Voxel importance in classifier ensembles based on sign consistency patterns: application to sMRI

摘要

This paper investigates a new measure of voxel importance based on analysing the sign consistency of voxels in an ensemble of linear SVM classifiers. The ensemble is endowed with a significant degree of diversity since the training set for each individual classifier is a random subsample of the initial training set. The importance of a voxel is proportional to the number of times that the voxel's weight has the same sign in all the classifiers of the ensemble. The multivariate nature of the method yields a robust importance pattern formed by clusters of voxels. The method is demonstrated with a MCI vs. control subjects classification task using the ADNI data.