Automated Differentiation of Bacterial from Fungal Keratitis Using Deep Learning

Abstract

Purpose : Infectious keratitis is a leading cause of blindness worldwide. Efforts to address this burden are limited by clinically significant gaps in our ability to rapidly diagnose the causative pathogen. Culture results are delayed and have poor sensitivity, and prior studies demonstrate that human experts are only accurate 66% of the time when differentiating bacterial from fungal keratitis based on clinical appearance. This results in delayed initiation of appropriate antimicrobial therapy and worsened visual outcomes. Herein we address this gap by developing and evaluating a deep learning artificial intelligence system to distinguish culture-proven bacterial and fungal keratitis using corneal photographs.

Methods : We established an image database by collating 1,010 external photographs from handheld digital cameras obtained as part of several randomized clinical trials conducted by the Aravind Eye Care System in India and the Proctor Foundation in San Francisco since 2006. 940 images were used to train the deep learning system, and 70 were reserved as an independent testing set. 50% of the training image set consisted of images from bacterial ulcers, and 50% were from fungal ulcers. We used a transfer learning approach adapting a pre-trained deep learning system (EyeMeter) built on the ResNet-50 convolutional neural network architecture.

Results : This deep learning system demonstrated 76% overall accuracy for differentiating bacterial and fungal corneal ulcers from photographs alone. For detecting bacterial keratitis, the system had 70% sensitivity and 80% specificity. For detecting fungal keratitis, the system had 80% sensitivity and 70% specificity.

Conclusions : A deep learning system trained using photographs from inexpensive handheld cameras was able to distinguish bacterial from fungal keratitis with greater accuracy than the published performance of expert human cornea specialists. Future incorporation of similar technology into mobile telemedicine platforms may reduce blindness from infectious keratitis, particularly in low- and middle-income countries where disease burden is highest but clinical and microbiologic expertise are scarce. Future directions include collection of additional images from geographically diverse sources to optimize and validate this technology.

This is a 2021 ARVO Annual Meeting abstract.