Detecting Abnormal Functional Communities (2014-2016)
Univariate analyses of fMRI activation patterns have allowed us to localize functional differences induced by neurological disorders. However, there is increasing evidence that complex pathologies reflect distributed impairments across multiple brain systems. Unfortunately, network analyses in clinical neuroimaging studies are largely limited to aggregate measures (e.g. centrality and degree distribution), which do not pinpoint a concrete etiological mechanism. Therefore, we propose a hierarchical Bayesian framework to detect both hyper- and hypo-synchronous functional communities within whole-brain task fMRI data.
Our approach deviates from prior literature in three crucial ways. First, we propose a unified generative model that describes the relationship between population templates and individual subject observations. Second, we perform community detection in the space of group-level functional differences, which allows us to identify both enhanced and impaired processes. Finally, our framework can simultaneously detect multiple abnormal communities of varying type. We have applied the model to an fMRI study of social perception in autism, and we are currently evaluating its efficacy in predicting the effects of early behavioral intervention in autism.
Identifying Foci of a Neurological Disorder (2011-2014)
Aberrations in functional connectivity inform us about disrupted neural communication, which may relate to certain brain pathologies. Although prior works have identified connections in the brain associated with a particular neurological disorder, such results are difficult to interpret and validate. This is because the bulk of our knowledge about the brain is organized around region properties (i.e., functional localization, tissue properties, morphometry) and not the connections between them. Moreover, it is nearly impossible to design non-invasive experiments that target and/or alter a particular connection between two brain regions.
We develop a unified framework that pinpoints region “foci”, whose connectivity patterns are most disrupted by the disorder. Using a probabilistic setting, we define a latent graph that characterizes the network of abnormal functional connectivity emanating from the affected foci. This generates population differences in the observed fMRI correlations. We present two versions of the model. The first variant considers the complete graph of pairwise functional connections. The second model uses neural anatomy as a substrate for modeling functional differences.
We have validated the models on a population study of schizophrenia, Alzheimer’s disease and Huntington’s disease. We have also adapted this framework for subject-specific analysis in order to identify the origins of epileptic seizures in the brain, and we have extended the model to extract multiple networks via a clustering prior. This modification has allowed us to disentangle social and language deficits in autism.
Joint Modeling of fMRI and DWI Data (2009-2011)
The interaction between functional and anatomical imaging modalities provides a rich framework for understanding the brain. In particular, correlations present in resting-state fMRI data are believed to reflect the intrinsic functional organization between regions. Similarly, tractography based on DWI data is used to estimate underlying white matter fiber tracts (anatomical pathways). However, relatively little progress has been made in fusing the complementary information provided by these modalities.
In this work, we propose a novel probabilistic framework to infer the relationship between fMRI and DWI data. The model is based on latent connectivity templates and makes intuitive assumptions about the data generation process. Additionally, we formulate a natural extension of the model to population studies.
Extracting Robust fMRI Measures for Clinical Diagnosis (2009-2010)
Pinpointing robust differences between a control and clinical population is crucial for understanding the effects of a disease. This task is especially difficult for resting-state fMRI data due to the considerable amount of noise/inter-subject variability and to the relatively small number of subjects found in clinical experiments. Univariate tests and random effects analysis are the standard techniques for population studies. However, this analysis ignores multivariate patterns within the data and is not robust.
Our approach is to apply feature selection/classification algorithms to identify predictive differences between the populations. We successfully employed the Gini Importance measure derived from Random Forests to identify discriminative functional connectivity anomalies induced by schizophrenia.
Data-Driven Functional Connectivity Analysis (2008-2009)
We investigate the use of data-driven clustering algorithms as complementary approaches to seed-based analysis. This allows us to partition the whole brain into an increasing number of clusters. We show that both K-Means and Spectral Clustering yield patterns that correspond to well-known functional systems.