[Insight-users] discrete hessian recursive gaussian filter

[Bradley Lowekamp]

Hello Jack,

I used these filter heavily a couple of years ago and they seems to work quite well. I unfortunately didn't really write good tests for them so I can not say they are ready to go with 100% confidence. But I think they should work fine with ITK 4.1+, and with a little tweaking 3.20. I just did a couple of updates to the git module.

Regarding streaming. The DiscreteHessian is fully streamable. However, the filter you mention, the DiscreteHessianRecursiveGaussian is semi-streamable. It will require it's full input. Internally this filter is composed of a series of RecursiveGaussians and then the DiscreteHessian. The first Gaussian needs to operate on the whole image and outputs a image of double, from there the filters are more streamable, and eventually result in a Hessian of just the requested region. And of course these filters are multithreaded.

You say you image is 10GB, you don't say what the pixel type is. If it is only uint8, the an image of double will be 80GB, so this filter combination should work for you even in that case.

Hope this is helpful,
Brad

On Nov 15, 2012, at 1:49 PM, Jack Lee <jack.lee at kcl.ac.uk> wrote:

> Dear Brad,
> 
> This question is a follow up to the thread:
> http://www.itk.org/pipermail/insight-users/2011-March/040411.html
> 
> What I'm trying to do is similar to what Kana described, i.e. use a 
> local hessian filter on a large image (>10Gb) to perform a multiscale 
> vesselness calculation. I have access to a multicore machine with large 
> RAM (256Gb) so ideally the filter would be both multithreaded and 
> streamable. Even with this RAM, processing the whole stack in one go is 
> probably impossible.
> 
> I'd appreciate it very much if you could update us on the latest 
> developments on your itkDiscreteHessianRecursiveGaussianImage filter. 
> Also will it strictly require the latest itk library?
> 
> 
> Many thanks,
> Jack