[Insight-users] vertebrae segmentation with GeodesicActiveContourShapePriorLevelSetFilter

[Juan Cardelino]

Giving a fast look at the results, I see the biggest problem on the
upper side of the shape, in that zone the boundary term is not good to
achieve the results, maybe increasing the weight of the shape term
could help. In addition, on the lower part of the image it seems that
the curve is attaching itself to the outer countour of the edge map
(where the white zone starts), and not to the correct boundary (thin
white countour). This could be because of the inflation (advection)
term pushing the curve to go outside. To correct this, you could
decrease the weight of the advection term, and/or shrinkig the initial
countour to be slightly smaller than the desired shape, in order to
have it inside of the object but very close.
This trials could let know you if there is any set of parameters well
suited for your image. I think that results are close, maybe will a
little tailoring of the parameters you could get a decent
segmentation. Again, I don't know the precision required in your
application, so I can not tell if it is achievable with this
technique.
I hope these comments help you.
Best regards,
                   Juan

On Wed, Jun 23, 2010 at 9:42 AM, Paul <pare85 at googlemail.com> wrote:
> I set the desired final segmentation as initial condition. Here some new
> pictures.
>
>
> edge map:
> http://250kb.de/xJNMujE
>
> new mean shape image and initial condition:
> http://250kb.de/9FtvL84
>
> segmentation result:
> http://250kb.de/Kwucded
>
> parameter
>  - propagation 0.5
>  - curvature 1
>  - advection 3
>  - shape 0.2
>
>
> Regards
> Paul
>
>
> Am 21.06.2010 14:04, schrieb Juan Cardelino:
>>
>> One good thing to try, is to use the desired final segmentation as
>> initial condition, and check if the evolution of the curve keeps it in
>> place. One interesting thing to see is the edge map, if it is good
>> enough it will tell you that you need less weight on the shape term.
>> On the contrary, if the boundary map isn't good enough, maybe it is
>> wise to raise the weight of the shape term.
>> Could you please post the new results? It will be interesting to see
>> how it goes.
>> Regards,
>>              Juan
>>
>> On Sun, Jun 20, 2010 at 6:25 PM, Paul<pare85 at googlemail.com>  wrote:
>>
>>>
>>> Hello Juan,
>>>
>>> thank you for your prompt reply. You are right. The number of iterations
>>> was
>>> too small. The segmentation results are better now but they still aren' t
>>> good enough. I will try some other parameter combinations, create a
>>> better
>>> mean shape image and check the edge image again. Maybe I can optimize
>>> some
>>> of this things.
>>>
>>> Best regards,
>>> Paul
>>>
>>>
>>>>
>>>> Hey Paul,
>>>> While I can't assure that this technique would work, I'm pretty
>>>> confident that is a good path to follow. However, as they say, the
>>>> devil is in the details, so maybe you will need some fine tuning to
>>>> make it work.
>>>> Regarding the particular problem with the results, I find interesting
>>>> to see the final number of iterations in each case. It seems to me
>>>> that your problem is that the default maximum number of iterations is
>>>> too small. You can tell that because the front stop advancing in a
>>>> quite arbritrary point, and not on an actual minimum of the energy.
>>>> The second example only confirms my thesis, because the only thing you
>>>> did is to increase the propagation speed, which partially aleviates
>>>> the problem, but you are still stopping at an arbritrarly point. You
>>>> can confirm this by looking at the console output of the program and
>>>> see the final number of iterations. If that happens to be the problem,
>>>> I suggest that you increase the number of iterations to a ridiculous
>>>> number, like 10k, and see at what number it stops. It will be also
>>>> useful if you post the output of the program.
>>>> I wish you luck with the problem, and let us know how it turned out.
>>>> Best regards,
>>>>                     Juan
>>>>
>>>> On Sat, Jun 19, 2010 at 10:34 AM, Paul<pare85 at googlemail.com>    wrote:
>>>>
>>>>
>>>>>
>>>>> Hello mailinglist,
>>>>>
>>>>> I' m using the GeodesicActiveContourShapePriorLevelSetFilter to segment
>>>>> 2D
>>>>> ct data of vertebrae. I can't find the right parameters to get a good
>>>>> segmentation of the typical vertebra shape. Do you think, it is
>>>>> possible
>>>>> to
>>>>> get a good segmentation of vertebrae with the
>>>>> GeodesicActiveContourShapePriorLevelSetFilter? The following links show
>>>>> some
>>>>> segmentation results, the input image and the mean shape image I'm
>>>>> using.
>>>>> Maybe you can give me some advice how I have to change the parameters.
>>>>> I
>>>>> tried some parameter combinations but did not get one satisfying
>>>>> result.
>>>>>
>>>>> input image:
>>>>> http://250kb.de/qJsgisV
>>>>>
>>>>> mean shape image:
>>>>> http://250kb.de/e3leuKZ
>>>>>
>>>>> segmentation result with seed points: 3 propagation=2 advection=3
>>>>> curvature=1 shape=0.5 distance=5
>>>>>
>>>>> http://250kb.de/oamCuus
>>>>>
>>>>> segmentation result with seed points: 5 propagation=4 advection=3
>>>>> curvature=1 shape=0.5 distance=5
>>>>>
>>>>> http://250kb.de/KyAGhW7
>>>>>
>>>>>
>>>>> Regards
>>>>> Paul
>>>>> _____________________________________
>>>>> Powered by www.kitware.com
>>>>>
>>>>> Visit other Kitware open-source projects at
>>>>> http://www.kitware.com/opensource/opensource.html
>>>>>
>>>>> Kitware offers ITK Training Courses, for more information visit:
>>>>> http://www.kitware.com/products/protraining.html
>>>>>
>>>>> Please keep messages on-topic and check the ITK FAQ at:
>>>>> http://www.itk.org/Wiki/ITK_FAQ
>>>>>
>>>>> Follow this link to subscribe/unsubscribe:
>>>>> http://www.itk.org/mailman/listinfo/insight-users
>>>>>
>>>>>
>>>>>
>>>>
>>>>
>>>
>>>
>>
>>
>
>