Deep learning approach for automatic diagnosis of Acanthamoeba and fungal keratitis based on in vivo confocal microscopy images
Mohammad Abolhosseini1 *, Mahmoud Essalat2 , Seyed Mohamadmehdi Moshtaghion3 , Thanh Huy Le2 , Mozhgan Rezaei Kanavi3
- Ocular Tissue Engineering Research Center, Research Institute for Ophthalmology and Vision Science, Shahid Beheshti University of Medical
Sciences, Tehran, Iran
- Department of Electrical Engineering, University of California , Los Angeles, United States, LA
- Ocular Tissue Engineering Research Center, Research Institute for Ophthalmology and Vision Science, Shahid Beheshti University of Medical Sciences, Tehran, Iran
Abstract: Infectious keratitis is a group of corneal disorders in which corneal tissues suffer inflammation and damage caused by pathogenic infections. Among these, Acanthamoeba keratitis (AK) and fungal keratitis (FK) are some of the most severe, with a high chance of permanent blindness without having an early and accurate diagnosis. In vivo confocal microscopy (IVCM) allows the imaging of different layers of the cornea, providing an important tool for an accurate diagnosis.
Methods: In this paper, we introduce the IVCM-Keratitis dataset consisting of sample images of AK, FK, as well as non-specific keratitis (NSK), and healthy corneas classes. We also use this dataset to develop multiple deep-learning models based on Convolutional Neural Networks (CNNs) for the automated diagnosis of infectious keratitis.
Results: Among these models, Densenet161, had the best performance with accuracy, precision, recall, and F1 score of 96.93%, 92.52%, 94.77%, and 93.55%, respectively.
Conclusion: The experimental results demonstrate the potential for deep learning-based models in early and automated diagnosis of AK and FK. We showed that these models can be used to highlight the areas of infection in the IVCM images and explain the reasoning behind their diagnosis using saliency maps as a technique used in eXplainable Artificial Intelligence (XAI).
