterial and fungal keratitis. The image set was randomly partitioned into training/validation (80%) and hold-out test sets (20%). Five-fold cross validation was used to identify optimal hyperparameters for model training. The model was developed in the python programming language with Tensorflow 2.0.
Results: A total of 426 subjects were included. Based on culture and smear results, 108 subjects were diagnosed with bacterial keratitis, 326 with fungal keratitis, and 8 with both bacterial and fungal keratitis (polymicrobial). MobileNet achieved an area under the curve of 0.84 in the hold-out test set, which is comparable to prior models trained for binary classification using images collated retrospectively from clinical trials.
Conclusions: These results demonstrate that deep learning models developed from prospectively collected, population-based datasets can accurately perform image-based identification of bacterial and fungal keratitis at or above the level of expert cornea specialists. This suggests a high potential for generalization of this technology to real world applications. This is also the first such model trained to identify polymicrobial infections consisting of both bacterial and fungal pathogens. Future iterations of these computer vision models may have utility for guiding antimicrobial therapy in the absence of microbiologic results.
Support: NIH (K12EY027720 and P30EY10572) and unrestricted departmental funding from Research to Prevent Blindness