[Insight-users] Registration & resample problem

[Luis Ibanez]

Hi Dean,

This sounds like a very interesting problem.

Having a strong anisotropic resolution introduces
a lot of difficulties for processing the data.

Resampling the volume in the orthogonal directions
may help... but still is not a substitute for the
real solution which is to acquire isotropic data sets.

In any case, Let's assume that you acquired the
three datasets and have succceeded registering them.
You will be in the situation where you have values
sampled with the three image acquisition grids,
and they are overlapped in space (more or less orthogonally).
In order to feed these values in a common interpolation scheme
you could use any of the KernelBased interpolators in ITK.
A large set of options is available:

http://www . itk . org/Insight/Doxygen/html/classitk_1_1KernelTransform . html

     - Thin Plates
     - Thin Plates R2 Log R
     - Elastic Body
     - Elastic Body Reciprocal
     - Volume


This will be computationally intensive (and will require
some serious memory) depending on the number of pixels
acquired from your 1cm^3 of tissue.  The advantage of
this method is that you simply dump all the pixel data
in the form of 4D points, where the first 3D components
are the space coordinates of the pixels in 3D (after
registration), and the fourth component is the gray
level intensity of the data set.

These 4D points will be provided in two groups as
4D landmarks (source and target landmarks). The
source landmarks will have their 4th component equal
to zero, while their homologous target landmarks will
have the 4th component equal to the intensity level
of the image.

For resampling the data you could still use the
ResampleImageFilter, since the KernelBased transform
are part of the Transform family.


Note that you will have to make sure that the 3 data
sets have a similar distribution of graylevel. You
may use the HistogramMatchingImagefilter
http://www . itk . org/Insight/Doxygen/html/classitk_1_1HistogramMatchingImageFilter . html
before you perform the registration. Since it is quite
likely that the three MRI dataset will have different
intensity distributions.


Resampling the data may require a significant amount
of effort and the result will be far from perfect (specially
if the anisotropy ratio is larger than 1:5). Just take a
quick look at the shapes of the Voronoi regions that will
result from combining the three orthogonal grids in a single
irregular grid and then look at the Delaunay regions which
is where the interpolation will take place. This will show
how inconsistent the final resampling will be.

So, if you still have a chance of convincing your MRI acquisition
colleagues of doing the right thing an acquiring isotropic data,
you may save a lot of effort, suffering and frustration.

It is unfortunate that the image acquisition crowd tends to dismiss
the fundamental principles of the Shannon theorem when it comes to
acquiring images in 3D and then expect the image processing crowd
to do magic with the data as if the Nyquist limit where only defined
in 1D.

The visual inspection of high-resolution in-planes slices only
create a false-sense of security regarding the actual information
contained in the image.

If we were working in the military and looking for mines in a field,
I doubt somebody will suggest to sample the terrain at every 10cm in
the X direction and at every 100cm in the y direction... The idea
seems to be quite natural for our colleagues acquiring CT and MRI
scans though.




Regards,



   Luis


---------------------------------
dean . inglis at camris . ca wrote:
> The problem I have is to image a small
> volume of tissue (~ 1 cm^3) at high
> spatial resolution with 3DGRE MRI. 
> In such a data set, the (isotropic) 
> in-plane resolution is >> than the slice 
> thickness. The constituent features within 
> the volume of tissue are themselves quite 
> small; on average about half the size of 
> the in-plane dimension.  To improve the 
> overall spatial resolution, I am 
> contemplating acquiring an additional data
> set, scanned in one of the other two anatomical 
> planes (i.e., dat set 1 is axial, data set 2
> is coronal or saggital).  My proposal is to 
> employ rigid body registration to align the two 
> data sets and then resample/interpolate them 
> into one data set having increased and isotropic 
> spatial resolution. There are other larger
> scale anatomical features present in the 
> volume that are readily distinguishable, so
> the registration phase should not be problematic. 
> Does anyone have any suggestions for a
> pipeline or some filters that may be
> appropriate for the second phase involving 
> genaration of the resampled/interpolated/
> combined data set?
> 
> regards,
> Dean
> 
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk . org
> http://www . itk . org/mailman/listinfo/insight-users
>