Comparison of independent component analysis and conventional hypothesis-driven analysis for clinical functional MR image processing.

Background and Purpose: With independent component analysis (ICA), regions of activation can be identified on functional MR (fMR) images without a priori knowledge of expected hemodynamic responses. The purpose of this study was to compare the results of fMR imaging data processed with spatial ICA with results obtained with conventional hypothesis-driven analysis.
Methods: Eleven patients with focal cerebral lesions and one with agenesis of the corpus callosum were enrolled. Each patient performed text-listening, finger-tapping, and word-generation tasks. Conventional activation maps were generated by fitting time courses of each voxel to a boxcar reference function. Maps were created from the same image data with ICA techniques. To compare the maps, a concurrence ratio (CR) (number of voxels activated on both maps to number of voxels activated on either map) was calculated.
Results: In the ICA analysis, maps with appropriate spatial and temporal features for auditory, sensorimotor, or language cortices were identified in most patients. Images processed with ICA resembled images processed with conventional means. In patients who moved or performed the task incorrectly, ICA produced a map that resembled the expected activation pattern but differed from the conventional image. CRs averaged 70% for all comparisons in the 12 patients.
Conclusion: fMR imaging maps for auditory, sensorimotor, and language tasks produced with ICA and conventional techniques were similar in most cases. Differences were consistent with the particular characteristics of the method. In data sets corrupted by motion or incorrect task performance, ICA may produce more accurate maps.