[ITK-users] Splitting and repasting a volume with overlap

Bradley Lowekamp blowekamp at mail.nih.gov
Wed Mar 18 09:14:53 EDT 2015 

Hello Pol,

I believe I have already suggested reading the latest ITK Software Guide, and the section "How To Write a Filter" which details how the pipeline works. This will cover the GenerateInputRequestedRegion method and how it interacts with the Pipeline and it critical for understanding the pipeline. Your case really does sound like the ITK pipeline, and would benefit as being a filter. You described your output as "connected components", I am not sure if that is a image or another reduced structure. Does the output represent the whole thing in a compact fashion or does it represent a small part which can also be streamed?

The may add more complications to your situation but it may be the right fit. I have written a external module "itkStreamingSinc"[1], which provides a base class for algorithms which are both streamed and threaded. I have used it for things like computing bounding boxes and image statistics on images which are hundreds of gigabytes in an out of core fashion. I have also wanted to write one which produces a an itk RLE LabelObject. It works good for algorithms that can consume a chunk and produce a reduced structure.

Hope that helps,
Brad

[1] https://github.com/blowekamp/itkStreamingSinc



On Mar 18, 2015, at 8:58 AM, Pol Monsó Purtí <lluna.nova at gmail.com> wrote:

> Hello matt,
> 
> I wasn't aware of that possibility, thanks for the heads up!It doesn't apply to our case becaues I haven't created a filter but rather a class that takes a region and gives me the splitted regions that later on I feed to an extract filter (or the copy algorithm Bradley pointed out to repaste the results back). Creating a filter doesn't really fit well, the 'low-level' is based on machine learning so it trains and then predicts, and it requires several minutes depending on the size of the image etc. I didn't judge appropriate to create a Filter out of it. 
> 
> Could you elaborate a little bit on what you were proposing? Let's say I have a set of preprocessing filters, my processing and a set of postprocessing filters. The output would be a set of connected components. The volume doesn't necessarily fit in memory (specially the processing) so we have to stream. Because the chunks need to overlap, are you saying that I could place,  in the pipeline, a Filter that is derived from the StreamingImageFilter that re-implements GeneratInputRequestedRegion() to have an overlap of 30 voxels, and that would should do it? where should I place such a filter and how does interact with the pre/post-processing?
> 
> I will re-read the article on streaming, but maybe it would help me if you pinpoint the particularities of the approach you suggested.
> 
> Cheers,
> 
> Pol
> 
> 
> 
> 
> 2015-03-17 20:22 GMT+01:00 Matt McCormick <matt.mccormick at kitware.com>:
> Hi Pol,
> 
> You may already be aware of this, but the chunks will overlap if you
> implement a GenerateInputRequestedRegion() method on your filter that
> pads the output RequestedRegion.
> 
> HTH,
> Matt
> 
> On Tue, Mar 17, 2015 at 1:13 PM, Pol Monsó Purtí <lluna.nova at gmail.com> wrote:
> > Hello Matt,
> >
> > As I've told Bradley, I can't use streaming as-is because I need overlap
> > between chunks and because the low-level library would need a proper itk
> > image with correct size, region, etc.
> > which I guess it wouldn't be provided by streaming.
> >
> > But most definetely, right now I use 'streaming' by using the ExtractFilter
> > at the end. I believe that streams too. I was actually more interested in
> > having a filter/object that computes the splits for me. I've implemented it,
> > but it's far from perfect. For example, it broke when the input image had a
> > non-zero index. (It's fixed now, but still, I'm sure the itk's Splitter
> > filter does a better job, if only could I have overlap ;))
> >
> > I guess the way to go would be to take the splitter (probably
> > ImageRegionSplitterMultidimensional now that I look at it) and extend it to
> > support overlap. If I do it, I'll let you know in case it is of interest to
> > anyone. For now I'll stick to my class.
> >
> > Cheers,
> >
> > Pol
> >
> > 2015-03-16 16:05 GMT+01:00 Matt McCormick <matt.mccormick at kitware.com>:
> >>
> >> Hi Pol,
> >>
> >> As a side note, limitation of memory usage is built into the streaming
> >> pipeline design of ITK.  All computation of appropriate regions should
> >> occur if the filters are implemented correctly.  To take advantage of
> >> streaming, place the StreamingImageFilter [1] at the end of your
> >> pipeline. To use less memory, set the ReleaseDataFlagOn() [2] on the
> >> component filters.
> >>
> >> HTH,
> >> Matt
> >>
> >> [1] http://www.itk.org/Doxygen/html/classitk_1_1StreamingImageFilter.html
> >>
> >> [2]
> >> http://www.itk.org/Doxygen/html/classitk_1_1ProcessObject.html#ae732e9163879ae559ccd0b5957141f97
> >>
> >> On Mon, Mar 16, 2015 at 10:20 AM, Bradley Lowekamp
> >> <blowekamp at mail.nih.gov> wrote:
> >> > You may find this method useful:
> >> >
> >> >
> >> > http://www.itk.org/Doxygen/html/structitk_1_1ImageAlgorithm.html#a5ec5bcac992cb3b1302840636530ba20
> >> >
> >> > Brad
> >> >
> >> > On Mar 16, 2015, at 10:16 AM, Pol Monsó Purtí <lluna.nova at gmail.com>
> >> > wrote:
> >> >
> >> > Mmh, I believe I misread the RegionOfInterestImageFilter documentation,
> >> > confusing dimension with size. That should work.
> >> >
> >> > I still have the computation of the splitting regions problem to handle,
> >> > but
> >> > maybe with this filter is easier than with the ExtractImageFilter. Or
> >> > maybe
> >> > it just removes step 7?
> >> >
> >> > Is there a way to ease the repasting back to the original region with
> >> > it?
> >> > How should I use the physical space index?
> >> >
> >> > 2015-03-16 14:25 GMT+01:00 Pol Monsó Purtí <lluna.nova at gmail.com>:
> >> >>
> >> >> Hello Bradley,
> >> >>
> >> >> Thank you for your answer and suggestion.
> >> >>
> >> >> Answering your question: It's not for multi-threading purposes, I have
> >> >> to
> >> >> split the image due to memory limitations.
> >> >>
> >> >> I would use extractImage directly and do streaming, but the client and
> >> >> the
> >> >> low-level libraries might not work correctly if I do that. I have to
> >> >> provide
> >> >> itk images disconnected from the pipeline and with proper indexes and
> >> >> sizes.
> >> >>
> >> >> The RegionOfInterestImageFilter would have the same memory footprint,
> >> >> and
> >> >> it would still be tricky to compute the regions of interest, due to the
> >> >> overlapping.
> >> >>
> >> >>
> >> >>
> >> >> 2015-03-16 14:00 GMT+01:00 Bradley Lowekamp <blowekamp at mail.nih.gov>:
> >> >>>
> >> >>> Hello,
> >> >>>
> >> >>> The RegionOfInterestImageFilter[1] is very similar to the
> >> >>> ExtractImageFilter, but returns an image whose starting index is
> >> >>> always
> >> >>> zeros.
> >> >>>
> >> >>> The process you described is remarkable close to what occur in each
> >> >>> filter for the multi-threading process.  Also what occurs in the
> >> >>> StreamingImageFilter is very similar, but just copies the region.
> >> >>>
> >> >>> As you didn't mention you motivation for chunking the data this way be
> >> >>> it
> >> >>> for threading or memory requirements, I can't fully advise.
> >> >>>
> >> >>> Brad
> >> >>>
> >> >>>
> >> >>> [1
> >> >>>
> >> >>> ]http://www.itk.org/Doxygen/html/classitk_1_1RegionOfInterestImageFilter.html
> >> >>> [2]
> >> >>>
> >> >>> https://github.com/InsightSoftwareConsortium/ITK/blob/b184194869ac6407fcbb4e309710894cfc84466f/Modules/Core/Common/include/itkStreamingImageFilter.hxx#L194-L213
> >> >>>
> >> >>>
> >> >>> On Mar 16, 2015, at 7:37 AM, Pol Monsó Purtí <lluna.nova at gmail.com>
> >> >>> wrote:
> >> >>>
> >> >>> > Hello everyone,
> >> >>> >
> >> >>> > We need to split a volume into pieces, overlaped by a certain
> >> >>> > amount,
> >> >>> > do some computation and then repaste them.
> >> >>> >
> >> >>> > The computation requires the pieces' region indexes to be set to
> >> >>> > zero.
> >> >>> > The original volume might not have index at zero, but be a subregion
> >> >>> > itself.
> >> >>> >
> >> >>> > What is the best way to achieve this?
> >> >>> >
> >> >>> > I've thought of the following pipeline:
> >> >>> >
> >> >>> > 1. imageRegionSplitterSlowDimension
> >> >>> > 2. Store the index and size of each region
> >> >>> > 3. manually increase the size of each region (and modify index)
> >> >>> > 4. crop with the largestPossibleRegion of the original volume (to
> >> >>> > prevent requesting more than available)
> >> >>> > iterate over regions:
> >> >>> >       5. extractImageFilter
> >> >>> >       6. disconnect pipeline for each region
> >> >>> >       7. reset indexes
> >> >>> >       8. perform computation
> >> >>> >       9. crop with the regions in step 1 to remove the added overlap
> >> >>> >       10. restore indexes
> >> >>> > 11. pasteImageFilter
> >> >>> >
> >> >>> > I'll have to think of a smart way to deal with the index
> >> >>> > restore/cropping of steps 9 and 10.
> >> >>> >
> >> >>> > I somehow feel somebody will have already dealt with this... And
> >> >>> > there's certainly a better way to do it.
> >> >>> > _____________________________________
> >> >>> > Powered by www.kitware.com
> >> >>> >
> >> >>> > Visit other Kitware open-source projects at
> >> >>> > http://www.kitware.com/opensource/opensource.html
> >> >>> >
> >> >>> > Kitware offers ITK Training Courses, for more information visit:
> >> >>> > http://www.kitware.com/products/protraining.php
> >> >>> >
> >> >>> > Please keep messages on-topic and check the ITK FAQ at:
> >> >>> > http://www.itk.org/Wiki/ITK_FAQ
> >> >>> >
> >> >>> > Follow this link to subscribe/unsubscribe:
> >> >>> > http://public.kitware.com/mailman/listinfo/insight-users
> >> >>>
> >> >>
> >> >>
> >> >> 2015-03-16 14:00 GMT+01:00 Bradley Lowekamp <blowekamp at mail.nih.gov>:
> >> >>>
> >> >>> Hello,
> >> >>>
> >> >>> The RegionOfInterestImageFilter[1] is very similar to the
> >> >>> ExtractImageFilter, but returns an image whose starting index is
> >> >>> always
> >> >>> zeros.
> >> >>>
> >> >>> The process you described is remarkable close to what occur in each
> >> >>> filter for the multi-threading process.  Also what occurs in the
> >> >>> StreamingImageFilter is very similar, but just copies the region.
> >> >>>
> >> >>> As you didn't mention you motivation for chunking the data this way be
> >> >>> it
> >> >>> for threading or memory requirements, I can't fully advise.
> >> >>>
> >> >>> Brad
> >> >>>
> >> >>>
> >> >>> [1
> >> >>>
> >> >>> ]http://www.itk.org/Doxygen/html/classitk_1_1RegionOfInterestImageFilter.html
> >> >>> [2]
> >> >>>
> >> >>> https://github.com/InsightSoftwareConsortium/ITK/blob/b184194869ac6407fcbb4e309710894cfc84466f/Modules/Core/Common/include/itkStreamingImageFilter.hxx#L194-L213
> >> >>>
> >> >>>
> >> >>> On Mar 16, 2015, at 7:37 AM, Pol Monsó Purtí <lluna.nova at gmail.com>
> >> >>> wrote:
> >> >>>
> >> >>> > Hello everyone,
> >> >>> >
> >> >>> > We need to split a volume into pieces, overlaped by a certain
> >> >>> > amount,
> >> >>> > do some computation and then repaste them.
> >> >>> >
> >> >>> > The computation requires the pieces' region indexes to be set to
> >> >>> > zero.
> >> >>> > The original volume might not have index at zero, but be a subregion
> >> >>> > itself.
> >> >>> >
> >> >>> > What is the best way to achieve this?
> >> >>> >
> >> >>> > I've thought of the following pipeline:
> >> >>> >
> >> >>> > 1. imageRegionSplitterSlowDimension
> >> >>> > 2. Store the index and size of each region
> >> >>> > 3. manually increase the size of each region (and modify index)
> >> >>> > 4. crop with the largestPossibleRegion of the original volume (to
> >> >>> > prevent requesting more than available)
> >> >>> > iterate over regions:
> >> >>> >       5. extractImageFilter
> >> >>> >       6. disconnect pipeline for each region
> >> >>> >       7. reset indexes
> >> >>> >       8. perform computation
> >> >>> >       9. crop with the regions in step 1 to remove the added overlap
> >> >>> >       10. restore indexes
> >> >>> > 11. pasteImageFilter
> >> >>> >
> >> >>> > I'll have to think of a smart way to deal with the index
> >> >>> > restore/cropping of steps 9 and 10.
> >> >>> >
> >> >>> > I somehow feel somebody will have already dealt with this... And
> >> >>> > there's certainly a better way to do it.
> >> >>> > _____________________________________
> >> >>> > Powered by www.kitware.com
> >> >>> >
> >> >>> > Visit other Kitware open-source projects at
> >> >>> > http://www.kitware.com/opensource/opensource.html
> >> >>> >
> >> >>> > Kitware offers ITK Training Courses, for more information visit:
> >> >>> > http://www.kitware.com/products/protraining.php
> >> >>> >
> >> >>> > Please keep messages on-topic and check the ITK FAQ at:
> >> >>> > http://www.itk.org/Wiki/ITK_FAQ
> >> >>> >
> >> >>> > Follow this link to subscribe/unsubscribe:
> >> >>> > http://public.kitware.com/mailman/listinfo/insight-users
> >> >>>
> >> >>
> >> >
> >> >
> >> >
> >> > _____________________________________
> >> > Powered by www.kitware.com
> >> >
> >> > Visit other Kitware open-source projects at
> >> > http://www.kitware.com/opensource/opensource.html
> >> >
> >> > Kitware offers ITK Training Courses, for more information visit:
> >> > http://www.kitware.com/products/protraining.php
> >> >
> >> > Please keep messages on-topic and check the ITK FAQ at:
> >> > http://www.itk.org/Wiki/ITK_FAQ
> >> >
> >> > Follow this link to subscribe/unsubscribe:
> >> > http://public.kitware.com/mailman/listinfo/insight-users
> >> >
> >
> >
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://public.kitware.com/pipermail/insight-users/attachments/20150318/61ab0640/attachment.html>

More information about the Insight-users mailing list