[Insight-users] [insight-users] 3D sequence registration . ( Fighthing Superstition )

[Luis Ibanez]

Hi Marco,


Thanks for the detailed description of the problem
that you are working on.


Please note that:


0) The is *no such thing* as the "most suitable" or the "best"
    metric for a given registration problem.

    The popularization of the misguided notion that there is
    a "best method" for solving image processing problems is
    the result of the *decadence* of the publishing system in
    our field.  Those who claim that they have a "best method"
    for any kind of problem are confused or are trying to
    confuse you; probably to try to sell you something that
    you don't need to buy.

    The desire to claim that a method is better than any other
    one was born from absence of scientific rigor in publications
    that disregard the importance of reproducibility in the scientific
    method, and are simply dedicated to fulfilling the needs of
    academics for filling up their yearly quotas of intellectual
    "production" by publishing a require N number of papers a year.

    First of all, *all* methods have parameters, and by changing
    the numerical values of these parameters you get *entirely*
    different results. Therefore any claim that "Method A" is
    better than "Method B", but doesn't list the set of parameters
    used on each method, has to produced in a context of ignorance,
    incompetence or malice.

    An honest an educated claim will instead report
    something like:

      Method A (as implemented in the source code that you can
      download from website Xa), when run on the images K, L
      (that you can download from the website Y), using the
      set of parameter (p1=0.5,p2=2.3.....etc) produce the
      following results (R1), while

      Method B (as implemented in the source code that you can
      download from website Xb), when run on the images K, L
      (that you can download from the website Y), using the
      set of parameter (q1=0.7,q2=7.3.....etc) produce the
      following results (Z1)

      and in the context of the application (AA: which can be
      radiation treatment, or surgical planning, or teaching,
      ...) we consider that the result R1 is better than Z1.

    That is a serious technical report,
    while the simple claim:

             "Method A" is better than "Method B"

    is simply a *superstition*.


    Please,
    demand from authors and editor of technical publications
    to raise their standards to the at least the basic
    scientific level that undergrad students will learn in
    Physics 101.

    Please,
    Refuse to collaborate in the dissemination of superstitious
    beliefs.



That being said,


1) Mean squares may still be able to register your images,
    if the propagation of the contrast agent between Image N
    and image N-1 is not too large, and if there are enough
    regions of high intensity contrast (not agent contrast)
    that could guide the metric down the registration process.

    But of course, only a set of experiments that include
    a process of parameter fine-tunning will tell for sure
    if that is the case for your specific set of images.


2) The Demons algorithm is essentially using a Mean Squares
    metric. The motion that it computes is represented as
    a vector field, where you get a displacement vector per
    pixel, and the vector field is smoothed using Gaussians.


3) BSpline and MutualInformation:  Both of these classes
    have a set of about ten parameters combined

    * number of grid point in the Spline
    * spline order
    * number of samples in the metric
    * number of histogram bins
    * Sigmas for the distributions

    not to mention the parameters of the optimizer that
    are typically 3 to 5.

    If the registration is not capturing large deformations,
    you may have to consider a multi-resolution approach to
    the BSpline grid, such as the one used in

         DeformableRegistration15.cxx


4) You may also find useful to try the Demons multi-resolution
    examples in

              DeformableRegistration16.cxx
              DeformableRegistration17.cxx



Specific suggestions will require that you tell us more
about the images that you are registering

Anatomy : (brain ? lung ? liver ?)
Modality : MRI ? CT ?

Screen shots of the images and the current results would
be great.

BTW, you may be interested in trying the application VV
http://www.midasjournal.org/browse/publication/288
that is designed for visualizing the result of these kind
of registration problems.



     Regards,



         Luis



---------------------
marco giordano wrote:
> Hi all,
> 
> I am doing a registration of a 3D temporal sequence that consist of 40
> frames (Vol1 .... Vol40).
> .
> The sequence shows the contrast uptake in vessels and human tissues
> and during the sequence some squeezing and non rigid motion occurs .
> 
> Normally I register all the frame ( Vol2 ... Vol 40) to the first
> frame (Vol1) or use a concatenation of registration (register VolN to
> Vol(N-1)registered)
> 
> For such kind of problem I heard that the Mutual Information is the
> most suitable metric.
> 
> In fact a metric as meansquaredifference would try to minimize the
> difference in intensities, but the intensities are already different
> for
> the fact that contrast changes the intensity in each frame.
> 
> So far I tried:
> 
> -DemonsAlgorithm DeformableRegistration2.cxx (registration is good but
> the some changes occur in the intensities )
> Anyway I do not undestand what kind of metric is used here and for
> what kind of motion is most suitable for this algorithm
> 
> -Bspline with Mutual Information DeformableRegistration8.cxx ( here
> strong motion is not completely removed ).
> 
> I wanted to ask if anyone has suggestion on how to carry on and if
> there is some parameters that can be changed
> e.g:
> 
> -the grid spaces etc.
> -number of iteration
> -the optimizer
> 
> Any help would be appreciated
> 
> PS: I would like to cite ITK whenever the results are suitable for an article.
> 
> Thanks
>