Functional inferences vary with the method of analysis in fMRI

Neuroanatomic substrates of specific cognitive functions have been inferred from anatomic distributions of activated pixels during fMRI studies. With declarative memory tasks, interest has focused on the extent to which various medial temporal lobe anatomic structures are activated while subjects encode new information. The aim of this project was to examine how commonly used variations in fMRI data processing methods affect the distribution of activation in anatomically defined medial temporal lobe regions of interest (ROIs) during a complex scene-encoding task. ROIs were drawn on an MRI anatomic template formed from 3D SPGR scans of eight subjects combined in Talairach space. Separate ROIs were drawn for the posterior and anterior hippocampal formation, parahippocampal gyrus, and entorhinal cortex. Twelve different activation maps were created for each subject by using four correlation coefficients and three cluster volumes. Friedman's two-way ANOVA by ranks was used to test the hypothesis that the distribution of activated pixels among defined anatomic ROIs varied as a function of the data processing method. By simply varying the combination of correlation coefficient and cluster volume, significantly different distributions of activation within named medial temporal lobe structures were obtained from the same fMRI datasets (P < 0.015; P < 0.001). The number of subjects studied (n = 8) is in a range commonly found in the literature yet this clearly resulted in spurious associations between processing parameter variations and activation distribution. Using data processing methods that are independent of the arbitrary selection of cutoff values for thresholding activation maps may reduce the likelihood of obtaining spurious results.