Evaluation of keratoconus detection from elevation, topography and pachymetry raw data using machine learning

Abstract

Purpose : To assess the performance of machine learning algorithms in detecting keratoconus (KCN) from corneal parameters in a tomography dataset, such as elevation, topography, and pachymetry.

Methods : We developed numerous machine learning models to detect keratoconus from corneal parameters. Elevation, topography and pachymetry dataset were obtained from 5881 eyes of 2800 patients in Brazil using a high-resolution rotating Scheimpflug camera system for anterior segment analysis (Pentacam ® HR – Oculus, Optikgeräte GmbH). The accuracy of models was computed using each dataset of elevation, topography and pachymetry parameters separately. 10-fold cross validation of the area under the receiver operating characteristic curve (AUC) was used to evaluate the accuracy of different models. Thus, 3 independent datasets were created. Each of them was evaluated, and performance evaluation related to KCN detection was observed.

Results : A total of 1726 eyes were normal, and 4155 eyes were diagnosed as KCN. Figure 1 presents the distribution of the anterior cornea curvature radius (ACCR) of the cornea parameter versus the mean radius of cornea curvature (MRCC) in the 7 to 9 mm area parameter of normal versus KCN eyes. The cubic support vector machine (SVM) outperformed all other machine learning classifiers with an AUC of 1 for detecting KCN using elevation parameters only (Figure 2, ROC on left and confusion matrix on the right panel). The highest accuracy of classifiers for detecting KCN using pachymetry only and topography only parameters were 96.6% and 95.2%, respectively.

Conclusions : The results suggest that the cubic support vector machine (SVM) using elevation parameters provide the highest accuracy in detecting normal from KCN cases. This algorithm might be of help for detecting KCN patients in ophthalmological clinical sets.

This is a 2021 ARVO Annual Meeting abstract.

 

Figure 1. Scatter plot of ACCR parameter versus MRCC parameter of normal eyes and eyes with KCN.

 

Figure 2. ROC curve and the confusion matrix of the machine learning classifier applied on elevation parameters only.