A comparative study of α-Divergence based NMF techniques for FMRI analysis

The main objective of fMRI analysis methods is to detect the Blood Oxygenation Level Dependent (BOLD) from fMRI sequences. Algorithms which are discussed here are known as data-driven methods. The main advantage of these types of algorithms over data-based methods is that there is no need for prior information. Here, we focus on one of the powerful matrix factorization algorithms which has been recently applied to fMRI called Non-negative Matrix Factorization (NMF) [1]. There exist many different NMF techniques in the literature and no comprehensive assessment of their performances on fMRI data has been reported. So, in this work the performance in terms of BOLD detection, using a-Deivergence based methods are investigated and then compared with the commonly used Euclidean distance based method. The aim is to highlight the advantages that such techniques can have in practice. We explored the performance of these techniques for two types of real fMRI and also synthetic data. We observed that the a-Divergence based methods are also applicable to fMRI data and reveal acceptable performance.