Análise computacional da biomecânica corneal para diagnóstico de ceratocone

Abstract

keratoconus. The images were segmented for identification and conversion into vectors for representation of the anterior surface, apparent posterior surfaces, apparent pachymetry and composition of the previous data. The vectors were chained (batch images), simplified with Wavelet and submitted to MLP, k-NN, Logistic Regression, Naïve Bayes and Fast LargeMargin, in addition the vectors were rearranged as 2D histograms for neural network application with Deep Learning. The evaluation of the classifications was done with the score equal to the product of the sensitivity multiplied by the specificity, with confidence interval between 0.7843 and 1 and level of significance 0.0157. Exams of 686 normal eyes and 406 eyes with keratoconus in degrees from I to IV, from exam bases from Europe and Brazil, were used for training and validation of applied data. The best models identified were apparent pachymetry on batch images, with wavelet level 4 and processed with fast large margin in the European database, with a score of 0.8247, sensitivity of 89.5% and specificity of 92.14%; and 2D histogram of apparent pachymetry, with LeNET5, at the Brazilian database, with a score of 0.8361, sensitivity of 88.58% and specificity of 94.39%. It is concluded that biomechanical models can be used to diagnose keratoconus.