Modeling and Nonlinear Analysis in fMRI via Statistical Learning - Advance Image Processing in Magnetic Resonance Imaging - page 565

ST-SVR t -test t -test

t > 7.8, p < 10 − 11 t > 4.2, p < 10 − 4

t > 2.3, p < 0.012

(a) (b) (c)

FIGURE 18.7 (See color insert following page 306 ). Comparison of ST-SVR and t -test for

real fMRI data from a visuospatial task (color activation maps).

The 3 × 3 × 1 window covers a brain region whose physical size is almost isotropic

10 mm 3 ). Visual comparisons in Figure 18.7 with results directly using t -
test on presmoothed data (with empirically chosen FWHM

(9.4 × 9.4 ×

= 6.25 mm × 6.25 mm)

reveal that the SVR approach ( W-model 1) leads to greater spatial extent in the

intraparietal sulcus (IPS) with potentially better delineation and localization of the
underlying spatial activation, in agreement with the underlying anatomy. Note that at
the bottom of each slice in Figure 18.7 are the respective t values and p values for
threshold. When the same t threshold used for SVR ( t

=

7.8) is used for the t -test, no
activations are detected. For the t -test in Figure 18.7c, we intentionally further decreased
the t threshold to t

>

2.3 and tried to detect more IPS activation regions, which, however,
lead to a more blurred spatial extent rather than localized spatial activation as in Figure
18.7a as well as some false activations outside of the cerebellum. The associated time
course for an activated voxel from the SVR method for this data is shown in Figure 18.8.
The multirun SVR algorithm is tested using a block-design fMRI motor
experiment on one subject (male). T2*-weighted images were acquired using

>

FIGURE 18.8 (See color insert following page 306 ). Time courses of an activated voxel

for the real fMRI data in Figure 18.7. (Horizontal axis: temporal frame index; vertical
axis: fMRI data intensity).