Search

Understanding the origins of intrinsic time-varying functional connectivity remains a challenge in the neuroimaging community. However, some associations between dynamic functional connectivity (dFC) and behavioral traits have been observed along with gender differences. We propose a permutation testing framework to investigate dynamic differences between groups of subjects. In particular, we investigate differences in fractional occupancy, state persistency and the full transition probability matrix. We demonstrate our framework on resting state functional magnetic resonance imaging data from 820 healthy young adults from the Human Connectome Project considering two prominent dFC models, namely sliding-window k-means and the Gaussian hidden Markov model. The variables showing consistent significant dynamic differences were limited to gender and the degree of motion in the scanner. We observe for the data considered that a large sample size (here 500 subjects) is needed to to draw reliable conclusions about the significance of those variables. Our results point to dynamic features providing limited information with regard to behavioral traits despite a relatively large sample size.

Dynamic functional connectivity has become a prominent approach for tracking the changes of macroscale statistical dependencies between regions in the brain. Effective parametrization of these statistical dependencies, referred to as brain states, is however still an open problem. We investigate different emission models in the hidden Markov model framework, each representing certain assumptions about dynamic changes in the brain. We evaluate each model by how well they can discriminate between schizophrenic patients and healthy controls based on a group independent component analysis of resting-state functional magnetic resonance imaging data. We find that simple emission models without full covariance matrices can achieve similar classification results as the models with more parameters. This raises questions about the predictability of dynamic functional connectivity in comparison to simpler dynamic features when used as biomarkers. However, we must stress that there is a distinction between characterization and classification, which has to be investigated further.

Modern functional and diffusion magnetic resonance imaging (fMRI and dMRI) provide data from which macro-scale networks of functional and structural whole brain connectivity can be estimated. Although networks derived from these two modalities describe different properties of the human brain, they emerge from the same underlying brain organization, and functional communication is presumably mediated by structural connections. In this paper, we assess the structure-function relationship by evaluating how well functional connectivity can be predicted from structural graphs. Using high-resolution whole brain networks generated with varying density, we contrast the performance of several non-parametric link predictors that measure structural communication flow. While functional connectivity is not well predicted directly by structural connections, we show that superior predictions can be achieved by taking indirect structural pathways into account. In particular, we find that the length of the shortest structural path between brain regions is a good predictor of functional connectivity in sparse networks (density less than one percent), and that this improvement comes from integrating indirect pathways comprising up to three steps. Our results support the existence of important indirect relationships between structure and function, extending beyond the immediate direct structural connections that are typically investigated.

Many data-driven approaches exist to extract neural representations of functional magnetic resonance imaging (fMRI) data, but most of them lack a proper probabilistic formulation. We propose a group level scalable probabilistic sparse factor analysis (psFA) allowing spatially sparse maps, component pruning using automatic relevance determination (ARD) and subject specific heteroscedastic spatial noise modeling. For task-based and resting state fMRI, we show that the sparsity constraint gives rise to components similar to those obtained by group independent component analysis. The noise modeling shows that noise is reduced in areas typically associated with activation by the experimental design. The psFA model identifies sparse components and the probabilistic setting provides a natural way to handle parameter uncertainties. The variational Bayesian framework easily extends to more complex noise models than the presently considered., Comment: 10 pages plus 5 pages appendix, Submitted to ICASSP 17

Dynamic functional connectivity (dFC) has recently become a popular way of tracking the temporal evolution of the brains functional integration. However, there does not seem to be a consensus on how to choose the complexity, i.e. number of brain states, and the time-scale of the dynamics, i.e. the window length. In this work we use the Wishart Mixture Model (WMM) as a probabilistic model for dFC based on variational inference. The framework admits arbitrary window lengths and number of dynamic components and includes the static one-component model as a special case. We exploit that the WMM framework provides model selection by quantifying models generalization to new data. We use this to quantify the number of states within a prespecified window length. We further propose a heuristic procedure for choosing the window length based on contrasting for each window length the predictive performance of dFC models to their static counterparts and choosing the window length having largest difference as most favorable for characterizing dFC. On synthetic data we find that generalizability is influenced by window length and signal-tonoise ratio. Too long windows cause dynamic states to be mixed together whereas short windows are more unstable and influenced by noise and we find that our heuristic correctly identifies an adequate level of complexity. On single subject resting state fMRI data we find that dynamic models generally outperform static models and using the proposed heuristic points to a windowlength of around 30 seconds provides largest difference between the predictive likelihood of static and dynamic FC.

Functional Magnetic Resonance Imaging (fMRI) is widely used to gain a better understanding of the human brain's functional organization. As fMRI data are high dimensional it is challenging to analyse using conventional methods thus low-rank approximations such as principal component analysis (PCA), and independent component analysis (ICA) is often applied as a preprocessing step before any additional analysis. Low-rank methods generally require that the rank or latent dimensionality is known beforehand. When this is not the case a range of plausible dimensionalities have to be tested and compared. Furthermore, in an fMRI-context it is not fully understood how information from multiple subjects should best be incorporated when applying dimensionality reduction. We propose a Bayesian group principal component analysis (Group-BPCA) model with an automatic relevance determination (ARD) prior to determine the number of active components supported by the data. All subjects share the same spatial maps (components), but the uncertainties on these maps as well as the noise is subject specific. We find an approximate solution using the mature variational Bayesian framework and develop a fast and scalable implementation using a graphical processing unit (GPU). We test the model on fMRI data from 29 healthy subjects performing a block-design fingertapping experiment. The model identified 10 active components. Neither variational Bayesian PCA on temporally concatenated data nor Group-BPCA, where uncertainties on the spatial maps are shared, leads to pruning components, but provide better generalization in two of three scenarios. We show that the right level of subject variability is highly dependent on the chosen validation scheme.

Independent component analysis (ICA) is a widely used blind source separation method for decomposing resting state functional magnetic resonance imaging (rs-fMRI) data into latent components. However, it can be challenging to obtain subject-specific component representations in multi-subject studies. Independent vector analysis (IVA) is a promising alternative approach to perform group fMRI analysis, which has been shown to better capture components with high inter-subject variability. The most widely applied IVA method is based on the multivariate Laplace distribution (IVA-GL), which assumes independence within subject components coupled across subjects only through shared scaling. In this study, we propose a more natural formulation of IVA based on a Normal-Inverse-Gamma distribution (IVA-NIG), in which the components can be directly interpreted as realizations of a common mean component with individual subject variability. We evaluate the performance of IVA-NIG compared to IVA-GL and similar decomposition methods, through the application of two types of simulated data and on real task fMRI data. The results show that IVA-NIG offers superior detection of components in simulated fMRI data. On real fMRI data with low inter-subject variability we find that all methods identify similar and plausible components.

Functional Magnetic Resonance Imaging has become a central measuring modality to quantify functional activiation of the brain in both task and rest. Most analysis used to quantify functional activation requires supervised approaches as employed in statistical parametric mapping (SPM) to extract maps of task induced functional activations. This requires strong knowledge and assumptions on the BOLD response as a function of activitation while smoothing in general enhances the statistical power but at the cost of spatial resolution. We propose a fully unsupervised approach for the extraction of task activated functional units in multi-subject fMRI data that exploits that regions of task activation are consistent across subjects and can be more reliably inferred than regions that are not activated. We develop a non-parametric Gaussian mixture model that apriori assumes activations are smooth using a Gaussian Process prior while assuming the segmented functional maps are the same across subjects but having individual time-courses and noise variances. To improve inference we propose an enhanced split-merge procedure. We find that our approach well extracts the induced activity of a finger tapping fMRI paradigm with maps that well corresponds to a supervised group SPM analysis. We further find interesting regions that are not activated time locked to the paradigm. Demonstrating that we in a fully unsupervised manner are able to extract the task-induced activations forms a promising framework for the analysis of task fMRI and resting-state data in general where strong knowledge of how the task induces a BOLD response is missing.

We evaluate the infinite relational model (IRM) against two simpler alternative nonparametric Bayesian models for identifying structures in multi subject brain networks. The models are evaluated for their ability to predict new data and infer reproducible structures. Prediction and reproducibility are measured within the data driven NPAIRS split-half framework. Using synthetic data drawn from each of the generative models we show that the IRM model outperforms the two competing models when data contain relational structure. For data drawn from the other two simpler models the IRM does not overfit and obtains comparable reproducibility and predictability. For resting state functional magnetic resonance imaging data from 30 healthy controls the IRM model is also superior to the two simpler alternatives, suggesting that brain networks indeed exhibit universal complex relational structure in the population.

The shift invariant multi-linear model based on the CandeComp/PARAFAC (CP) model denoted ShiftCP has proven useful for the modeling of latency changes in trial based neuroimaging data [17]. In order to facilitate component interpretation we presently extend the shiftCP model such that the extracted components can be constrained to pertain to predefined frequency ranges such as alpha, beta and gamma activity. To infer the number of components in the model we propose to apply automatic relevance determination by imposing priors that define the range of variation of each component of the shiftCP model and learning the hyper-parameters of these priors during model estimation.

To overcome poor signal-to-noise ratios in neuroimaging, data sets are often acquired over repeated trials that form a three-way array of space×time×trials. As neuroimaging data contain multiple inter-mixed signal components blind signal separation and decomposition methods are frequently invoked for exploratory analysis and as a preprocessing step for signal detection. Most previous component analyses have avoided working directly with the tri-linear structure, but resorted to bi-linear models such as ICA, PCA, and NMF. Multi-linear decomposition can exploit consistency over trials and contrary to bi-linear decomposition render unique representations without additional constraints. However, they can degenerate if data does not comply with the given multi-linear structure, e.g., due to time-delays. Here we extend multi-linear decomposition to account for general temporal modeling within a convolutional representation. We demonstrate how this alleviates degeneracy and helps to extract physiologically plausible components. The resulting convolutive multi-linear decomposition can model realistic trial variability as demonstrated in EEG and fMRI data.

Non-negative matrix factorization (NMF), i.e. V ap WH where both V, W and H are non-negative has become a widely used blind source separation technique due to its part based representation. The NMF decomposition is not in general unique and a part based representation not guaranteed. However, imposing sparseness both improves the uniqueness of the decomposition and favors part based representation. Sparseness in the form of attaining as many zero elements in the solution as possible is appealing from a conceptional point of view and corresponds to minimizing reconstruction error with an L0 norm constraint. In general, solving for a given L0 norm is an NP hard problem thus convex relaxation to regularization by the L1 norm is often considered, i.e., minimizing (1/2||V - WH||F 2 + lambda||H||1).An open problem is to control the degree of sparsity lambda imposed. We here demonstrate that a full regularization path for the L1 norm regularized least squares NMF for fixed W can be calculated at the cost of an ordinary least squares solution based on a modification of the least angle regression and selection (LARS) algorithm forming a non-negativity constrained LARS (NLARS). With the full regularization path, the L1 regularization strength lambda that best approximates a given L0 can be directly accessed and in effect used to control the sparsity of H. The MATLAB code for the NLARS algorithm is available for download.

Non-negative matrix factorization (NMF) has become a widely used blind source separation technique due to its part based representation and ease of interpretability. We currently extend the NMF model to allow for delays between sources and sensors. This is a natural extension for spectrometry data where a shift in onset of frequency profile can be induced by the Doppler effect. However, the model is also relevant for biomedical data analysis where the sources are given by compound intensities over time and the onset of the profiles have different delays to the sensors. A simple algorithm based on multiplicative updates is derived and it is demonstrated how the algorithm correctly identifies the components of a synthetic data set. Matlab implementation of the algorithm and a demonstration data set is available.

Non-linear hemodynamic models express the BOLD signal as a nonlinear, parametric functional of the temporal sequence of local neural activity. Several models have been proposed for this neural activity. We identify one such parametric model by estimating the distribution of its parameters. These distributions are themselves stochastic, therefore we estimate their variance by epoch based leave-one-out cross validation, using a Metropolis-Hastings algorithm for sampling of the posterior parameter distribution.