Research Abstracts - 2007

Comparison of Data-Driven Analysis Methods for Identification of Functional Connectivity in fMRI

Yongwook Bryce Kim & Polina Golland

Background

Since its development in the early 1990s, functional magnetic resonance imaging (fMRI) has played a tremendous role in visualizing human brain activity for the study of mechanisms of human brains and clinical practice. Acquired from a regular magnetic resonance machine with special parameter settings, this non-invasive imaging method measures changes in blood flow, which in turn are an indication of neural activity. This new imaging technique has generated a large volume of new high dimensional data and hence, the need for new image analysis methods. Much of the work in fMRI data analysis has revolved around the detection of activation at different locations. In addition to localizing activity, we are also interested in how the "activated" voxels are related and connected to each other. Functional connectivity [1,2], the central theme of this project, concerns these functional interactions and coordinated activations among different parts of the brain.

Traditionally, the regression based, hypothesis-driven approach has been used to detect functional connectivity. A "seed" region of interest has to be first selected by the user, and then the network is defined as the areas where their correlations with the seed time course exceed a pre-defined threshold. This method can work well when the goal is identifying regions that co-activate with a certain part of the brain. Hypothesis-driven methods require prior information of the protocol and hypothesis of an experiment to model the expected hemodynamic response. Moreover, an arbitrary choice of the correlation threshold should be made, which is directly related to the statistical significance level.

Problem Setting

Recently, there has been an increasing number of fMRI experiments that investigate the brain activity in a more natural, near protocol-free setting, such as responding to audio-visual input like a movie or rest state scanning. Unlike traditional protocol-based experiments, these new complex experiments do not contain a well-defined onset protocol. Instead, paradigm-free, data-driven exploratory methods such as principal component analysis (PCA) [3,4], independent component analysis (ICA) [5,6], and clustering algorithms [7,8] can naturally provide an alternative to testing each voxel's time course against a hypothesis. They explore the data to find "interesting" components or underlying sources. Structures or patterns in the data, which are difficult to identify a priori, such as unexpected activation and connection, motion related artifacts, and drifts, may be revealed by these components. However, the direct relationship among the data-driven methods is largely unknown and the performance in correctly detecting and classifying functionally connected regions depends on various theoretical and experimental factors.

The main goal of this research is to understand the factors that contribute to the differences in the identification of functional connectivity based on clustering, PCA, and ICA. To our knowledge, this is the first work in the context of functional connectivity, which provides a detailed comparison of PCA, ICA, and clustering, both theoretically (in terms of generative models) and experimentally. With ever increasing volume of complex experimental fMRI data, we believe that our work will provide a better understanding of the functional brain networks and a direction for further analysis.

Collaborators

Professor Bruce Fischl (Harvard Medical School) and Professor John Gabrieli (Dept. of Brain and Cognitive Sciences, MIT)

Research Support

This research is supported under the National Alliance for Medical Image Analysis, the Morphometry Biomedical Informatics Research Network, the Neuroimaging Analysis Center, the NIH NINDS R01 grant on Computational Modeling of Shape Distributions, the Engineering Research Center for Computer-Integrated Surgical Systems and Technology.

References:

[1] B. Biswal, F. Z. Yetkin, V. M. Haughton, and J. S. Hyde, Functional connectivity in the motor cortex of resting human brain using echo- planar MRI. J. Magn. Reson. Med., vol. 34, 537-541, 1995.

[2] D. Cordes, V. Haughton, J. Carew, K. Arfanakis, and K. Maravilla. Hierarchical clustering to measure connectivity in fMRI resting-state data. J. Magn. Reson. Med., vol. 20(4): 305-317, 2002.

[3] K. Friston. Functional and effective connectivity in neuroimaging: a systhesis. Human Brain Mapping, vol. 2:56-78, 1994.

[4] M. Tipping and C. Bishop. Probabilistic Principal Component Analysis. J. Royal Stat. Society, Series B, vol. 61: 611-622, 1999.

[5] A. Bell and T. Sejnowski. An information-maximization approach to blind separation and blind deconvolution. Neural Computation, vol 7: 1129-1159, 1995.

[6] M. J. McKeown, S. Makeig, G. G. Brown, T.-P. Jung, A. A. Kindermann, A. Bell, and T. Sejnowski. Analysis of fMRI data by blind separation into independent spatial components. Human Brain Mapping, vol. 6:160-188, 1998.

[7] A. Dempster, N. Laird, and D. Rubin. Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical Society, Series B , 39(1):1-38, 1977.

[8] C. Goutte, P. Toft, E. Rostrup, F. Nielsen, and L. K. Hansen. On clustering fMRI time series. NeuroImage, vol. 9:298-310, 1999.