Comparison of Independent Component Analysis and Conventional Hypothesis-Driven Analysis for Clinical Functional MR Image Processing
Michelle A. Quigleya,
Victor M. Haughtona,
John Carewb,
Dietmar Cordesb,
Chad H. Moritza and
M. Elizabeth Meyerandb
a Department of Radiology University of Wisconsin, Madison
b Department of Medical Physics, University of Wisconsin, Madison
View larger version (66K):
[in a new window]
|
FIG 1. Text-listening task in a patient with cortical dysplasia involving the left occipital lobe.
A–C, fMR images. The z score map processed with conventional analysis (A) and that processed with ICA (B) are similar. Both show bilateral activation in the auditory cortices. The intersect map of pixels identified in A and B is shown in C; it demonstrates a 91% CR between the maps in A and B.
D, Time course plot from an activated voxel in the conventional analysis shows that the changes in signal are temporally correlated with task performance.
E, Temporal pattern of a selected independent component in the ICA with the highest z score closely resembles the time-course plot of the activated voxel in D. A .70 correlation between the time course of the ICA component and the boxcar reference function was found.
| |
View larger version (63K):
[in a new window]
|
FIG 2. Finger-tapping task in a patient with a left temporal glioma.
A–C, fMR images. The image processed with reference function analysis (A) shows activation in the sensorimotor cortices and sensorimotor area. The image processed with ICA (B) shows a similar pattern of activation. The intersect map of pixels identified on both the image processed conventionally and that processed with ICA (C) has a high CR (83%). Note the similarity of the three maps.
D, Time course plot of the activated voxel shows that the changes in signal are temporally correlated with task performance.
E, Temporal pattern of a selected ICA component (representing the highest z score) has a .64 correlation between the time course of the component and the boxcar reference function.
| |
View larger version (56K):
[in a new window]
|
FIG 3. Word-generation task in a patient with a left arteriovenous malformation.
A–C, fMR images. The image processed by using the reference function (A) and the image processed with ICA (B) are similar. The intersect map of pixels (C) shows that 80% of the pixels identified with the methods were the same (80% CR between the maps).
D, Time course plot selected from an activated voxel shows that the fluctuation in signal is temporally related to task performance.
E, Temporal pattern of a selected ICA component has a pattern similar to that of the time course in the activated pixel.
| |
View larger version (53K):
[in a new window]
|
FIG 4. fMR images in a patient with a right meningioma. For this task, predominant activation typically is seen in the sensorimotor cortex of one hemisphere.
A, Image processed with reference function analysis. The patient moved the left hand according to the on-off commands; however, the patient moved the right hand when instructed to move right hand and when instructed to move the left hand. Therefore, this image shows anomalous activation.
B, However, the image processed with the ICA component specific for activation in the right hemisphere shows the expected unilateral activation pattern.
| |
View larger version (79K):
[in a new window]
|
FIG 5. Findings in a patient with focal seizures and a history of resection of a primitive neuroectodermal tumor and presumptive postoperative complications. When performing the finger-tapping task with either hand, the patient appeared to move.
A–D, fMR images. The z score maps processed with reference to the standard boxcar (A and C) show activation in the sensorimotor cortex and motion-related artifact. The z score maps based on an independent component identified with ICA (B and D) show activation with less motion artifact.
E, The patient’s motion in the inferior-superior direction while performing the finger-tapping task is documented on this graph of inferior-superior motion versus time.
| |
View larger version (69K):
[in a new window]
|
FIG 6. fMR images obtained with the text-listening task in a patient with a cortical dysplasia involving the left occipital lobe. (Images in the leftmost column are those in Figure 1, with the addition of the next three consecutive sections.)
Top row, The z score maps obtained with conventional hypothesis-driven analysis.
Middle row, The z score maps obtained with ICA show good agreement with those in the top row.
Bottom row, The intersect map reflects an 87% CR for these four consecutive sections.
| |
View larger version (55K):
[in a new window]
|
FIG 7. fMR images obtained with a text-listening task in the patient with a left temporal glioma.
Top row, Consecutive z score map obtained with conventional analysis.
Middle row, Consecutive z score map obtained with ICA.
Bottom row, Concurrence in 81% with for these six sections, as reflected in the intersect map.
| |
View larger version (46K):
[in a new window]
|
FIG 8. fMR images in a right-handed patient with a right parasylvian arteriovenous malformation (depicted in the figure) performing the finger-tapping task. In this case, ICA revealed two independent components for the right and left sensorimotor patterns. Therefore, two comparisons were made. The set of images on the left shows the right sensorimotor component, while the set on the right shows the left component. The conventional z score map (top row), spatial ICA map (middle row), and intersect map (bottom row) for the right sensorimotor cortex show a concurrence of 59% for the three sections shown. The left motor sensoricortex reflects a 49% concurrence for the two sections shown.
| |
View larger version (88K):
[in a new window]
|
FIG 9. fMR images in a left-handed patient with a left posterior parietal arteriovenous malformation (depicted in the Figure) performing the finger-tapping task. Note that the four consecutive sections of the conventional z score map (top row), spatial ICA map (middle row), and intersect map (bottom row) are similar. In this case, just one independent component showed the left and right sensorimotor cortices in the same component. This component even included activation in the cerebellum that was found in both analyses. The CR in this four-section example was 73%.
| |
