Abstract
Purpose :
It is often difficult to diagnose infectious keratitis with anterior segment examination alone. Several AI models are developed using high-quality image data, but such high-quality data is difficult to obtain constantly in daily practice. In this study, we attempted to create an AI that can predict the cause of infectious keratitis based on “real world” quality images.
Methods :
A total of 336 slit-lamp microscopic images (high-quality) of 4 categories of infectious keratitis (bacterial, fungal, viral, and acanthamoeba keratitis) were obtained at the Sankara Nethralaya. Additionally, 238 anterior segment images (wide ranges of quality) were obtained from the Internet. From a total of 564 image datasets, 484 images were used to create image classification AI models by transfer learning with EfficientNet. The performance of the AI was evaluated using 80 images that were not used for training.
Results :
Three AI models were created to predict diagnoses as 1) acanthamoeba keratitis among acanthamoeba, viral, bacterial, and fungal keratitis, 2) viral keratitis among viral, bacterial, and fungal keratitis, and 3) bacterial keratitis among bacterial and fungal keratitis. Accuracies were 83.0% for acanthamoeba keratitis, 80.0% for viral keratitis, and 70.0% for bacterial, and fungal keratitis. F-values were 70.6% for acanthamoeba keratitis, 77.5% for viral keratitis, and 70.7% for bacterial, and fungal keratitis.
Conclusions :
Current result shows the feasibility of creating AI models utilizing a mixture of varied quality images as a dataset, that is, high-quality slit-lamp images obtained at the hospital and images obtained from the Internet. Future development by obtaining larger numbers of images might improve the performance of AI, and allow AI-assisted diagnosis of infectious keratitis based on the “real world” anterior segment images which are not always acquired at high quality.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.