Detection of Keratoconus and Forme Fruste Keratoconus Through Supervised Machine Learning Algorithms

Keratoconus is an ocular disease that causes progressive corneal thinning. This irreversible thinning distorts the curvature of cornea, impairs visual acuity and sometime leads to loss of vision. Refractive surgery on keratoconic eyes results in complications rather than a cure for ophthalmic power. Therefore, it is crucial to identify keratoconus in its early stage. Moreover, identifying early signs of keratoconus clinically is difficult without expertise. Hence, the goal of this research is to detect keratoconus alongside ‘Forme fruste’ keratoconus considering ‘Forme fruste keratoconus’ as an eligible candidate for early detection of the disease. The first step towards the objective was carried out through comparing the keratoconus eyes with normal eyes with considerable 100% accuracy using supervised machine learning algorithms - Naive Bayes, k-Nearest Neighbour, Logistic Regression, Support Vector Machine, Decision Tree, and Random Forest. The remarkable results achieved in binary classification encouraged us to perform multiclass classification using the same supervised machine learning techniques with near to 100% of accuracy in segregating Keratoconus, forme fruste keratoconus and normal eye data. The real time corneal tomographies with corneal features were used herein that were obtained from the SIRIUS CSO keratometer. Both types of classification used 21 out of the 28 corneal features. Amongst all these supervised machine learning classifiers used herein, the Decision Tree and Random Forest were found to perform significantly well over others with significant accuracy.