Abstract
Purpose :
In vivo confocal microscopy (IVCM) has become a useful tool in diagnosing infectious keratitis, with different ideologies having distinct appearances on confocal imaging. Our retrospective study aims to evaluate the efficacy of deep learning in discerning between viral keratitis, acanthamoeba keratitis (AK), and non-infectious normal control IVCM images of the cornea.
Methods :
In a retrospective design, 53,900 confocal eye images taken between August 2021 and August 2023 were analyzed: 25,114 images from 69 individuals without corneal pathology, 9,941 images from 9 patients diagnosed with viral keratitis, and 18,845 images from 9 patients diagnosed with acanthamoeba keratitis. Diagnoses were made clinically by a board-certified ophthalmologist. Two Advanced Convolution (ResNet50) and Vision Transformer (FastViT_SA12) neural networks were employed for the classification of the three groups. The training set incorporated 44,702 images, with the test set containing 9,198 images. Algorithmic performance metrics included average One-vs-Rest group Area Under the Curve (AUC), accuracy, sensitivity, specificity, Positive Predictive Value (PPV), and Negative Predictive Value (NPV).
Results :
The FastViT model yielded an AUC of 0.852 and an accuracy of 69.7%. The ResNet50 model achieved an AUC of 0.843 and an accuracy of 69.8%. When identifying AK, the two models averaged a specificity of 87.2%, sensitivity of 64.5%, PPV of 67.4%, and NPV of 82.4%. When identifying viral keratitis, the two models averaged a specificity of 86.2%, sensitivity of 48.4%, PPV of 50.1%, and NPV of 77.7%.
Conclusions :
Our research highlights the significant potential of machine learning in accurately classifying confocal eye images. The performance in accuracy and sensitivity metrics of our models indicates a promising approach for the early diagnosis of infectious keratitis. This is crucial for timely clinical intervention and improved patient prognosis.
This abstract was presented at the 2024 ARVO Imaging in the Eye Conference, held in Seattle, WA, May 4, 2024.