[Insight-users] preparation about segementation

[Luis Ibanez]

Hi Ouyang,

The most important preparation step for segmenting the
Visible Human data set is to make sure that you have
enough coffee available    :-)



About your questions:

1) The following options have been used when segmenting
    color images:

   A - Using the Statistical classification approaches.
       This is 'classification' in general. You will
       find multiple examples of this approach in the
       directory:

         Insight/Validation

      in the two subdirectories:

           - StatisticalClustering
           - TissueClassification

      The reports of these validation studies are available
      in the InsightDocuments cvs module:

               InsightDocuments/Validation

    B - Simply ignore color, and use a single channel.
        For registration, it has been found that the
        red channel provide enough information for
        achieving multi-modality registration.

    C - Modify methods like Confidence connectedness in
        order to replace the one-dimensional intensity
        range criterion with a three dimensional one,
        at the style of a Mahalanobis distance.

   I would suggest you to explore option "C".



2) Every anatomical structure has different
    characteristics. You will have to focus in particular
    organs and decide what sections of particular organs
    are of your interest. "Segmentation" is a totally
    ill-defined problem, for example, when you mention
    'Chest', it is unlikely that two doctors will give you
    the same definition of what 'chest' is.  Does it finish
    in the diaphragm muscle ? doees chest include the collar bones ?
    does chest include the shoulders ?

    Even when you focus in a particular organ, there are
    organ components that have totally different tissue
    characteristics. For example, if you talk about lung
    segmentation.. do you want the ventricles to be part
    of the lungs ?  If you want to segment the heart, do
    you consider the arteries part of the heart ?

    You will have to define well the particular segmentation
    task that can have some anatomical meaning, only then
    you can start trying segmentation methods.

    Among the most popular methods in ITK, you may want to
    try:

    a) Watersheds
    b) LevelSet methods:
        FastMarching
        ShapeDetection
        GeodesicActiveContours
        Threshold + levelsets
        Canny + LevelSets
        Laplacian + LevelSets
     c) Region growing methods
          Confidence connectedness
          Threhold connectedness
          Fuzzy Connectedness
          (note that there is a version of Fuzzy
          connectedness that is designed for RGB data).
     d) Deformable models
     e) Cellular Aggregates



3)  The most common preparing step are noise reduction
     approaches. In particular you want to use any of the
     'edge-preserving' smoothing methods described in the
     software guide.  Note that there are specific methods
     for RGB data.  They are described in section 5.5.3
     pdf-page 128.
     http://www.itk.org/ItkSoftwareGuide.pdf

     I would suggest you to get familiar with the

     VectorGradientAnisotropicDiffusionImageFilter  and
     VectorCurvatureAnisotropicDiffusionImageFilter

     Examples for both of these filters are availabe in
     Insight/Examples/Filtering.

     Note that 'edge-preserving' is not an absolute,
     if you apply enough iteration of those filters, you
     will equally wash out all the edges,... so,
     use with moderation ! (in any case the computational
     demands of these filters will prevent you from going
     too far.)


4)  Make more coffe...


5)  We have talked about the need of a document with
     'Recipies for Segmentation'. Contrary to what we
     would like to think, Image processing is not an
     exact science, and it is hard to anticipate what
     combinations of methods will work in particular
     structures. So we will be interested in collecting
     users-experiences with particular methods, image
     modalities and anatomical structures.



Regards


    Luis


-------------------
ouym99 wrote:
> Hi, Luis, Hi. Everyone.
> I am about to segment a complete set of visible human male data using ITK
> software.
> Based on the acknowledgement of it as a challenging job, I sincerely need your
> suggestions in the following three aspects:
> 1)	How can I get segmented regions the same color as the original image. Since the
> image is in tiff24 format, and that many methods in ITK, such as
> ConnectedThresholdImageFilter/ WatershedImageFilter, etc, could only result in
> binary segmented images, so how can I get a color segmentation in ITK?
> 2)	Which method, generally speaking, is more often used to segment the particular
> part of the VHP? Namely that, which method to segment the limb, bone, muscle;
> which method to segment trunk, chest, abdomen? And so on.
> 3)	What should the preparing steps to be? Right now , I thought the preparations
> like: acquire of the anatomic structure of the human body; browse of the image set
> so as to make a plan to segment from the easiest part to the hardest; a plan about
> how to create a serial number. And , from your point of view, what else
> preparation should I make?
> 
> Sincerely yours,
> Ouyang
> 
> 
> _______________________________________________
> Insight-users mailing list
> Insight-users@public.kitware.com
> http://public.kitware.com/mailman/listinfo/insight-users
>