[ITK-users] Cardiac Deformation using Segmentation and Registration of Ultrasound Images

[Matt McCormick]

Hi Thanos,

On Thu, Aug 3, 2017 at 1:00 PM, thanos thanos <thanosxania at gmail.com> wrote:
> Hello everyone,
>
> Dženan, Andras and Matt thank you very much for your answers.
> To be honest I've tried already in the past registration with Slicer using
> Demons, B-Splines and also Elastix and Plastimach.
>
> Dženan, I read your last sentence already many times but I can't really
> understand what you mean by "initializing registration of time point n+1 by
> resulting transform from time point n."
>
> Matt, thank you for sharing your thoughts. I already looked at the
> AnisotropicDiffusionLBR paper and I have also tried the online interactive
> figures
> (https://insightsoftwareconsortium.github.io/ITKAnisotropicDiffusionLBR/).
> Seems to work nicely, but I wasn't sure how to build the project. I already
> configured my ITK by setting the Module_AnisotropicDiffusionLBR at the CMake
> and then I tried to build the CoherenceEnhancingDiffusion from the
> C:\ITK\src\Modules\Remote\AnisotropicDiffusionLBR\examples but I got "
> Cannot open include file: 'CoherenceEnhancingDiffusionFilter.h': No such
> file or directory " so I guess I am missing something..

Thanks for the note. I will look into this issue and address it next week.


> I tried in the past to perform segmentation using the
> GeodesicActiveContourImageFilter (first for 2D images) but I miserably
> failed to get a descent result. I then tried to extend it to 3D but I think
> that even though I built the project I was not getting valid results (it was
> some months ago so I don't remember exactly).
>
> Now, on important question that I have is how important is it to have a
> perfect segmentation in order to perform registration? Is it maybe enough to
> just get a just descent segmentation that includes the whole heart even if
> that means that the active contours stop before the reach the actual
> boundary?
> I guess I have to first proceed in a good segmentation before going back to
> the registration problems..

In general, "it depends," but segmentation and registration problems
can be viewed as coupled. In general we need to identify structures
before we can register them. Preprocessing filtering step that reduce
noise could be viewed in this respect.


> Also, I suppose that the registration should not be done to the filtered
> images but to the original in order to preserve the speckle noise which
> includes some kind of information. Is that correct?

Right. The speckle is not actually noise. It is signal that does not
change and indicates the location of tissue. Some motion tracking
techniques, like ultrasound strain imaging, use the speckle signal
exclusively. However, some registration techniques may be confounded
by local minima in the cost function caused my high frequency content.
And, the speckle can change when motion is large and the ultrasound
beam encounters the tissue at a different angles.


> I will have a look within the next days to the last two approaches that you
> suggested, although, with a first glance, it wasn't quite obvious how the
> PointSetToPointSetMetric works, but I will try to study it a bit more.

There are some tests in the ITK test suite that demonstrate how to use
the new point set registration framework.


Hope this helps,
Matt