Classification method evaluation in tangent space for statistical shape analysis

Abstract

The purpose of this study was to evaluate the performance of several classifiers from the literature on data in the tangent space at the statistical shape analysis context . Additionally, simulations were conducted considering three scenarios: (1) data without the principal component analysis (PCA) application ; and (2) data with the PCA application, where the components explained variations from 70% to 75% and from 90% to 95%. Simulation results showed the performance of the classifiers decreases when there is low concentration data, with significant accuracy gains observed when applying PCA in most of the scenarios examined. The next step was to perform classification using four datasets of real data, considering the same scenarios as in the simulation study. In these applications, the best results were observed in databases where the average shapes were significantly different between the groups. Conversely, the worst performances were observed in data related to magnetic resonance imaging of schizophrenic patients, with a maximum accuracy of 85.7%.