Dictionary learning algorithms for functional magnetic resonance imaging

Abstract

The detection of regional activity and estimation of communication networks are two critical features to enable functional understanding of a human brain. In this regard, functional magnetic resonance imaging (fMRI) has emerged as a powerful noninvasive neuroimaging modality to investigate not only the neural activity within the different isolated brains regions but also how different brain regions communicate with each other during cognition and resting state. Similar to other imaging modalities, the functional connectivity networks for fMRI data are usually determined by following a data-driven approach, which typically consists of statistical similarity measures such as estimation of correlation matrix, and machine learning methods such as matrix decomposition and exploratory techniques. Unlike hypothesis-driven approach for activation detection that requires prior knowledge about the shape of hemodynamic response function (HRF), the main advantage of data-driven approach is its independence from predetermined paradigm knowledge, hence, its applicability to both activation detection and functional connectivity estimation. However, the trained bases from the standard decomposition techniques and data transformation bases from exploratory techniques are not very meaningful and can give rise to issues such as computational inefficiency and inability to retrieve spatiotemporal ground truth with sufficient conviction, respectively. To remedy these problems, matrix decomposition is adapted to fMRI data by considering sparse assumption based on the spatial characteristics and autocorrelation assumption based on temporal characteristics of the fMRI data to develop novel sparse dictionary learning algorithms for this work. In proposed algorithms, better assumptions about underlying structure of the data are taken into account, which yield bases that are relatively closer to the artificially generated ground truth. For every proposed algorithm, an empirical study is introduced for quantitative and comparative analysis with existing data-driven techniques. To validate the proposed algorithms, they are applied to six different fMRI data-sets and it has been shown that they outperform existing state state-of-the-art-methods in terms of statistical performance of map and network retrieval for activation detection and functional connectivity, respectively.