Utilization of automated machine learning approach toward detection of granular corneal dystrophy using slit-lamp photographs

Abstract

Purpose : This index study aims to apply automated machine learning (AutoML) techniques for the diagnosis of granular corneal dystrophy, a rare inherited condition characterized by progressive proteinaceous deposition in the corneal stroma.

Methods : Patients with a confirmed diagnosis of granular corneal dystrophy and slit-lamp photos of their affected eye(s) were enrolled in the study. Individuals with concomitant corneal conditions, ungradable imaging data, or uncertain diagnoses were excluded from the study. Slit-lamp photos depicting granular corneal dystrophy were obtained from the Cornea Service at Byers Eye Institute, Stanford University, and Shiley Eye Institute, University of California San Diego (UCSD). Slit-lamp photos of normal corneas were retrieved from the same institutions' image databases. These images underwent resizing and cropping, focusing solely on the cornea. Subsequently, we employed a deep learning model, utilizing Vertex-AI, the AutoML platform developed by Google (Menlo Park, CA). The area under the precision-recall curve (AUPRC) was plotted and sensitivity, specificity, positive predictive value (PPV), and accuracy (AC) were calculated.

Results : The model was trained using a dataset comprising 89 images, consisting of 72 granular corneal dystrophy cases and 17 healthy corneas. Seventy-one images were used for training, 9 were used for validating, and 9 were used for testing the model. AUPRC for the dataset was found to be 0.96 (Figure 1). The sensitivity, specificity, PPV, and AC of the model were 86%, 100%, 100%, and 88.64%. Figure 2 shows the confusion matrix heatmap of the trained model.

Conclusions : A clinician-derived machine learning model developed without coding was able to differentiate granular corneal dystrophy from corneas without known pathologies with high accuracy. This model has the potential for further enhancement, which could enable it to differentiate between various types of corneal dystrophies, aiding physicians in the diagnosis of corneal diseases.

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

 

Figure 1. Precision-recall curve and Precision-recall by threshold for the trained model using Vertex-AI.

 

Figure 2. Confusion matrix heatmap of the trained model.