Re: [vmtk-users] dealing with large data

[Luca A. <an...@ma...>]

Hi Joong-Ho,
  let's say there are several levels at which the problem of large  
files can be tackled.
One thing you must be aware of is that level sets take a lot of  
memory, mainly for storing
the original image, the gradient magnitude image, and the gradient  
image (for the advection term).
Just these take up 4x the memory of the original image. There's not a  
whole lot you can do about this.
However, you managed to segment the first chunk of data, which is  
promising.
The second thing is that colliding fronts takes even more memory  
during the initialization, so expect
your memory usage to go up. For your information, fast marching takes  
than colliding fronts, thresholding
less than fast marching and isosurface less than thresholding.
Another suggestion is on the system: free up space on the hard drive  
for hosting your virtual memory file.
It's better if your PC starts swapping than just killing the  
application because memory is full.

Now, on to the solutions:
It is definetly possible to segment in chunks and then merge after  
the fact. Or you could segment on lower
resolution images and then optimize non interactively on full  
resolution. Let's explore the various possibilities.

1. Segmenting one vessel at a time on the full image and merging the  
results:
Use vmtklevelsetsegmentation as usual. Then merge the resulting  
levelsets files with
vmtkimagecompose -ifile foo_ls1.vti -i2file foo_ls2.vti -operation  
min -ofile foo_ls12.vti
Eventually, you can do it in sequence
vmtkimagecompose -ifile foo_ls1.vti -i2file foo_ls2.vti -operation  
min --pipe vmtkimagecompose -i2file foo_ls3.vti -ofile foo_ls123.vti
and so on.

2. Generating small VOIs, segmenting all vessels in each VOI and  
merging the results:
Use vmtklevelsetsegmentation as usual on each small VOI. I suggest to  
generate VOIs that have at least a small overlap,
so the segmented vessels will merge seamlessly.
Once you have your segmentations on the small VOIs, you have to  
sample them on a larger canvas and then merge them
as in the previous point.
vmtkimagereslice -ifile foo_ls1.vti -rfile originalimage.vti - 
background -4.0 -ofile foo_ls1re.vti
Here original image is the full image: it's just used to provide the  
origin/spacing/extent information for the resampling, i.e. the
size and position of the canvas.
Do this for every image, then merge all re images as in 1.

3. Segmenting the full image at low resolution and optimizing the  
result on high resolution non-interactively:
this is the solution I suggest the most. Probably your acquired image  
has a resolution that goes well beyond what you actually
need to depict the aorta and the iliacs. Of course the more  
resolution the marrier, but you could use a lower resolution image
to perform your segmentation quickly and interactively, without  
bringing down the system in the process. Then resample the
ls file to full resolution and use it as initiallevelsets for the  
final non-interactive optimization of your segmentation on the full
resolution image. Let's start by resampling the image:
vmtkimagereslice -ifile originalimage.vti -spacing 1.0 1.0 1.0 -ofile  
resampledimage.vti
Then run vmtklevelsetsegmentation on resampledimage.vti (here I set  
isotropic resolution of 1.0mm, but you can of course choose
otherwise).
Once you're done and you've got your foo_ls.vti levelsets file,  
resample it at full resolution
vmtkimagereslice -ifile foo_ls.vti -rfile originalimage.vti -ofile  
foo_lsre.vti

and, last, run vmtklevelsetsegmentation

vmtklevelsetsegmentation -ifile originalimage.vti - 
initiallevelsetsfile foo_lsre.vti -iterations 300 -advectionweight  
1.0 -ofile foo_ls.vti
In this way, no rendering window will come up. Level sets will evolve  
with the iterations and weights you specify at command line,
and will be initialized with foo_lsre.vti

I hope you can find the right solution for your needs among the three  
of these.

Best

Luca


--
Luca Antiga, PhD
Biomedical Technologies Laboratory,
Bioengineering Department,
Mario Negri Institute
email: an...@ma...
web: http://villacamozzi.marionegri.it/~luca
mail: Villa Camozzi, 24020, Ranica (BG), Italy
phone: +39 035 4535-381



On Feb 3, 2007, at 2:10 AM, Joong-Ho Won wrote:

> Hi Luca,
>
> Thanks for the notice. I will use 3.0.1 next time I compile VMTK (hope
> it's not too soon :) ).
>
> Now I'm working on real datasets. They are scans of abdominal aorta  
> and
> iliac arteries. Each dataset is made of 600-1000+ dicom files.
>
> As you may have guessed, there is a memory problem. vmtkimagereader
> couldn't read more than 495 slices. So I dropped some chunks and was
> able to load them. VOI selector worked well, and the next task is  
> level
> set segmentation.
>
> It seem to be working, but after segmenting the first branch, when I
> tried initial segmentation, VMTK reported out of memory and nothing
> happened. That same thing happened even though I save the first
> segmentation and continued segmentation using -levelsetsfile option.
>
> My desktop has 2GB of memory, and this is the physical limit of the
> hardware. Maybe the only solution is to split them into several parts
> and merge the results.
>
> So I wonder if it is possible:
>
> 1. merge segmentation files off-line, that is, outside of
> vmtklevelsetsegmentation.
>
> 2. merge centerlines
>
> It seems 2 should be enough for my purpose, but when I tested with
> splitting, the origins didn't match. It appears that  
> vmtkimagereader has
> -origin option, but they are not for dicom format.
>
> Could you figure out a solution? Thanks.
>
> Regards,
>
>     Joong-Ho.
>
>
>
> Luca Antiga wrote:
>> Hi Joong-Ho,
>>  after checking with Luis Ibanez, I found out that the problem has  
>> been
>> fixed in the current ITK stable version (3.0.1)
>> I compile it and I confirm that the problem is gone.
>> Ciao
>>
>> Luca
>>
>
> ---------------------------------------------------------------------- 
> ---
> Using Tomcat but need to do more? Need to support web services,  
> security?
> Get stuff done quickly with pre-integrated technology to make your  
> job easier.
> Download IBM WebSphere Application Server v.1.0.1 based on Apache  
> Geronimo
> http://sel.as-us.falkag.net/sel? 
> cmd=lnk&kid=120709&bid=263057&dat=121642
> _______________________________________________
> vmtk-users mailing list
> vmt...@li...
> https://lists.sourceforge.net/lists/listinfo/vmtk-users