[Insight-users] Anisotropic diffusion ( was RE: ShapeDetection)

[Joshua Cates]

Hi Christina,

The different varieties of anisotropic diffusion that we have implemented
are documented in the header files of the code and are also described on
the www.itk.org webpage under the Documentation, Manual Pages section. =20
You will find an overview of the mathematics involved and some references. =
=20

Take a look at the documentation for itkAnisotropicDiffusionFunction.h for
the most detailed info.  I apologize for the missing equations in the
online documentation, this was supposed to have been fixed some time ago.

Briefy, here is a breakdown of the different filters:

Curvature: Less blurring at regions of high curvature.  Better for
preserving smaller image features but a slower technique.

Gradient:  Preserves regions with higher gradient magnitude.  If you find
the curvature variant doesn't give the results you want, try this one.

The vector filters are just implementations of the above for vector-valued
data, rgb data, etc.  They require an images of itk::Vectors.  Sounds like
for MRI data you probably won't need these.

(AnisotropicDiffusionImageFilter is just a base class for the specific
types of diffusion and has no functionality in itself.  It cannot be
instantiated. )

For parameters, start with about 10-15 iterations of the filters using
Conductance parameters of 1.0 and TimeStep of 0.125.  You can lower the
time step value, but values much higher than 0.125 will not produce stable
solutions.

Let me know if you have any more questions or any suggestions,

Josh

______________________________
 Josh Cates=09=09=09
 School of Computer Science=09
 University of Utah
 Email: cates@cs.utah.edu
 Phone: (801) 587-7697
 URL:   www.cs.utk.edu/~cates


On Wed, 20 Mar 2002, Lydia Ng wrote:

> Can someone help Christina as to the differences
> in the differ favors of anisotropic diffusion?
>=20
> She is hoping to use the anisotropic diffusion to remove
> texture and enhance MR images of the liver.
>=20
> Cheers,
> Lydia
>=20
> > -----Original Message-----
> > From: Christina Lee [mailto:dreamie222@yahoo.com]
> > Sent: Wednesday, March 20, 2002 2:28 PM
> > To: Lydia Ng
> > Subject: RE: ShapeDetection
> >=20
> >=20
> >=20
> > Hi Lydia,
> >=20
> > How are you?
> > I think I'll try to preprocess the MR images with
> > anisotropic diffusion as you suggested. Although I'm
> > not exactly sure what kind of anistropic diffusion
> > filter that I need (there seems to be three
> > -curvature, gradient & vector!?)
> >=20
> > I couldn't find the source code (or the cxx file) for
> > itkAnisotropicDiffusionImageFilter.h & I did a
> > thorough search on the files that I downloaded from
> > the ITK website. Are there DLL or LIB files for the
> > derived classes that I'm not aware of?=20
> >=20
> > Hope to hear from you soon.
> >=20
> > Thank you so much for your help!
> >=20
> > Christina
> > --- Lydia Ng <lng@insightful.com> wrote:
> > > Hi Christina,
> > >=20
> > > The ShapeDetection program is not very sophisticated
> > > in
> > > that it only uses image gradient information to
> > > generate
> > > an edge potential map. Ideally, the edge potential
> > > map is
> > > zero at an "edge" and one in homogenous areas.
> > >=20
> > > A fastmarching algorithm is used to grow out
> > > a contour from a seed point. The rate of expansion
> > > of the=20
> > > contour is determined by the values in the edge
> > > potential map.=20
> > >=20
> > > I think the algorithm is working well for CT=20
> > > because the edges are nice and strong.
> > > I am not familiar with MR images of the liver, but
> > > if there are
> > > high gradient textures within the liver then these
> > > might=20
> > > interfere with the expansion of the contour.
> > >=20
> > > The other issue is that the computed edge potential
> > > map
> > > is far from ideal. For example, if the boundary
> > > gradient isn't
> > > strong, the associated edge potential will be
> > > some non-zero value and the contour will eventually
> > > leak=20
> > > across the boundary as time goes on.
> > >=20
> > > What kind of output are you getting when you look
> > > at it at various time crossing thresholds?
> > >=20
> > > I guess you can also try some preprocessing with
> > > something
> > > like anisotropic diffusion to get rid of some of the
> > > texture
> > > and enhance the boundaries.
> > >=20
> > > - Lydia
> > >=20
> > > > -----Original Message-----
> > > > From: Christina Lee [mailto:dreamie222@yahoo.com]
> > > > Sent: Wednesday, February 27, 2002 2:30 PM
> > > > To: Lydia Ng
> > > > Subject: ShapeDetection
> > > >=20
> > > >=20
> > > >=20
> > > > Hi Lydia,
> > > >=20
> > > > How are you?=20
> > > > I've been running the ShapeDetection program on
> > > some
> > > > MR images (to segment out the liver) & I haven't
> > > had
> > > > any luck. The program worked pretty well for the
> > > CT
> > > > data but not for the MR data - I wonder if it's
> > > > because the MR images are more textured. Any
> > > comments
> > > > or suggestions?
> > > >     =20
> > > > Hope to hear from you soon.
> > > >=20
> > > > Thanks,
> > > >=20
> > > > Christina
> > > >=20
> > > > __________________________________________________
> > > > Do You Yahoo!?
> > > > Yahoo! Greetings - Send FREE e-cards for every
> > > occasion!
> > > > http://greetings.yahoo.com
> > > >=20
> >=20
> >=20
> > __________________________________________________
> > Do You Yahoo!?
> > Yahoo! Movies - coverage of the 74th Academy Awards=AE
> > http://movies.yahoo.com/
> >=20
> _______________________________________________
> Insight-users mailing list
> Insight-users@public.kitware.com
> http://public.kitware.com/mailman/listinfo/insight-users
>=20