Detecting Keratoconus using Machine Learning Models

Abstract

One of the major progressing sectors due to the introduction of technology has been healthcare. Diagnosis of patients has improved by manifolds. Keratoconus is a rare disease where it affects the patient’s cornea. There are ongoing researches around the world to find a solution that is accessible and practical. Our objective is to detect whether a person is suffering from Keratoconus or not. This huge volume of important data cannot be handled manually, hence use of concepts like machine learning, data analysis, data mining, etc. play an important role. To evaluate accuracy of Machine learning models like Inception V3, VGG16, MobileNet V2, ResNet 50 using color coded corneal maps. The authors have implemented these models and compared their performances amongst each other and thus select the best fit model. The training set contains of 1050 images and comprising of 1051 Normal eyes and 862 Suspect eyes. The models were implemented in python language on the Google Colab Platform. These models are providing a range of 75-95% accuracies depending on the different models. The highest accuracy was obtained by Inception V3 which was 95%. The dataset were corneal maps recorded using Scheimpflug imaging system. Based on the classification of the parameters of the corneal maps, the input data was sorted on the basis of severity and also predicting how likely the patient is to suffer from keratoconus