Evaluation of a Machine-Learning Classifier for Keratoconus Detection Based on Scheimpflug Tomography

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning

Background:

• Keratoconus (KC) is a progressive corneal disease requiring accurate detection.
• Early identification of forme fruste (FF) KC is crucial for timely intervention.
• Current KC classification methods have limitations in accuracy and objectivity.

Purpose of the Study:

• To assess the performance of a support vector machine (SVM) algorithm for automated corneal pattern identification.
• To objectively classify corneal patterns using 22 parameters from Pentacam measurements.
• To compare the SVM algorithm's classification accuracy against established keratoconus detection methods.

Main Methods:

• Utilized Pentacam data from 860 eyes across five groups: KC, FF, astigmatic, post-refractive surgery, and normal.
• Developed an SVM algorithm in Weka using 22 corneal parameters for classification.
• Performed cross-validation for three tasks: KC vs. normal, FF vs. normal, and all five groups.

Main Results:

• Achieved 98.9% accuracy for KC versus normal eyes, with 99.1% sensitivity and 98.5% specificity.
• Discriminated FF versus normal eyes with 93.1% accuracy, 79.1% sensitivity, and 97.9% specificity.
• Classified all five groups with 88.8% accuracy, a weighted average sensitivity of 89.0%, and 95.2% specificity.

Conclusions:

• The SVM algorithm demonstrates high accuracy in identifying keratoconus and forme fruste.
• Results are comparable or superior to existing single-parameter methods for corneal disease classification.
• The automated approach offers an objective and efficient tool for diagnosing corneal abnormalities.