LEARN STRUCTURAL AND RESTING STATE FUNCTIONAL CONNECTIVITY PATTERNS

LEARN STRUCTURAL AND RESTING STATE FUNCTIONAL CONNECTIVITY PATTERNS FROM TASK-BASED FMRI DATA Xi Jiang Computer Science Department The University of Georgia jiang@cs. uga. edu Introduction Resting state f. MRI (R-f. MRI) has been widely used for exploring functional networks of the human brain. Large-scale brain networks constructed from R-f. MRI data are informative about global properties of the human brain. However, the properties of functionally-specialized sub-networks, such as the working memory system, cannot be directly assessed from the large-scale networks. We propose to perform task-based f. MRI to identify functional networks, and then use them as reliable data to learn consistent structural and resting state functional connectivity patterns. Our experimental results show that brain sub-networks identified by task-based f. MRI have consistent structural and resting state functional connectivity patterns, indicating their potential roles as prior models to guide and constrain the subnetwork identification from large-scale networks in the absence of task-based f. MRI datasets. Fig. 2. Fig. 1. Flowchart • Background /Related Work The human brain is believed to be functionally segregated or specialized. For studying higher cognitive functions and neurological diseases, the identification of functional networks has gained increasing interest in recent years. In particular, Rf. MRI has been increasingly used for exploring functional networks of the human brain. Under the premise that low-frequency oscillations in R-f. MRI time courses between spatially distinct brain regions are suggested to reflect the functional architecture of the brain, large-scale brain networks constructed from R-f. MRI data are informative about global properties of the human brain. However, the properties of functionally-specialized sub-networks such as the working memory, attention and emotion sub-networks cannot be directly assessed from the large-scale networks. In the literature, data-driven algorithms have been widely used to identify the functional sub-networks from R-f. MRI data. However, these data-driven approaches might be sensitive to the parameters used, and the identification of consistent subnetworks across individuals is still an open problem. Moreover, whether the functionally-specialized sub-networks have consistent structural connectivity patterns has raised much interest. • (1) (2) The entropy is minimized using a gradient descent approach: (3) • The result is reproducible (Fig. 3 and Fig. 4). ROI quality measurement and optimization We define G as a criterion for ROI optimization and unreliable ROI removal: (4) Fig. 4. Weight variability distribution Fig. 6. (a) Structural connectivity matrices; Fig. 5. G values for 4 test subjects (b) resting state functional connectivity matrices Identification of consistent structural and resting state functional sub-networks has been a challenging problem due to the lack of prior models and the sensitivity to clustering parameters used. We proposed a novel experimental and computational paradigm to solve this problem. It can be used as prior models to guide and constrain the sub-network identification from large-scale networks in the absence of task-based f. MRI datasets. References (6) 1. Faraco, C. “Mapping the working memory network using the OSPAN task”, Neuro. Image 47(1): S 105, 2009. 2. Bernard Ng. “Discovering sparse functional brain networks using group replicator dynamics (GRD)”, IPMI 2009. (7) Acknowledgments (5) • Fig. 3. Reproducibility of network Discussion and Contributions The entropy is given by: Approach As summarized in Figure 1, our overall strategies include 4 major steps: Functional network identification via task-based f. MRI To identify the working memory network, each participant performed a modified version of the OSPAN task [1]. Totally, we identified 16 high activated regions (Fig. 2). Consistent sub-network identification We identify the consistent working memory sub-network via replicator dynamics approach [2] incorporated with group information: Working memory network The result is shown in Figure 5. Measure the structural and resting state functional connectivity patterns This work is finished under the guidance of my advisor, Dr. Tianming Liu. I also would like to thank all my collaborators in the lab and the co-authors of this paper.