[ITK] Segmentation of PET image

[Timothee Evain]

If you have good enough markers maybe a denoising isn't mandatory since your high signal area seems homogeneous.
But if you need it, I would recommend a AnisotropicDiffusion filter (aka Perona-malik) : http://www.itk.org/Doxygen/html/classitk_1_1GradientAnisotropicDiffusionImageFilter.html
Quite computationally heavy, but it's great to smooth noisy flat areas while sparing edges. 
You have to set only the conductance parameter (~ "smoothing force") and the number of iterations.
However, I've never tried it on PET data.

Regarding the coarse segmentation, k-means sound way heavier to me than just a threshold but I'm no specialist.
By the way I have forgotten to precise that I was thinking to an Otsu thresholding and not just an empirical one.
http://www.itk.org/Doxygen/html/classitk_1_1OtsuMultipleThresholdsImageFilter.html

Good luck!

Tim 


----- Mail original -----
De: "Cyril Jaudet" <drcjaudet at gmail.com>
À: "Timothee Evain" <tevain at telecom-paristech.fr>
Envoyé: Vendredi 31 Juillet 2015 13:17:43
Objet: Re: [ITK] Segmentation of PET image

Exactly it is FDG PET image so it rely on glycolysis consumption. In fact
there is a paper on a Growing method (E. Day et al., 2009) for this kind of
structure but it requires a calibration procedure, so not very easy too use
in a multicentric setup.  I'll try the method you suggest also do you think
that i have to use a prefilter to decrease the noise as bilateral filtering
for example and an border enhancement filter as gradient magnitude image
filter. Geets and al. in 2007 have implemented a simlar methode but it need
denoising and deconvolution before watershed segmentation so you really
have to tune it up also as i want to perform texture based analysis
modifying the image may not be a good solution.
For the coarse segmentation maybe just use a k means classification as a
prestep will be a good idea?

Thank you for your reply i'll spare the result as soon as i have try the
watershed segmentation.
Best regards,
Cyril



2015-07-31 12:22 GMT+02:00 Timothee Evain <tevain at telecom-paristech.fr>:

> So I guess your radioactive tracker is some kind of glucose that gonna
> yield a high signal on tumors?
> If so, maybe you could:
> - use this information to threshold your images to get a coarse
> segmentation of high signal structures
> - take this coarse segmentation as landmarks to try a watershed with
> markers :
> http://www.itk.org/Doxygen/html/classitk_1_1MorphologicalWatershedFromMarkersImageFilter.html
>
> Don't forget to add a marker for the background also.
>
> If you can't rely only on tumor signal, I suggest that you use the bladder
> as a hint to define a small region of interest around/under it to target
> the tumoral area.
>
> Hope this helps,
>
> Tim
>
> ----- Mail original -----
> De: "Cyril Jaudet" <drcjaudet at gmail.com>
> À: "Timothee Evain" <tevain at telecom-paristech.fr>
> Cc: community at itk.org
> Envoyé: Vendredi 31 Juillet 2015 11:44:11
> Objet: Re: [ITK] Segmentation of PET image
>
> Hello Timothee,
>
> I plan to segment rectal tumor. So it is a specific structure but it is
> near the bladder who have a high signal in PET. Here an image with the
> green label segmented with an adaptative (need calibration curve) contrast
> based method.
>
>
> [image: Images intégrées 1]
>
> Thank you,
> Cyril
>
> 2015-07-31 11:23 GMT+02:00 Timothee Evain <tevain at telecom-paristech.fr>:
>
> > Hello Cyril,
> >
> > What are you trying to segment ?
> > Do you expect a full image segmentation or are you targeting specific
> > structures?
> >
> > Tim
> >
> > ----- Mail original -----
> > De: "Cyril Jaudet" <drcjaudet at gmail.com>
> > À: community at itk.org
> > Envoyé: Vendredi 31 Juillet 2015 10:24:30
> > Objet: [ITK] Segmentation of PET image
> >
> > Hello ITk community,
> >
> > I am looking for a robust method to segment positron emissopn tomography
> > image in a multicenter study. The image are blurry and with a lot of
> noise.
> > One approach who seems reproductible but need a calibration of each PET
> is
> > a simple region growing algorithme.
> > One old itk approach seems very interesting by combining fuzzy connected
> > image filter and Voronoi segmentation image filter. However if i
> understand
> > well the fuzzy method were patented.
> > Do you know if there is a similar way for perfoming these kind of
> > segmentation?
> >
> > Thank's you,
> > Cyril Jaudet, PhD
> > UZ Brussel
> >
> > _______________________________________________
> > Community mailing list
> > Community at itk.org
> > http://public.kitware.com/mailman/listinfo/community
> >
>