Biomedical Engineering Reference
In-Depth Information
3D method performs a non-linear multimodality registration of MRI acquisition
of different subjects. The similarity measure that we use incorporates robust
estimators whose utility is twofold: on the one hand we want to limit the influence
of the acquisition noise, on the other hand, we want to cope with possible
modifications of structures' topology [75].
Since the luminance of MR images might not be directly comparable, we
propose an intensity correction scheme that is anatomically consistent [71].
This correction method will be described in Section 8.3.2. Then volumes to be
registered are rigidly aligned by maximizing mutual information (described in
Section 8.3.2).
Many tasks in computer vision may be expressed as the minimization of a cost
function. The optimization is often difficult to achieve, because the cost function
is non-convex and because the optimization involves a very large number of
variables. Therefore efficient iterative multigrid (or multilevel) approaches have
been developed [67, 93] and applied in motion estimation [48] and in early vision
[133].
To take into account large deformations, we use a multiresolution optimiza-
tion scheme. Besides, at each resolution level, we use a multigrid minimization
to accelerate the algorithm and improve the quality of the estimation. Within
this hierarchical approach, we designed an adaptive partition of the volume to
refine the estimation on the regions of interest and avoid useless efforts else-
where. An anatomical segmentation of the cortex is introduced and used in two
ways: at each resolution level, we initialize the partition as an octree subdivision
based on the segmentation, and the segmentation mask is used in the subdivision
criterion which controls the refinement of the estimation.
The method will first be extensively presented in Section 8.3.2. We will also
present an extension of this method to multimodal data [72] in Section 8.3.2.
Results on synthetic and real data will then be presented in Section 8.3.3.
8.3.2
Method
8.3.2.1
General Formulation
The optical flow hypothesis, or brightness constancy constraint, introduced by
Horn and Schunck [76], assumes that the luminance of a physical point does
not vary much between the two volumes to register. It amounts to zeroing the
