Diagnosis Of Keratoconus Using Optical Coherence Tomography And Machine Learning Algorithms

To classify with machine learning models into four categories: normal, keratoconus in its fruste form, incipient keratoconus, moderate-advanced keratoconus.

A public database of 3162 patients in total was used, of which 264 belonged to the healthy group, 2595 to the fruste keratoconus group, 221 to the incipient keratoconus group and 82 to the moderate-advanced keratoconus group

All 3162 patients underwent optical coherence tomography SS-1000 CASIA OCT Imaging Systems (Tomey, Japan). The database included all topographic, pachymetric, anterior chamber, aberrometric data from 1 to 8 millimeters in total 448 variables. To train and evaluate all the classification models, the 448 variables of the 3162 patients were divided into 80% for the training part and the remaining 20% for the test part.

Results: The classification machine learning models that exceeded 90% overall accuracy are presented. LGMB Classifier obtained 97% accuracy, XGB Classifier 96%, Logistic Regression 94%, Baggin Classifier 94%, SGD Classifier 93%, SVC 95%, Extra Trees Classifier 95%, Decision Tree Classifier 92%, Linear SVC 92%, Passive Aggressive Classifier 93%.

Conclusions: The main utility of these algorithms is the detection of keratoconus in its fruste form or incipient keratoconus in healthy patients and to be able to decide with greater certainty the possibility of performing refractive techniques of ablation of the corneal surface such as photorefractive keratectomy (PRK) or laser-assisted in situ keratomileusis (LASIK).