[Insight-users] Are there segmented visible human data?

[Weiguang Guan]

Hi Luis,

I appreciate your time, and the segmentation recipes help me improve my
segmentation. It would be very useful if there was a complete collection
of recipes for segmentation of different organs/tissues in images acquired
by different imaging modalities.

Although I agree with you on most of the points you raised I do believe
there are some kind of ground-truths. In other words I believe there is
"the segmentation" of a specific organ given a specific dataset. The
question is how to apply a carefully-chosen, well tuned-up segmentation
process to the dataset in order to get "a segmentation" close "enough" to
"the segmentation" with as little user interaction as possible.

When someone says to segment lungs out of CT scan he means the "sematic
lungs". There could be a ground-truth as long as there was no dispute
about what lungs are composed of. The lungs are distinguishable from
surrounding tissues/organs in at least one physical property except for
its biological function.

The illposedness of image segmentation lies in the fact that a medical
device sensors one or more physical properties (eg, absorption of x-ray in
CT scan) and the object of interest may not be distiguishable from others
in those sensored properties (like You cannot tell whether a bottle
contains liquor or pure water by merely looking at it).  This could be the
case that the object has similiar sensored properties to its surroundings.
What makes it even harder to delineate the object is noise added and
quantization during image acquisition process. 

If the separability becomes a question due to lack of supports from image
more user interatives must be heavily involved, where the segmentation
tool is sort of degraded to a editing tool :-) as one tries to get the
imagined model with a paintbrush. However, there is no criteria to measure
separability just like there is not a well-accepted standard to evaluate a
segmentation.

Weiguang

-- 
========================================================
Weiguang Guan, Research Engineer
RHPCS, McMaster University
========================================================
On Wed, 24 Nov 2004, Luis Ibanez wrote:

> 
> Hi Weiguang,
> 
> 
> 1) You will find products with labelled images
>     from the Visible Human and Visible Woman
>     datasets in the NLM-NIH page:
> 
> http://www.nlm.nih.gov/research/visible/products.html
> 
>     You will also find animations and didactic material at
> 
> http://www.nlm.nih.gov/research/visible/animations.html
> 
> 
> 
> 2) Yes, we have segmented structures such as
> 
>               Liver, Lungs and Heart
> 
>     BUT !!!
> 
>     Note that segmentation is an *ILL-POSED* problem.
> 
>     There is no such thing as "the segmentation" of an organ.
>     Instead you can only get "a segmetation" of an organ.
> 
>     This means that there are many techniques and many
>     parameters for each segmentation technique. If you
>     want to avoid frustration in your segmentation
>     endeavors you *MUST* first clearly define the
>     purpose of your segmentation efforts, so you have a
>     criterion that tells how much is "good enough".
> 
>     A full segmentation of an anatomical structure is
>     rarely feasible. This is due to the continuous
>     nature of biological tissues, and the fact that
>     organs are ultimately made out of cells.
> 
> 
>     The only real segmentation method is "Dissection"
>     You don't want to do that unless you are an M.D.    :-)
> 
> 
> 
> 3) To address you question:
> 
>     Here are some recomendations of segmentation "recipies"
>     since segmentation is not a science but rather and art
>     quite similar to cooking:
> 
> 
>     Lung:  get CT data sets,
>            and use the itkOtsuThresholdImageFilter
>            followed by itkConnectedThresholdImageFilter.
>            be prepared to smooth them with Erosion/Dilation
>            and/or BinaryMedian filter.
> 
> 
>     Liver:  get CT data sets, *avoid* large interslice spacings.
> 
>            - use Watershed
>            - or use ThresholdSegementation Level Set
>            - or use FastMarching Level Set
> 
>            Beware that a liver is composed of several
>            segments and that the anatomical shape of the
>            Liver is *very* variable, since the liver is
>            actually filling in the space left by other
>            organs.
> 
> 
>     Heart:  CT or MR could be used, but note that the distinction
>             between the muscle walls and the surrounding organs
>             is quite subtle.
> 
>             Reject energetically any datasets that has traces
>             of motion artifacts. Keep in mind that cardiac MR
>             is acquired by takin one slice per carciac cycle
>             so technically speaking each slice is in a different
>             time, although they are (hopefully) in the same phase
>             of the cardiac cycle.
> 
>             Note also the "heart" is a very complex organ.
>             You *must* decide whether your definition of heart
>             include the connected veins and arteries or not
>             and whether you want to get the details of both
>             the internal and external walls.
> 
> 
>     Spleen: sorry no recipy for spleen, except the warning
>             the is very close to the liver and very similar
>             in intensities in CT.
> 
> 
> 
> About using RGB data, yes, that helps, but beware that color
> doesn't always have as much information as the human eyes+brain
> tend to believe.  You will find multiple methods in ITK for
> segmenting color images. They include
> 
>        Region growing
>        Statistical classification
>        Fuzzy connectedness
>        Watersheds
> 
> 
> The images in the cover of the ITK software guide were segmented
> from the CT and RGB dataset of the Visible Woman.
> 
> 
> Keep in mind that once you "think" that you are done with
> your segmentation, your supervisor is going to ask you for
> the "Validation". At that point you will ask yourself why
> you didn't went to Law school or rather formed a music band
> instead of working in medical imaging.
> 
> The current standards for validation come down to comparing
> your segmentation with the ones produced by human operators
> using delineation methods.
> 
> 
> Before you start writing code, get an Anatomy Atlas, and
> have a chat with a doctor. That will help you understand
> the structures that you are trying to segment and what
> pieces of them are really important.
> 
> 
> 
>     Regards,
> 
> 
>        Luis
> 
> 
> 
> ---------------------
> Weiguang Guan wrote:
> 
> > Hi ITK users,
> > 
> > Is there a well-labeled visible human image data (each voxel is classified
> > to as the organ or tissue it belongs to), either as a commercial product
> > or in public domain? I would appreciate it very much if you could give me
> > some information. Thank you very much in advance.
> > 
> > Has anyone succeeded in segmenting liver, kidney, heart, spleen out
> > of 3D CT scan with ITK? Do you think it would be easier to segment them in 
> > MRI or visible RGB images than in CT? 
> > 
> > Weiguang
> > 
> 
> 
> 
>