Modeling and Nonlinear Analysis in fMRI via Statistical Learning - Advance Image Processing in Magnetic Resonance Imaging - page 568

The output of the ST-SVR is a continuous high-dimensional regressed function,
from which the discrete regressed fMRI data can be formed. Therefore, explicit
interpolations required in motion correction, subject normalization, and slice-
timing correction are embedded in the ST-SVR learning framework. The errors
introduced by these interpolations are avoided. This desirable feature is not
available for other methods.
The performance of the presented method is dependent on the validity of the
explicit models. Without other prior temporal information, the on-off boxcar
function in Equation 18.8 is used. Another way is to use a generic function, such
as a gamma or Gaussian function, to model the time course (6, 33). In order to
have reliable prior temporal models, we plan to learn the model functions (or
hemodynamic response functions for event-related fMRI data) from the ST-SVR-
restored data. Instead of relying on a generic parametric model (e.g., gamma,
Gaussian), we would estimate the temporal models through statistical shape
learning without assuming a specific shape of the hemodynamic response. The
learned hemodynamic model can be incorporated to improve the specificity and
sensitivity of fMRI signal detection.
Currently, the proposed ST-SVR method is validated by applying the conven-
tional t -test on the SVR-restored fMRI data for activation detection. We are also
interested in incorporating decision making (activation detection) into ST-SVR
regression, by reformulating the SVR optimization objective function so that sta-
tistical clustering criteria can be optimized as well during data regression. In fact,
validating the hemodynamic response estimation from ST-SVR-restored data is
another way of evaluating the method. In addition, it would be interesting to further
pursue the advantages of the nonlinear system analysis using this ST-SVR approach
in exploring neuronal and hemodynamic responses as well as their interactions.

ACKNOWLEDGMENTS

The author would like to thank Drs. Lawrence Staib, Todd Constable, and Robert
Schultz for their in-depth discussions and suggestions, and for providing the fMRI
data concerning this work. The author is also grateful to Drs. Karl Friston, Will
Penny, and John Ashburner for their valuable comments that improved the quality
of this work.

REFERENCES

1. Backfriender, W., Baumgartner, R., Stamal, M., Moser, E., and Bergmann, H.

(1996). Quantification of intensity variations in functional MR images using
rotated principal components. Phys. Med. Biol. 41: 1425-1438.

2. Benali, H., Pelegrini-Issac, M., and Kruggel, F. (2001). Spatiotemporal covariance

model for medical images sequences: application to functional MRI data. in
Information Processing in Medical Imaging, 197-203.

3. Birn, R.M., Saad, Z.S., and Bandettini, P.A. (2001). Spatial heterogeneity of the

nonlinear dynamics in the fMRI BOLD response. Neuroimage 14: 817-826.