Visualization of cortical activity using fMRI
Imaging Technology Department, HP Labs, and Stanford University*
Summer 1996
Sponsored by the HP Labs Grassroots Basic Research Program
Teo, Sapiro & Wandell, "Creating connected representations of cortical gray matter for functional MRI visualization", IEEE transactions on medical imaging 16 (6), 852-863, 1997
Background
The goal of this summer project was to investigate and to improve existing techniques involved in the visualization of human cortical activity using functional magnetic resonance imaging (fMRI). In conventional magnetic resonance imaging (MRI), a volume of data representing the anatomical structure of the cortex is reconstructed. Functional MRI, on the other hand, indirectly measures the amount of neural activity in different regions of the cortex. Cortical activity occurs largely in the gray matter regions of the cortex. Thus, most visualization techniques require prior segmentation of this gray matter component.
Figure 1: MRI of the occipital pole of the left hemisphere of a human brain.
Gray matter segmentation is difficult because gray matter forms only a thin (2 mm or 2 pixel) layer on the surface of the cortex, which itself is inundated with numerous ridges and folds (as shown in Figure 2). Currently, much of gray matter segmentation is performed manually. This process is tedious and requires several days even for a trained person. On the other hand, most automatic gray matter segmentation methods that have been proposed do not consider anatomical constraints and often produce segmentations that are anatomically incorrect. We propose a new semi-automatic method of gray matter segmentation that takes into consideration these anatomical requirements.
Figure 2: Photograph of a cross section of an actual human cortex.
Method
Our proposed method consists of four steps. In the first step, white matter is segmented. White matter is segmented because it is less noisy and suffers less from partial volume effects. Figure 3 shows the results of segmenting the white matter in Figure 1. The maximum aposteriori probability (MAP) estimate coupled with a structural prior is used to determine segmentation. The structural prior is applied using a novel technique involving anisotropic smoothing.
In the second step, the white matter connected component corresponding to the cortex is selected using a 3D connected components algorithm. After this single connected component is extracted, the topology of that component is verified in the third step. Since, in the anatomy, gray matter borders on white matter, the topology of gray matter places constraints on the topology of white matter. Specifically, because gray matter is a sheet, the white matter component cannot have cavities or handles. Cavities are identified using a flood-filling algorithm. Handles are detected by computing the Euler characteristic of the white matter.
In the final step, the gray matter classification is grown out from the boundary of the white matter component (as shown in Figures 4 and 5). Since gray matter is at most 2 mm (or 2 pixel) thick in this direction, only two layers of gray matter classification are grown. Connectivity between gray matter voxels are determined from the connectivity of their parents from which they were grown. Thus, gray matter voxels may be adjacent in the volume but not connected on the gray matter surface.
Figure 3: Segmentation results of white matter.
Figure 4: Segmentation results of gray matter.
Figure 5: Segmentation results of gray matter overlaid on original MR image.
Figure 6: Flattened gray matter of the left occipital lobe of a human cortex. Gray matter was segmented semi-automatically by the proposed algorithm. Colors represent the 3D distance of the original gray matter voxel from a fiducial plane. Brighter colors denote larger distances.
Conclusions
Manual segmentation of the occipital pole of the left hemisphere (less than one eighth of the cortex) takes about 18 hours typically while the semi-automatic procedure proposed here requires only about 15 minutes. Furthermore, the procedure ensures that the gray matter is topologically/anatomically correct in that no tears or self-intersections occurs. Once the highly convoluted gray matter regions are segmented, they can then be unfolded and flattened so that regions hidden deep in the folds become completely visible (as shown in Figure 6). Functional MR images can then be overlaid on this flattened representation so as to give a better understanding of neural function in the different areas of the cortex.
Acknowledgements: Hagit Hel-Or (Stanford) for the original version of the segmentation tool. Steve Engel (Stanford) for the brain unfolding software. Geoff Boynton and Jon Demb (Stanford) for expertise on anatomy and f/MRI. Tom Malzbender (HP Labs) for discussions on graphics and visualization. Daniel Lee, Alex Drukarev, and the Grassroots Research Program Committee (HP Labs) for supporting this research.