[Insight-users] Anisotropic Diffusion filtering
luis.ibanez at kitware.com
Sun Aug 5 10:26:30 EDT 2007
1) Could you post somewhere in a public web site
a slice of one of your 3D images ?
2) The numerical values recommended for the parameters
of this filter, are for a typical MRI image.
If your images happen to have a very different
appearance then you need to explore different settings
for the parameters.
I would strongly suggest you to do this first in a single
slice of the 3D image (use the ExtractImageFilter), since
it will be a lot faster to try many different combinations
of the filter parameters.
3) 50 iterations, should be more than enough for producing
a visible effect in the images... unless your images are
What is the modality of your images ? (CT? MRI? Ultrasound ?)
What is the content of the image ? brain?, liver?
4) The time step should be small. The critical values are in the
order of 0.625. Whenever you go higher than that the filter
becomes numerically unstable. You should probably stay below
half that value, just to be safe.
Note however, that there is a relationship between the time
step and the number of iterations. After all, this filter is
simulating the melting of the material in your image as if it
were a physical problem. This is a discretization of a continuous
physical model. If you use half of the time steps, you will need
double the iterations in order to cover the same process.
That being said, this is just a simile, there is no real melting
happening here, so you simply use the physical model as an analogy
for smoothing the image.
5) A Gaussian smoothing of 20-30 is *HUGE*.
What specific filter did you use for the smoothing ?
What is the pixel spacing of your images ?
Please give us more information,
rashed karim wrote:
> Hi everyone,
> I am trying to use the itkGradientAnisotropicDiffusionImageFilter class
> for smoothing my MRI images. To my dismay, my images are not responding
> to the filter even when I run 50 or so iterations with a conductance of
> about 3.0 (i.e. output image is the same as the input image with no
> visible difference). An earlier forum post had suggested to use a 3.0
> conductance level, but such a high conductance rate makes the
> output image unrecognizable (lots of salt-pepper noise).
> So has anyone used this filter before, and if so, what sort of
> parameter values (number of iterations, conductance and time step)
> should I be looking to use. Although I know that these values depend
> on the type of my images I am working with, but any suggestions will be
> very useful.
> Also in the documentation it says somewhere that the time step parameter
> is similar to the width of the gaussian kernel. I have tried values of
> 20.0 - 30.0. However, I get a warning message telling me that this sort
> of value is an "unstable time step" for my images. The message also
> suggests a timestep which is usually a very small value (~ 0.05). What i
> dont understand, is that I had no problems (isotropic)
> Gaussian smoothing my images with widths ranging from 20.0 - 30.0, then
> why does the anisotropic filter complain for such widths?
> Here is part of my code if it helps:
> typedef itk::Image< PixelType, 3 > InputImageType;
> typedef itk::Image< PixelType, 3 > OutputImageType;
> typedef itk::ImageFileReader< InputImageType > ReaderType;
> typedef itk::ImageFileWriter< OutputImageType > WriterType;
> typedef itk::GiplImageIO ImageIOType;
> typedef itk::GradientAnisotropicDiffusionImageFilter<InputImageType,
> OutputImageType > FilterType;
> ReaderType::Pointer reader = ReaderType::New();
> WriterType::Pointer writer = WriterType::New();
> ImageIOType::Pointer giplIO = ImageIOType::New();
> FilterType::Pointer filter = FilterType::New();
> reader->SetFileName( argv );
> writer->SetFileName( argv );
> // for GIPL files
> filter->SetNumberOfIterations( numberOfIterations );
> filter->SetTimeStep( timeStep );
> filter->SetConductanceParameter( conductance );
> Thanks for any suggestions,
> Rashed Karim.
> Insight-users mailing list
> Insight-users at itk.org
More information about the Insight-users