Biomedical Engineering Reference
In-Depth Information
functional images of the brain, fMRI is an appealing technique because it offers
a good trade-off between spatial and temporal resolution. To increase its tem-
poral resolution, echo-planar imaging (EPI) is used because it makes possible
to collect at least five slices per second at a reduced spatial resolution. The
drawback of this impressive acquisition rate is that it may introduce artifacts
and distortions in the data. More details about these distortions can be found in
[78].
If the distortions do not vary during the time series, they will not affect much
the detection of subtle signal changes, but they will perturb the localization of the
functional activity once being overlapped to the anatomical volume. It becomes
necessary to correct these geometrical distortions in order to accurately identify
activated areas.
Simulated Data.
To evaluate the multimodal registration method, we use
the simulated database of the MNI (Brainweb)
7
[34]. The T1-weighted MR vol-
ume is the reference volume (3% noise and 9% inhomogeneity), whereas the
T2-weighted MR volume is the floating volume.
From the T2-weighted MR volume, we extract a subvolume and we apply
a rigid transformation (three rotations and three translations). To simulate lo-
cal geometrical distortions, we apply a thin-plate spline [16] deformation to the
volume. The thin-plate deformation is computed by choosing one point in the vol-
ume and a displacement for this point. We choose a displacement of magnitude
5 voxels, with no privileged direction. Furthermore, the thin-plate deformation
field is constrained to be naught at the border of the volume.
After rigid registration (see Fig. 8.20), distortions are clearly visible. On the
axial view, ventricles are not well registered ; on the sagittal and coronal view, the
sagittal mid-plane is not well aligned. We then perform the multigrid non-rigid
registration from grid level 7 until grid level 5 to avoid useless computational
efforts. In this case we do not need to estimate a dense transformation, since the
distortions are rather smooth and regular. Furthermore, the statistical similarity
measure is only meaningful for a large number of voxels, i.e., for large cubes.
After non-rigid registration, the internal structures are accurately registered
(see ventricles on the axial view, and sagittal mid-plane on the coronal view for
instance).
7
http://www.bic.mni.mcgill.ca/brainweb
