Modelos computacionais baseados em aprendizagem de máquina para classificação do ceratocone por meio dos sinais biomecânicos fornecidos pelo Ocular Responder Analyser

Abstract

This work aims to develop computational models based on artificial intelligence to detect keratoconus, creating methods to predict the presence of keratoconus using machine learning. These models used corneal biomechanical data, from the signs of applanation and pressure the Ocular Response Analyzer exam (ORA). The goal is to build computational models to help the refractive surgeon decide between performing surgery or not in cases susceptible. The models were created using neural networks the types of MLP and RBF and tree decision and also Wavelet signal processing. The experiment was divided into two phases: the first to know and understand the signals identifying the most important features and regions, and the second using the wavelet signal processing to extract the relevant information signal. The various models developed in two phases were studied and compared. The best results were: for total accuracy rate of 93.03%, 93.95% rate of sensitivity and specificity rate of 99.26%. The conclusion is that the developed models can contribute to the refractive surgeon in his decision in the diagnosis of keratoconus grade I and II.