[Insight-users] MultiResolution registration : tunning the Translation
scaling
Luis Ibanez
luis . ibanez at kitware . com
Wed, 05 Nov 2003 11:48:15 -0500
Hi Jeese
The translation scaling is a critical parameter in
the registration process. You are certainly encouraged
to tune this value in order to register your images.
This parameter has an initial nominal value intended
to compensate the difference in dynamic range between
the translation components of the transform and the
set of {rotation/scaling/shearing} components.
Translations are expected to be in the range of half the
diagonal of the image (measured in millimeters), while
rotation/scaling/shearing parameters are in the ranges
of [-1:1], [1/2:2],[-1:1] respectively. Note that those
ranges indicate typical values, by no means they indicate
that the parameters are bounded to this range.
Once the translation scaling parameter has compensated for
this dynamic range difference, you still have the freedom
of tunning the parameter in order to regulate how much
rotation versus how much translation the optimizer will
try to compensate. In principle, this parameters should
be tunned differently for every registration problem
depending on the relative amount of rotation versus
translation in the image missalignment.
Note also that in and ideal multi-resolution application
this parameter should be tunned differently for every
resolution level. For example, at coarse resolution you
should concentrate in solving only translations since
the blurring and degradation of the image due to subsampling
probably will remove most of the anatomical details required
for solving rotation missalignemnts. Higher resolution levels
should gradually concentrate more on rotation and less in
translation since in principle the previous coarser resolution
level should have solved for the bulk of the translation
missalignment.
Please let us know if you have further questions,
Regards,
Luis
----------------------
iajesse wrote:
> Hi Luis,
>
> I have experimented according to your advices, some dataset were
> aligned, but others were misaligned. I'm not sure weather the dataset
> have problems. can i turn the parameter 18 (The scale applied to the
> translation parameters during optimization.)?
>
> */Luis Ibanez <luis . ibanez at kitware . com>/* wrote:
>
>
> Hi Jesse,
>
>
> When you use MultiResolution, it is convenient
> to set a very large number of iterations in the
> coarsest level as well as a small learning rate.
>
> Don't care much about the subsequent resolution
> levels until to succed to get convergence in the
> coarsest level.
>
> Once you get the first level to converge (that
> is, to stop before using all the allotted
> iterations), play increasing the learning rate
> and decreasing the number of iterations for this
> coasest level. The goal is to get it to converge
> using the largest possible learning rate.
> This will result in the shortest convergence time.
>
> Once you have done this tunning for the coarsest
> level, you can keep the parameters for this level
> and start doing the same tunning on the next
> resolution level.
>
>
>
> Regards,
>
>
> Luis
>
>
> -------------------
> iajesse wrote:
> > Hi Luis,
> >
> > The output file is ok now, but another question occur.
> >
> > Now, There are 30 patient dataset, each patient have T1 and T2, I
> want
> > to register the T1 to T2, and apply MultiResMIRegistration on them,
> > but about 10 patients are misaligned, I tune some parameter, such
> as:
> > The number of iterations, learning rate , but improvement is little.
> > Could you please give me some advice on how to tune parameter? for
> > example, which parameter should be changed first?
> >
> > Thanks
> >
> > Jesse
> >
> > */Luis Ibanez /* wrote:
> >
> >
> >
> > Hi Jesse
> >
> > The ITK terminology for registration has changed over time.
> >
> > In the early days we used to talk about "target" and "source"
> > image. Then we switch to ! "fixed" and "moving" which are the
> > terms used today, and is how you will find the descriptions
> > in the software guide.
> >
> > This application is one of the earliest on the toolkit and
> > it is therefore making use of the original terminology in
> > some sections.
> >
> >
> > If you look at the code of the application
> > (Ah, the pleasure of open source !), you will see in:
> >
> >
> > InsightApplications/MultiResMIRegistration/Common
> >
> > SimpleAppOutput.txx
> >
> > That the correspondences are:
> >
> > "target" ----> "fixed"
> > "source" ----> "moving"
> > "register" ----> "resampled moving"
> >
> >
> > Please find attached a modified version of this file
> > that will save the images as M! etaImages instead of
> > series of PGM files. MetaImage use a text header file
> > and a binary raw data image file.
> >
> > Use it to replace th! e one in your system. In that way
> > you will not need to go through the extra step of using
> > Mricro for viewing the images.
> >
> >
> >
> > Please let us know if you find further problems.
> >
> >
> > Regards,
> >
> >
> > Luis
> >
> >
> >
> >
> > ------------------------
> > iajesse wrote:
> > > Hi Luis,
> > >
> > > I have a problem on output of MultiResMIRegistration ,
> > >
> > > There are three dataset in ourput,
> > >
> > > register000.pgm ....
> > > source000.pgm,,,
> > > and target000.pgm ...
> > >
> > > I convert them into raw image using Mricro, but find that the new
> > target
> > > image is not the old one, and source is not the old one. I think
> > that
> > > the target is the fixed, and source is floating, why target change?
> > >
> > > Best
> > >
> > > Jesse> >
> > =========================================
> >
> >
>
> __________________________________________________
> Do You Yahoo!?
> Tired of spam? Yahoo! Mail has the best spam protection around
> http://mail . yahoo . com
>