[Insight-users] Selecting the sigma for the vesselness

[Dan Mueller]

On 26/09/2007, Chris McIntosh <cmcintos.sfu at gmail.com> wrote:
>
>
> rashedk wrote:
> > However, I dont understand one thing: Why did the author of the Manniseng
> > VED's ITK implementation compare the VED filter responses to that of the
> > multiscale vesselness response in his ITK journal paper? Link here:
> > http://insight-journal.org/dspace/bitstream/1926/558/5/VEDArticle2.pdf
> >
>
> I can't speak for the authors, but I believe its because the VED filter was
> originally presented as an improvement over the original frangi filter.
>
> > Dont they do different things? VED enhances vessels whereas multiscale
> > vesselness detects vessels?
> > What I am trying to get at is, if VED does a better job than the standard
> > vesselness filter, then how can we use it to detect vessels only.

In my experience vessel enhancement is usually a preliminary step for
segmentation. Even though a filter may be designed for "enhancement",
the resultant image could possess qualities which simplifies or
improves segmentation. For example: it may be possible to apply a
simple region growing operation on a VED enhanced image to segment the
vessels, whereas applying the same operation on a non-enhanced image
would give poor results.

Anyway, just my thoughts.

Cheers, Dan Mueller

> Can you define more explicitly what you mean by detect? Do you want the
> vessel radius for each vessel,
> or a binary segmentation of the image into vessel and non-vessel, or an
> image showing just the vessels (
> the result of masking the image with a binary segmentation).
>
> Thanks,
>
> Chris
>
> > Thanks
> > Rashed karim
> >
> >
> >
> > Chris McIntosh-2 wrote:
> >
> >> Hello Rashed,
> >>
> >> It appears I misunderstood your goal. It seems you are trying to
> >> determine the vessel radius
> >> at each voxel, and not filter the image. Or perhaps segment the entire
> >> vessel...
> >>
> >> Yes your interpretation of the VED filter is correct. As is your
> >> interpretation of Frangi's approach.
> >> I thought there was a multi-scale version of that filter in ITK already,
> >> but perhaps I was wrong.
> >>
> >> The Hessian3DToVesselnessMeasuerImageFilter, however, is not Frangi's
> >> filter.
> >> http://www.itk.org/Doxygen34/html/classitk_1_1Hessian3DToVesselnessMeasureImageFilter.html
> >>
> >> "3D Multi-scale line filter for segmentation and visualization of
> >> curvilinear structures in medical images", Yoshinobu Sato, Shin
> >> Nakajima, Hideki Atsumi, Thomas Koller, Guido Gerig, Shigeyuki Yoshida,
> >> Ron Kikinis.
> >>
> >> is the filter that is used there.
> >>
> >> Cheers,
> >>
> >> Chris
> >>
> >>
> >>
> >> rashedk wrote:
> >>
> >>> Hi Chris,
> >>>
> >>> Doesn't Manniseng's VED filter *enhance* the vessels rather than
> >>> detecting
> >>> them? It does do multi-scale analysis, but goes one step further by
> >>> applying
> >>> the appropriate diffusion (iso- or aniso-) process. So the output of the
> >>> VED
> >>> filter should be the same image with the vessels enhanced and not the
> >>> vesselness. Please, correct me if I am wrong.
> >>>
> >>> What I also don't understand is why
> >>> Hessian3DToVesselnessMeasureImageFilter
> >>> does not implement a multiscale approach? If it is Frangi's vesselness
> >>> which
> >>> is implemented here, then his paper did suggest returning the highest
> >>> response by varying the sigma.
> >>>
> >>> Regards,
> >>> Rashed
> >>>
> >>>
> >>>
> >>> Chris McIntosh-2 wrote:
> >>>
> >>>
> >>>> Hello,
> >>>>
> >>>> I'm not sure if this helps, but just in case.
> >>>>
> >>>> In general, using a constant sigma for an image is troublesome because
> >>>> the ideal scale for each vessel
> >>>> is the actual radius of that vessel (in voxels). So you need a different
> >>>> scale at each voxel.
> >>>>
> >>>> I would suggest looking at:
> >>>> http://hdl.handle.net/1926/558
> >>>>
> >>>> They implement a powerful multi-scale technique for vessel filtering,
> >>>> where each pixel's response
> >>>> is tested under a range of scales and the maximal response is taken.
> >>>>
> >>>>
> >>>> Cheers,
> >>>>
> >>>> Chris
> >>>>
> >>>> rashedk wrote:
> >>>>
> >>>>
> >>>>> Hi everyone,
> >>>>>
> >>>>> I am currently using the vesselness filters to detect vessels. I use a
> >>>>> pipeline with HessianRecursiveGaussianImageFilter and
> >>>>> Hessian3DToVesselnessMeasureImageFilter filters which works fine.
> >>>>> However, I
> >>>>> am using a trial and error approach for selecting the sigma. In a
> >>>>> previous
> >>>>> post I came across an approach whereby sigma selection can be done
> >>>>> effectively by observing the responses of a single pixel that is known
> >>>>> to
> >>>>> be
> >>>>> part of a vessel (a.k.a multi-scale analysis). How do I constraint the
> >>>>> above
> >>>>> filters to only process a single pixel and not the entire image?
> >>>>>
> >>>>> Thanks,
> >>>>>
> >>>>> Rashed Karim