Image Registration Methods in MRI - Advance Image Processing in Magnetic Resonance Imaging - page 192

more robust in respect to those directly related to voxel values. In MR image
processing, signal changes in the image of the same anatomical district can often
occur, so the use of information-based metrics can be useful. Signal changes can
be related to the use of a contrast agent (i.e., first-pass cardiac and brain perfusion)
and to different acquisition device settings (i.e., images of the same patient
acquired in different times).

7.4 THE INTERPOLATION EFFECT

IN THE REGISTRATION PROBLEM

When subvoxel translation or image rotation is involved, image interpolation (also
defined as resampling ) is also required to obtain the roto-translated image. In the
interpolation operation, the coordinate grid is defined by the voxel locations in the
reference image. The voxel values of the roto-translated floating image must be
recomputed in the coordinate grid. Several methods have been proposed for inter-
polation of medical images; an extended review can be found in [16,17]. The main
interpolation methods are truncated and windowed sinc, nearest neighbor, linear,
quadratic, cubic B-spline, and Lagrange and Gaussian interpolations. Because the
interpolation operation has to be repeated for each computation of the similarity
function, both interpolation accuracy and computational complexity are important
in the choice of the interpolation method. In the MRI field, some interpolation
algorithm optimized for MR images [18] have been proposed. These algorithms
use sinc-based interpolation, taking into account the bandwidth of the MR signal
to find the best sinc shape. These methodologies, although more effective in respect
to traditional methods, are usually too slow to be adopted in the solution of the
registration problem. Instead, the interpolation required to obtain the final, registered
image is performed only once, and the choice of an accurate interpolation method
is appropriate.
The main trouble with interpolation operation is that it can modify the gray-level
values, affecting the evaluation of the similarity function. This effect can lead to
incorrect registration results if a histogram-based metric such as MI or NMI is used.
Often, the image dynamic is reduced to avoid this effect.
The simplest interpolation algorithm is the nearest neighbor interpolation, in
which the new voxel values are recomputed as the value of the closest neighboring
voxel. This algorithm preserves the gray values of the original voxels but makes the
registration metric insensitive to intravoxel misalignment because image movements
less than half of the pixel size do not modify the interpolated images.
A more effective interpolation algorithm is trilinear interpolation (bilinear in
2-D images), in which the values of recomputed voxels are evaluated as the
weighted sum of the neighboring voxels. This technique introduces new gray
values in the interpolated image.
When the similarity function is based on joint histogram computation as in
MI, it is preferable to use an interpolation technique that preserves the gray-level
distribution. This technique, called trilinear partial volume distribution (PV)