An Artificial Intelligence-Based Model to Detect Acanthamoeba Keratitis on Confocal Microscopy

Abstract

Purpose : Acanthamoeba keratitis (AK) has one of the worst prognoses among infectious keratitides (IKs), with late diagnosis and misdiagnosis being significant contributors to this. The application of in-vivo confocal microscopy (IVCM) holds promise in aiding in AK diagnosis, but the interpretation of images requires specialized training. This diagnostic research study based on retrospective IVCM images aims to leverage AI-based models for early and accurate diagnosis of AK and to discriminate AK from non-acanthamoebal corneal infiltrations.

Methods : The dataset comprises 2,764 IVCM images from patients who were seen at Massachusetts Eye and Ear with a culture-confirmed diagnosis of AK. 2,507 IVCM images allocated for training the algorithm and 257 images for evaluation by the algorithm. The images were divided into two principal categories: 1,402 training and 173 evaluation images for AK, and 1,265 training and 173 evaluation images for other types of corneal infiltrations. Labeling was conducted by two experienced raters, with only cases that received unanimous agreement included in the training dataset. An AI-based model, employing convolutional neural networks (CNN), was developed to diagnose AK and distinguish it from other causes of corneal infiltrations.

Results : The inter-rater agreement on presence of AK on IVCM was 84.0%. The model showed a reasonable performance, with an accuracy rate of 79.4% and a Receiver Operating Characteristic - Area Under Curve (ROC-AUC) score of 0.77 for detecting non-acanthamoebal keratitis. It also had an accuracy rate of 74% with a ROC-AUC of 0.77 for identifying AK. The overall mean accuracy was 76%. The model achieved a recall (sensitivity) of 0.76 with a 95% Confidence Interval (CI) of 0.73-0.79 and a specificity of 0.76 with a 95% CI of 0.73-0.79.

Conclusions : This study presents an AI model utilizing CNN as a promising tool in early and accurate diagnosis of AK based on IVCM images. This approach may improve early diagnosis of AK where IVCM is available and may improve visual outcomes in this vision-threatening disease.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

A sample input image and its corresponding feature maps, extracted from the first layer of the designed network.

 

Architecture of the deep convolutional neural network (CNN) model. The CNN model diagnoses IK and then differentiates Acanthamoeba and other types of keratitides.