Kangjoo Lee

A Data-Driven Sparse GLM for fMRI Analysis Using Sparse Dictionary Learning With MDL Criterion

We propose a novel statistical analysis method for functional magnetic resonance imaging (fMRI) to overcome the drawbacks of conventional data-driven methods such as the independent component analysis (ICA). Although ICA has been broadly applied to fMRI due to its capacity to separate spatially or temporally independent components, the assumption of independence has been challenged by recent studies showing that ICA does not guarantee independence of simultaneously occurring distinct activity patterns in the brain. Instead, sparsity of the signal has been shown to be more promising. This coincides with biological findings such as sparse coding in V1 simple cells, electrophysiological experiment results in the human medial temporal lobe, etc. The main contribution of this paper is, therefore, a new data driven fMRI analysis that is derived solely based upon the sparsity of the signals. A compressed sensing based data-driven sparse generalized linear model is proposed that enables estimation of spatially adaptive design matrix as well as sparse signal components that represent synchronous, functionally organized and integrated neural hemodynamics. Furthermore, a minimum description length (MDL)-based model order selection rule is shown to be essential in selecting unknown sparsity level for sparse dictionary learning. Using simulation and real fMRI experiments, we show that the proposed method can adapt individual variation better compared to the conventional ICA methods.

Quantification of CMRO2 without hypercapnia using simultaneous near-infrared spectroscopy and fMRI measurements

Estimation of the cerebral metabolic rate of oxygen (CMRO(2)) and cerebral blood flow (CBF) is important to investigate the neurovascular coupling and physiological components in blood oxygenation level-dependent (BOLD) signals quantitatively. Although there are methods that can determine CMRO(2) changes using functional MRI (fMRI) or near-infrared spectroscopy (NIRS), current approaches require a separate hypercapnia calibration process and have the potential to incur bias in many assumed model parameters. In this paper, a novel method to estimate CMRO(2) without hypercapnia is described using simultaneous measurements of NIRS and fMRI. Specifically, an optimization framework is proposed that minimizes the differences between the two forms of the relative CMRO(2)-CBF coupling ratio from BOLD and NIRS biophysical models, from which hypercapnia calibration and model parameters are readily estimated. Based on the new methods, we found that group average CBF, CMRO(2), cerebral blood volume (CBV), and BOLD changes within activation of the primary motor cortex during a finger tapping task increased by 39.5 +/- 21.4%, 18.4 +/- 8.7%, 12.9 +/- 6.7%, and 0.5 +/- 0.2%, respectively. The group average estimated flow-metabolism coupling ratio was 2.38 +/- 0.65 and the hypercapnia parameter was 7.7 +/- 1.7%. These values are within the range of values reported from other literatures. Furthermore, the activation maps from CBF and CMRO(2) were well localized on the primary motor cortex, which is the main target region of the finger tapping task.

Detection of abnormal resting-state networks in individual patients suffering from focal epilepsy: an initial step toward individual connectivity assessment

The spatial coherence of spontaneous slow fluctuations in the blood-oxygen-level dependent (BOLD) signal at rest is routinely used to characterize the underlying resting-state networks (RSNs). Studies have demonstrated that these patterns are organized in space and highly reproducible from subject to subject. Moreover, RSNs reorganizations have been suggested in pathological conditions. Comparisons of RSNs organization have been performed between groups of subjects but have rarely been applied at the individual level, a step required for clinical application. Defining the notion of modularity as the organization of brain activity in stable networks, we propose Detection of Abnormal Networks in Individuals (DANI) to identify modularity changes at the individual level. The stability of each RSN was estimated using a spatial clustering method: Bootstrap Analysis of Stable Clusters (BASC) (Bellec et al., 2010). Our contributions consisted in (i) providing functional maps of the most stable cores of each networks and (ii) in detecting “abnormal” individual changes in networks organization when compared to a population of healthy controls. DANI was first evaluated using realistic simulated data, showing that focussing on a conservative core size (50% most stable regions) improved the sensitivity to detect modularity changes. DANI was then applied to resting state fMRI data of six patients with focal epilepsy who underwent multimodal assessment using simultaneous EEG/fMRI acquisition followed by surgery. Only patient with a seizure free outcome were selected and the resected area was identified using a post-operative MRI. DANI automatically detected abnormal changes in 5 out of 6 patients, with excellent sensitivity, showing for each of them at least one “abnormal” lateralized network closely related to the epileptic focus. For each patient, we also detected some distant networks as abnormal, suggesting some remote reorganization in the epileptic brain.

Sparse SPM: Group Sparse-dictionary learning in SPM framework for resting-state functional connectivity MRI analysis.

Recent studies of functional connectivity MR imaging have revealed that the default-mode network activity is disrupted in diseases such as Alzheimers disease (AD). However, there is not yet a consensus on the preferred method for resting-state analysis. Because the brain is reported to have complex interconnected networks according to graph theoretical analysis, the independency assumption, as in the popular independent component analysis (ICA) approach, often does not hold. Here, rather than using the independency assumption, we present a new statistical parameter mapping (SPM)-type analysis method based on a sparse graph model where temporal dynamics at each voxel position are described as a sparse combination of global brain dynamics. In particular, a new concept of a spatially adaptive design matrix has been proposed to represent local connectivity that shares the same temporal dynamics. If we further assume that local network structures within a group are similar, the estimation problem of global and local dynamics can be solved using sparse dictionary learning for the concatenated temporal data across subjects. Moreover, under the homoscedasticity variance assumption across subjects and groups that is often used in SPM analysis, the aforementioned individual and group analyses using sparse dictionary learning can be accurately modeled by a mixed-effect model, which also facilitates a standard SPM-type group-level inference using summary statistics. Using an extensive resting fMRI data set obtained from normal, mild cognitive impairment (MCI), and Alzheimers disease patient groups, we demonstrated that the changes in the default mode network extracted by the proposed method are more closely correlated with the progression of Alzheimers disease.

SPARK: Sparsity-based analysis of reliable k-hubness and overlapping network structure in brain functional connectivity.

Functional hubs are defined as the specific brain regions with dense connections to other regions in a functional brain network. Among them, connector hubs are of great interests, as they are assumed to promote global and hierarchical communications between functionally specialized networks. Damage to connector hubs may have a more crucial effect on the system than does damage to other hubs. Hubs in graph theory are often identified from a correlation matrix, and classified as connector hubs when the hubs are more connected to regions in other networks than within the networks to which they belong. However, the identification of hubs from functional data is more complex than that from structural data, notably because of the inherent problem of multicollinearity between temporal dynamics within a functional network. In this context, we developed and validated a method to reliably identify connectors and corresponding overlapping network structure from resting-state fMRI. This new method is actually handling the multicollinearity issue, since it does not rely on counting the number of connections from a thresholded correlation matrix. The novelty of the proposed method is that besides counting the number of networks involved in each voxel, it allows us to identify which networks are actually involved in each voxel, using a data-driven sparse general linear model in order to identify brain regions involved in more than one network. Moreover, we added a bootstrap resampling strategy to assess statistically the reproducibility of our results at the single subject level. The unified framework is called SPARK, i.e. SParsity-based Analysis of Reliable k-hubness, where k-hubness denotes the number of networks overlapping in each voxel. The accuracy and robustness of SPARK were evaluated using two dimensional box simulations and realistic simulations that examined detection of artificial hubs generated on real data. Then, test/retest reliability of the method was assessed using the 1000 Functional Connectome Project database, which includes data obtained from 25 healthy subjects at three different occasions with long and short intervals between sessions. We demonstrated that SPARK provides an accurate and reliable estimation of k-hubness, suggesting a promising tool for understanding hub organization in resting-state fMRI.

Statistical parametric mapping of FMRI data using sparse dictionary learning

Statistical parametric mapping (SPM) of functional magnetic resonance imaging (fMRI) uses a canonical hemodynamic response function (HRF) to construct the design matrix within the general linear model (GLM) framework. Recently, there has been many research on data-driven method on fMRI data, such as the independence component analysis (ICA). The main weakness of ICA for fMRI is its restrictive assumption, especially independence. Furthermore, recent study demonstrated that sparsity is more important than independency in ICA analysis for fMRI. Hence, we propose sparse learning algorithm, such as K-SVD, as an alternative, that decomposes the dictionary-atoms using sparsity rather than independence of the components. For the fMRI finger tapping task data, we employed the K-SVD algorithm to extract the time-course signal atoms of brain activation. The activation maps using trained dictionary as a design matrix showed tightly localized signals in a small set of brain areas.

Sparse dictionary learning for resting-state fMRI analysis

Recently, there has been increased interest in the usage of neuroimaging techniques to investigate what happens in the brain at rest. Functional imaging studies have revealed that the default-mode network activity is disrupted in Alzheimers disease (AD). However, there is no consensus, as yet, on the choice of analysis method for the application of resting-state analysis for disease classification. This paper proposes a novel compressed sensing based resting-state fMRI analysis tool called Sparse-SPM. As the brains functional systems has shown to have features of complex networks according to graph theoretical analysis, we apply a graph model to represent a sparse combination of information flows in complex network perspectives. In particular, a new concept of spatially adaptive design matrix has been proposed by implementing sparse dictionary learning based on sparsity. The proposed approach shows better performance compared to other conventional methods, such as independent component analysis (ICA) and seed-based approach, in classifying the AD patients from normal using resting-state analysis.

Disruption, emergence and lateralization of brain network hubs in mesial temporal lobe epilepsy

Hubs of brain networks are brain regions exhibiting denser connections than others, promoting long-range communication. Studies suggested the reorganization of hubs in epilepsy. The patterns of connector hub abnormalities specific to mesial temporal lobe epilepsy (mTLE) are unclear. We wish to quantify connector hub abnormalities in mTLE and identify epilepsy-related resting state networks involving abnormal connector hubs. A recently developed sparsity-based analysis of reliable k-hubness (SPARK) allowed us to address this question by using resting state functional MRI in 20 mTLE patients and 17 healthy controls. Handling the multicollinearity of functional networks, SPARK measures a new metric of hubness by counting the number (k) of networks involved in each voxel, and identifies which networks are actually associated to each connector hub. This measure provides new information about the network architecture involving connector hubs and a realistic range of k-hubness. We quantified the disruption and emergence of connector hubs in individual epileptic subjects and assessed the lateralization of networks involving connector hubs. In mTLE, we found pathological disruptions of normal connector hubs in the mTL and within the default mode network. Right mTLE had remarkably higher emergence of new connector hubs in the mTL than left mTLE. Different patterns of lateralization of the salience network involving the abnormal hippocampus were found in right versus left mTLE. The temporal, cerebellar, default mode, subcortical and motor networks also contributed to the lateralization of hippocampal networks. We finally observed an asymmetrical connector hub reorganization and overall regularization of epilepsy-related resting state networks in mTLE, characterized by the disruption of distant connections and the emergence of local connections.

A data-driven spatially adaptive sparse generalized linear model for functional MRI analysis

A novel data-driven sparse generalized linear model (GLM) and statistical analysis method for fMRI is developed. Although independent component analysis (ICA) has been broadly applied to fMRI to separate spatially or temporally independent components, recent studies show that ICA does not guarantee independence of simultaneously occurred distinct activity patterns in the brain and sparsity of the signal has been shown to be more important. Motivated from the ICA and biological findings such as sparse coding in the primary visual cortex simple cells, we propose a compressed sensing based data-driven sparse GLM solely based upon the sparsity of the signal. It enables estimation of spatially adaptive design matrix from sparse signal components that represent synchronous neural hemodynamics. Furthermore, an MDL based model order selection rule can determine unknown sparsity for sparse dictionary learning.