[Insight-users] Image to Spatial Object Metric
Luis Ibanez
luis.ibanez at kitware.com
Mon Feb 2 14:58:31 EST 2009
Hi Antonio,
It seems that part of the difficulty that you are facing is
rooted in the fact that the metric that computes a binary
overlap between the spatial object and the image will have
a plateau for all the transform parameters that map the
spatial object inside the object in the image.
You may want to consider blurring the image, or computing
its distance map, as a way of breaking that plateau, and
converting it in to a function with a monotonic profile
that could be tracked by an optimizer.
The metric illustrated in the ModelToImage examples is
a very basic one...
Regards,
Luis
--------------------------
Antonio Foncubierta wrote:
> Hello again, it seems i could overcome some of my difficulties
> re-defining a metric. This is what I did:
>
> * First I tried a SpatialObjectToImageStatistis calculator, so that
> I would be able to sum the values of pixels of the image included
> inside the Spatial Object. I couln't make it, and struggled for
> several days on that. So I decided I would look for another
> solution that might not be as efficient, but could be used meanwhile.
> * So I did the following, I used an SpatialObjectToImageFilter and
> counted the pixels of that Spatial Object. This has given me a
> metric that uses the points inside the image ratio, and the pixels
> that includes the spatial object. With this I am able to register,
> but I face the following problem, which is why I ask you for some
> guidance:
>
> Given that my registration algorithm using this unefficient metric is
> quite slow, i would certainly need to have the right scales for dynamic
> range of paratemers. So what's the rule to choose them? I can see
> examples in the ITK Software Guide, but I'd rather have some more
> information about why those values were chosen.
>
> Thank you in advance,
>
>
> Antonio Foncubierta
>
>
> 2009/1/16 Antonio Foncubierta <antonio.foncubierta at gmail.com
> <mailto:antonio.foncubierta at gmail.com>>
>
> Thank you very much for your quick answer, Luis. As requested, I
> will give as many details as possible:
>
> 1. I create an itkImage using a SpatialObjectToImageFilter. This
> image contains an ellipsoid with radii 10 15 20, for instance.
> 2. I want to make a registration of a certain Spatial Object with
> that Image. The starting Spatial Object is another ellipsoid
> with radii, for example, 5, 5, 5. I set the Spatial Object
> offset in a way so that the ellipsoid centre fits the Image
> Gravity Center (I use a MomentCalculator to get that value).
> 3. I start the registration method using an Affine Transform,
> because the starting ellipsoid needs a scale change.
> 4. The result is that the optimizer doesn't have anything to
> optimize, because the SimpleModelToImageMetric GetValue method
> returns the number of Spatial Object points in the Image
> (every point in the image has value=1). Therefore, the metric
> is already maximized. The result is a small ellipsoid fully
> contained inside the image.
> 5. In some other cases, I located the starting spatial object
> outside the gravity center of the image and then started the
> registration method. The result then was a randomly shaped
> ellipsoid (it depends on where I put the center and the
> original size of it) which has all of its points inside the
> image. The Optimizer returns the same parameters once it
> achieves a maximum in the metric, no matter how many
> iterations I allow the registration method to run.
>
> Those results have made me think that I must change the metric so
> that it has a maximum when all the points are inside the image and
> the spatial object volume is as high as possible.
>
> Any ideas?
>
> I also would like to apologize if you understood that I was
> reporting a problem in the registration methods or examples (which
> obviously work perfectly). My intention was to find what I am doing
> wrong or what I could probably use to solve _my_ problem.
>
> Thank you in advance,
>
> Antonio Foncubierta
>
>
> 2009/1/16 Luis Ibanez <luis.ibanez at kitware.com
> <mailto:luis.ibanez at kitware.com>>
>
>
> Hi Antonio,
>
> the ImageToSpatialObjectRegistrationMethod<> that is illustrated
> in the
> examples
>
> Insight/Examples/Registration
>
> ModelToImageRegistration1.cxx
> ModelToImageRegistration2.cxx
>
> should, in principle, work with any of the ITK Transforms.
>
> It is likely that the problem that you faced when using an
> Affine transform are more related to:
>
>
> A) Poor choice of parameter scaling values (the array
> that is passed to the Optimizer in order to compensate
> for the difference in dynamic range between Translation
> and Rotation/Scale components.
>
> B) Poor choice of Optmizer parameters.
>
>
> Of course, the option of writing your own Metric is always worth
> considering, but that is independent of the need for solving (A)
> and (B) properly.
>
>
> In general, when you report that something "doesn't work", it is
> very helpful to describe *exactly* what the result was, and how
> do you arrive to the conclusion that "it doesn't work".
>
> Please give us more detail on what you observed....
>
>
> Thanks
>
>
> Luis
>
>
>
> ---------------------------
> Antonio Foncubierta wrote:
>
> Hello,
>
> I am trying to develop a Model Based Registration Method,
> and I started using the example mentioned by the Itk
> Software Guide and included in the 3.6 version of Insight:
> ***ModelToImageRegistration1.cxx***. I have been trying to
> adapt the example to work with my data: 3D Images. Achieving
> that was not a big deal, but I started to struggle when I
> wanted the Transform to be an itkAffineTransform, which
> includes a scale change. The example explicitly says that
> the SimpleImageToSpatialObjectMetric is designed considering
> no scale changes, and it says it right, because the metric
> doesn't work as I required. (It does work as expected, though).
>
> So my main problem now is finding a way to define a metric
> that maximizes both the ****number of Spatial Object Points
> inside the Image**** and the ****Spatial Object Volume****.
> The first one is easy, as it is exactly what the above
> mentioned example does, but I haven't found a way to compute
> the Transformed SpatialObject Volume.
>
> And that's why I ask you if any of you have found a similar
> problem and how did you solve it. Also, if any of you know a
> not-as-simple metric which I could use.
>
> Thank you in advance,
>
> PS: The explanation migh be very blurry, so if you don't
> undersrtand anything of it, please, just ask. I'd rather
> repeat things in a different way, or attach images to not
> having help at all ;)
>
> Antonio Foncubierta
>
>
> ------------------------------------------------------------------------
>
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org <mailto:Insight-users at itk.org>
> http://www.itk.org/mailman/listinfo/insight-users
>
>
>
More information about the Insight-users
mailing list