Extreme Learning Machines for Feature Selection and Classification of Cocaine Dependent Patients on Structural MRI Data

Abstract

In this paper, we present a Computer Aided Diagnosis and image biomarker identification system for cocaine dependence, which selects relevant regions from a set of brain structural magnetic resonance images (sMRI). After sMRI volume preprocessing for spatial normalization, we compute Pearson's correlation between pixel values across volumes and the indicative variable, obtaining a volume of correlation values (VCV). We calculate the gradient of the VCV which is used to perform a watershed segmentation of the brain volume into regions. A region selection stage finds the most relevant watershed regions. We propose two different approaches to characterize region relevance: (a) a wrapper procedure using extreme learning machines (ELM), and (b) apply correlation distribution percentiles to select most discriminant regions. Next, we consider three different procedures to extract the image features corresponding to selected regions: (1) collecting the sMRI intensity values of all the voxels that compose each region, compute (2) the mean or (3) the median of the sMRI intensity value of the voxels contained in each selected region. Extracted feature vectors are used to build a classifier aiming to discriminate between cocaine dependent patients and healthy controls. We compare results of several classifiers: ELM, OP-ELM, SVM and 1NN. Also, we visualize the brain locations of selected regions, checking if these locations are in accordance with previous findings in the medical literature.