A tiny glimpse into the human brain using model-free analysis for resting-state fMRI data

Abstract: Resting-state functional Magnetic Resonance Imaging (fMRI) acquires four dimensional data that indirectly depicts human brain activity. Within these four dimensional datasets reside resting-state functional connectivity networks (RFNs), depicting how the human brain is organized functionally. This series of studies delve into the use of data-driven analysis methods for resting-state fMRI data. Their strengths were explored and their weaknesses tackled, both in their methodologies and applications, all in hope to gain a better understanding of the data, and thereby how the brain function. The journey begins through the usage of one of the most common data-driven analysis methods in use today: Independent Component Analysis (ICA). ICA requires no user input parameter apart from the input dataset and the number of output Independent Components (NIC). The requirement of the NIC, a priori, is troubling as the inherent number of Independent Components (ICs) that exists within non-simulated datasets is unknown, due to the existence of various noise and artefact sources to differing degrees. Furthermore, comparing datasets using ICA is problematic because of the inherently different dimensionality of different datasets. To investigate the effects of NIC on the ICA output results, a classification framework based on Support Vector Machines (SVM) was implemented to automatically classify ICs as either potential RFNs, or noise/artefact signal. This feature-optimized classification of ICs with SVM, or FOCIS, framework uses features derived from verbal instructions for manual visual inspection of ICs. With only few significant features selected through iterative feature-selection and a small training set, the classification framework performed well with over 98% in overall accuracy for group ICA output results. Analysis of different resting-state fMRI datasets using FOCIS indicated that the specification of NIC can critically affect the ICA results on restingstate fMRI data. These changes are complex and are individually different from one another, irrespective whether the IC is a potential RFN or artefact/noise signals. Applying this knowledge on group comparison studies, ICA was used to study migraine patients undergo kinetic oscillation stimulation treatment. The immediate effects of the treatment allows direct correlation of a patient’s pain levels with changes in their RFNs. Differences in RFNs that include areas in the midbrain and limbic system regulating the central nervous system were discovered in migraine patients compared to healthy control group. Overlapping areas were also shown to be affected by the treatment. These results provide supporting evidence for the hypothesis that the treatment affects and regulates the parasympathetic autonomic reflex, alleviating migraine symptoms. Hierarchical clustering is another data-driven analysis method that is almost devoid of all userinput parameters. The algorithm naturally stratifies data into a hierarchical structure. It is believed that brain function is hierarchically organized, so an algorithm which reflects this aspect is a seemingly excellent choice to use for analyzing the resting-state fMRI data. A hierarchical clustering analysis framework was developed to extract RFNs from resting-state fMRI data with full brain coverage at voxel level. The RFNs identified using hierarchical clustering conforms to those identified previously using other data processing techniques, such as ICA. An innate ability of the clustering algorithm is to naturally organize data into a hierarchical tree (dendrogram). This was fully utilized though extensions in the framework for cluster evaluation. Extending the hierarchical clustering framework with the cluster evaluation pipeline allowed extraction of functional subdivisions of known RFNs. This demonstrated that not only can hierarchical clustering be used to extract the modular organization at the scale of large systems for entire RFNs, but can also be used to derive the functional subdivision of RFNs and provide a consistent method of analysis at different levels of detail. The subnetworks extracted using hierarchical clustering reveals the intrinsic functional connectivity amongst the subnetworks within RFNs and provide clues for further exploring the potential for currently unknown functional junctions within RFNs.