A Comparison of Representation Learning Methods for Dimensionality Reduction of fMRI Scans for Classification of ADHD

Abstract

This paper compares three feature representation techniques used to represent resting state functional magnetic resonance (fMRI) scans. The proposed models of feature representation consider the time averaged fMRI scans as raw representation of image data. The effectiveness of the representation is evaluated by using these features for classification of Attention Deficit Hyperactivity Disorder (ADHD) patients from healthy controls. The dimensionality reduction methods used for feature representation are maximum-variance unfolding, locally linear embedding and auto-encoders. The classifiers tested for classification purpose were neural net and support vector machine. Using auto-encoders with four hidden layers along with a support vector machine classifier yielded a classification accuracy of 61.25% along with 65.69% sensitivity and 52.20% specificity.

Tasks

Reproductions