Uso da transformada de Wavelet e técnicas de aprendizado de máquina para criação de modelo computacional de auxí­lio ao diagnóstico de ceratocone baseado em parâmetros biomecânicos da córnea

Abstract

This study aimed to create computer models to aid the diagnosis of keratoconus using an edge segmentation algorithm, wavelet transforms and machine-learning techniques based on biomechanical parameters of the cornea derived from Corvis ST. 102 normal eyes, and 73 eyes with keratoconus grade I and II for training and validation of the models were included. Initially the 31 original equipment parameters were studied to assess its power in separating the control groups and research group. Then, it was used the Canny algorithm for edges segmentation of the selected images, so that 400 central points of these images could be processed with the wavelet transforms and later with machine learning techniques. The best result was achieved using the Haar wavelet transform and a Multilayer Perceptron Neural Network, reaching the sensitivity of 84.93%, a specificity of 89.22% and an Area Under the ROC Curve (AUC) of 0932. It concludes that the developed models can help make the diagnosis of early keratoconus.