Detecting Keratoconus from Corneal Imaging Data Using Machine Learning

Abstract

The prevalence of keratoconus is estimated to be 1 in 1,500 people. It's often linked to a decline in eyesight. Given its widespread occurrence, there is a pressing need for the creation of new instruments that can detect the illness at an early stage, therefore halting its course and preventing visual loss. The purpose of this research is to create and evaluate a machine learning method for early detection of keratoconus. For the purpose of detecting keratoconus, we implemented a number of machine learning algorithms and put them to the test on real-world medical data, such as corneal topography, elevation, and pachymetry parameters obtained from OCT-based topography machines at a number of corneal clinics in Japan. We used Matlab to create 25 unique machine learning models, with results ranging from 58% to 92.0% accuracy. Using a subset of eight corneal characteristics with the strongest discriminatory power, a support vector machine (SVM) algorithm achieved an accuracy of 94%. Especially in the preclinical and early phases of the illness, it may be difficult for doctors to monitor corneal condition and discover keratoconus using subjective assessments alone, but the suggested model may help. The algorithm may be utilized as either an add-on to existing corneal imaging equipment or as a standalone piece of software for diagnosing keratoconus at an early stage.