Assessment and optimization of functional MRI analyses

Numerous functional magnetic resonance imaging (fMRI) data analysis strategies have been used by different laboratories to enhance the detection of brain activation. A consensus has not been achieved regarding the relative statistical power of different strategies. In this report, we compared several commonly-used data analysis strategies for conventional (nonechoplanar imaging) fMRI data by evaluating statistical testing, spatial filtering, and thresholding. The strategies were assessed using synthetic fMRI images, which were produced by introducing focal "activations" of known intensity, size, and location into "resting" images of six normal volunteers. Three parametric statistical tests (paired t-test, independent t-test, and crosscorrelation coefficient) and three nonparametric statistical tests (Kolmogorov-Smirnov test, Wilcoxon signed rank test, and Mann-Whitney test) were evaluated in terms of sensitivity, specificity, and normality. The results indicated that the independent t-test and crosscorrelation coefficient were the most powerful statistical tests, performing identically well. The effects of smoothing on the detection of activation were assessed. Sensitivity of detecting brain activation can be enhanced by a factor of 8.7 by spatially filtering the raw image data. The optimal full width at half magnitude (FWHM) of the spatial filter was determined to be two pixels (2 mm) in this study. The conventional intensity-only thresholding (IOT) technique was compared with the combined use of the spatial-extent and intensity thresholding (SEe-IT). A highly significant increase in sensitivity (up to 30-fold) was obtained using the SEe-IT technique compared to the IOT technique. Finally, the SEe-IT method was parametrically optimized. A large extent threshold proved more sensitive for detecting brain activation.