[Insight-users] Connected component statistics

Gaetan Lehmann gaetan.lehmann at jouy.inra.fr
Mon Jul 25 03:06:37 EDT 2005 

Hi Bill,

It's seems that we have done quite the same work last week. I saw this  
weekend that you have implemented region and bounding box in  
LabelStatisticsImageFilter. I have also implemented bounding box (and  
center of gravity), but in a new filter. It's funny to see how people can  
need the same thing at the same time :-)

The new filter is based on LabelStatisticsImageFilter but I decided to  
create a new one because the computed values don't require pixel value,  
but only label, and so the filter take only one image as input instead of  
2 for LabelStatisticsImageFilter.

I think to work on perimeter/area this week, so it would be great to  
decide where it have to be implemented.

Regards,

Gaetan


On Fri, 22 Jul 2005 13:18:58 +0200, Gaetan Lehmann  
<gaetan.lehmann at jouy.inra.fr> wrote:

>
> Hi everyone,
>
> Here is the first work I have done on this filter. It's mainly a
> modification of LabelStatisticsImageFilter.
> Comment would be highly appreciated, especially for :
>   + Name of method/class
>   + Choice of class or type to represent data
>   + GetLabels method (not present in LabelStatisticsImageFilter)
>   + computation of center of gravity in pysical units
>
> I will work on computation of perimeter/area next week
>
> Regards,
>
> Gaetan
>
> On Wed, 20 Jul 2005 13:47:18 +0200, Miller, James V (Research)
> <millerjv at crd.ge.com> wrote:
>
>> A map should be used for the internal implementation.  Like you
>> suggested,
>> however, it is important to provide access methods (or queries) into the
>> map.
>> One reason this is particularly important is in MSVC++ you cannot pass a
>> reference
>> or iterators to map across DLL boundaries.
>>
>> Jim
>>
>> -----Original Message-----
>> From: Gaetan Lehmann [mailto:gaetan.lehmann at jouy.inra.fr]
>> Sent: Wednesday, July 20, 2005 3:01 AM
>> To: Richard.Beare at ieee.org; Sayan Pathak
>> Cc: insight-users at itk.org; Miller, James V (Research);
>> Karthik.Krishnan at kitware.com
>> Subject: Re: [Insight-users] Connected component statistics
>>
>>
>>
>> Hi,
>>
>> even if a map is a good thing for c++ developers, I'm not sure it's
>> really
>> usable for python (and java, and tcl) ones.
>> Or the filter must provide methods to access content of the map without
>> using the map directly. Something like methods in
>> LabelStatisticsImageFilter :
>>   RealType 	GetMinimum (LabelPixelType label) const	
>>   RealType 	GetMaximum (LabelPixelType label) const	
>>   RealType 	GetMean (LabelPixelType label) const	
>>   RealType 	GetMedian (LabelPixelType label) const	
>>   RealType 	GetSigma (LabelPixelType label) const	
>>   RealType 	GetVariance (LabelPixelType label) const	
>>   RealType 	GetSum (LabelPixelType label) const	
>>   unsigned long 	GetCount (LabelPixelType label) const	
>>
>> it would be great also to choose consistent name : in other filters, we
>> have the word "component" and the word "object" for the same thing.
>> Component seems better to me.
>>
>> Richard, are you working on this filter ?
>> I don't think we need to do the work twice :-)
>>
>> Gaetan
>>
>> On Wed, 20 Jul 2005 01:14:40 +0200, Richard Beare
>> <richard.beare at gmail.com> wrote:
>>
>>> Hi,
>>> I guess that the ordering can be made somewhat arbitary.
>>>
>>> Here's a summary of how the equivalent functions I used in my previous
>>> job:
>>>
>>> step 1 was labelling. This produced an integer image with each
>>> connected component having a separate label. Labels were consecutive
>>> and assigned in raster order.
>>>
>>> step 2 was computing binary statistics, given this label image. This
>>> was a single pass process that produced a table of results. Each row
>>> in the table corresponded to a connected component and included the
>>> label. The label was somewhat redundant until you started to
>>> manipulate the table.
>>>
>>> Subsequent steps were somewhat arbitary - they may include filtering
>>> objects based on combinations of these attributes to produce another
>>> label image, or colouring the regions according to some parameter
>>> (like area) or any number of other things.
>>>
>>> If the aim is to minimize the number of passes through the image then
>>> relabelling according to ordering of connected component pixels should
>>> be left as an option to applications that need it - perhaps some sort
>>> of substitution filter could be developed for that purpose.
>>>
>>> Starting on this might be a good introduction to filter writing for me.
>>>
>>> On 7/20/05, Sayan Pathak <SayanP at alleninstitute.org> wrote:
>>>>
>>>>
>>>> Hi Jim/Richard,
>>>> I like the idea of using
>>>> std::map<itk::CommponentStatistics> type storage. The first
>>>> entry of the map (the Key) could be the component label index.  It
>>>> would
>>>> probably be desirable to assign label for each of the connected
>>>> components
>>>> in descending order of their size (as it is done in the current
>>>> filter).
>>>>
>>>> Sayan
>>>>
>>>> > Date: Tue, 19 Jul 2005 09:18:11 -0400
>>>> > From: "Miller, James V (Research)" <millerjv at crd.ge.com>
>>>> > Subject: RE: [Insight-users] Connected component statistics
>>>> > To: <Richard.Beare at ieee.org>
>>>> > Cc: insight-users at itk.org, Karthik Krishnan
>>>> >       <Karthik.Krishnan at kitware.com>
>>>> > Message-ID:
>>>> >
>>>> <FA26BEF1EA775E4584FB34B91E14A1C403F34F82 at SCHMLVEM01.e2k.ad.ge.com>
>>>> > Content-Type: text/plain;     charset="iso-8859-1"
>>>>
>>>> >
>>>> > Richard,
>>>> >
>>>> > A one pass algorithm is what I would want ITK to have to compute
>>>> these
>>>> > statististics for all components.  Glad to hear you were already
>>>> thinking
>>>> > along those lines.
>>>> >
>>>> > As for a representation, I would start with an approach like the
>>>> RelabelComponentImageFilter
>>>> > and LabelStatisticsImageFilter.  After these filters run, you can
>>>> query
>>>> the filter for
>>>> > the statistics of interest for a given component.
>>>> >
>>>> > If this information needs to flow through the pipeline, then a
>>>> separate
>>>> data structure,
>>>> > say a std::map<itk::ComponentStatistics> could be used
>>>> and wrapped with a DataObjectDecorator
>>>> > to add a DataObject API to the std::map.  We would still have to
>>>> create an
>>>> itk::ComponentStatistics
>>>> > class (or some other named class) to store the statistics for a  
>>>> given
>>>> component.
>>>>
>>>>
>>>>
>>>> > Jim
>>>>
>>>> > > -----Original Message-----
>>>> > > From: Richard Beare [mailto:richard.beare at gmail.com]
>>>> > > Sent: Monday, July 18, 2005 7:37 PM
>>>> > > To: Miller, James V (Research)
>>>> > > Cc: Gaetan Lehmann; Karthik Krishnan; insight-users at itk.org
>>>> > > Subject: Re: [Insight-users] Connected component statistics
>>>> > >
>>>> > >
>>>> > > Hi everyone,
>>>> > >
>>>> > > It seems that there are a few starting points around already. The
>>>> > > measures we computed in my previous job could all be done in a
>>>> single
>>>> > > pass and consolidated at the end. I think that most of them have
>>>> > > already been covered in the discussion.
>>>> > >
>>>> > > The only thing I'm not sure about now is an ITK specific question  
>>>> -
>>>> > > how should the statistics be represented. The previous function of
>>>> > > this type that I used produced a table of statistics as output,
>>>> with
>>>> > > one row for each connected component. This table was easily
>>>> > > represented as an image. I suspect that this approach wouldn't be
>>>> > > appropriate in ITK, but I'm not sure what the right mechanism is,
>>>> > > since I'm still learning my way around.
>>>>
>>>> On 7/18/05, Miller, James V (Research) <millerjv at crd.ge.com> wrote:
>>>> > Gaeton,
>>>> >
>>>> > Center of gravity and bounding box could be added to the
>>>> RelabelStatisticsImageFilter
>>>> > since they are simple calculations.  However, it is a bit of a
>>>> side-effect.  The
>>>> > volume is calculated in the RelabelStatisticsImageFilter because it
>>>> is
>>>> used by the
>>>> > filter to order the labels.  Since it was already calculated, it  
>>>> made
>>>> sense to allow
>>>> > the user to access it.
>>>> >
>>>> > The best place for these measures and the more complicated measures
>>>> would
>>>> be a class
>>>> > similar to the LabelStatisticsImageFilter. The filter should be
>>>> called
>>>> something like
>>>> > LabelMomentsImageFilter (and calculate ImageMoments as well as
>>>> geometric
>>>> moments. Or maybe
>>>> > something like LabelShapeImageFilter.
>>>> >
>>>> > I do not see any reason to use the ImageSpatialMaskObject for this
>>>> task.
>>>> Usually, you
>>>> > would want to calculate these features for EVERY component.  So it
>>>> best to
>>>> run through a
>>>> > filter that does it.
>>>> >
>>>> > Jim
>>>> >
>>>> > -----Original Message-----
>>>> > From: Gaetan Lehmann [mailto:gaetan.lehmann at jouy.inra.fr]
>>>> > Sent: Monday, July 18, 2005 4:57 AM
>>>> > To: Miller, James V (Research); Karthik Krishnan;
>>>> Richard.Beare at ieee.org
>>>> > Cc: insight-users at itk.org
>>>> > Subject: Re: [Insight-users] Connected component statistics
>>>> >
>>>> >
>>>> >
>>>> > Hi,
>>>> >
>>>> > I'm also interested in basic measures such as volume, center of
>>>> gravity
>>>> > and bounding box.
>>>> > I think to work on it this week, so it should be great to define
>>>> where to
>>>> > implement methods to get those informations.
>>>> >
>>>> > RelabelComponentImageFilter is already able to give volume, so it
>>>> should
>>>> > be a good place to add other very simple informations like center of
>>>> > gravity and bounding box.
>>>> > It don't seem to be the good place for more complex info such as
>>>> > perimeter/surface... perhap's ImageMaskSpatialObject is a better
>>>> place ?
>>>> >
>>>> > WDYT ?
>>>> >
>>>> > Gaetan
>>>> >
>>>> > On Fri, 15 Jul 2005 20:39:46 +0200, Miller, James V (Research)
>>>> > <millerjv at crd.ge.com> wrote:
>>>> >
>>>> > > The LabelStatisticsImageFilter will calculate statistics for each
>>>> > > component.
>>>> > > Currently, the LabelStatisticsImageFilter calculate mean,  
>>>> variance,
>>>> sum,
>>>> > > count,
>>>> > > and histograms.
>>>> > >
>>>> > > We plan to add another filter that calculated the moments (either
>>>> image
>>>> > > moments or
>>>> > > geometric moments) of each component. I have been waiting for it  
>>>> to
>>>> be
>>>> > > completed and
>>>> > > may have to just write it myself.
>>>> > >
>>>> > > Also, the RelabelComponentImageFilter allows you to query the
>>>> volume of
>>>> > > a component.
>>>> > >
>>>> > > Jim
>>>> > >
>>>> > > -----Original Message-----
>>>> > > From: insight-users-bounces+millerjv=crd.ge.com at itk.org
>>>> > >
>>>> [mailto:insight-users-bounces+millerjv=crd.ge.com at itk.org]On
>>>> Behalf Of
>>>> > > Karthik Krishnan
>>>> > > Sent: Friday, July 15, 2005 10:10 AM
>>>> > > To: Richard.Beare at ieee.org
>>>> > > Cc: insight-users at itk.org
>>>> > > Subject: Re: [Insight-users] Connected component statistics
>>>> > >
>>>> > >
>>>> > >
>>>> > >
>>>> > > Richard Beare wrote:
>>>> > >
>>>> > >> Hi,
>>>> > >>
>>>> > >> I've found ConnectedComponentImageFilter and
>>>> > >> RelabelConnectedImageFilter, but I haven't been able to find
>>>> anything
>>>> > >> that returns measures of the connected components -- e.g.
>>>> area/volume,
>>>> > >> perimeter/surface area, centre of gravity, ellipse moments and
>>>> minimum
>>>> > >> enclosing rectangles.
>>>> > >>
>>>> > >>
>>>> > > You could threshold the labelled image to get each connected
>>>> component
>>>> > > and...
>>>> > >
>>>> > > If you are using CVS ITK, I had put a method recently :
>>>> > >
>>>> ImageMaskSpatialObject::GetAxisAlignedBoundingBoxRegion()
>>>> > > that computes the minimum axis aligned bounding box of a 3D binary
>>>> image.
>>>> > >
>>>> > > ImageMomentsCalculator can help you with moments, CG etc.
>>>> > >
>>>> > > An appropriate class to add methods to compute volume, perimeter,
>>>> > > surface area etc may be theImageMaskSpatialObject class . You  
>>>> could
>>>> > > write them and put them back into ITK :)
>>>> > >
>>>> > > I don't think ITK has many CAD measurement methods.
>>>> > >
>>>> > >> Also, what is the recommended approach for making greyscale
>>>> measures
>>>> > >> of regions defined by a connected component image?
>>>> > >>
>>>> > >> Thanks
>>>> > >>
>>>> > >> Richard
>>>> > >> _______________________________________________
>>>> > >> Insight-users mailing list
>>>> > >> Insight-users at itk.org
>>>> > >> http://www.itk.org/mailman/listinfo/insight-users
>>>> > >>
>>>> > >>
>>>> > >>
>>>> > > _______________________________________________
>>>> > > Insight-users mailing list
>>>> > > Insight-users at itk.org
>>>> > > http://www.itk.org/mailman/listinfo/insight-users
>>>> > > _______________________________________________
>>>> > > Insight-users mailing list
>>>> > > Insight-users at itk.org
>>>> > > http://www.itk.org/mailman/listinfo/insight-users
>>>> >
>>>> >
>>>> >
>>>> > --
>>>> > Gaetan Lehmann <gaetan.lehmann at jouy.inra.fr>
>>>> > Tel: +33 1 34 65 29 66
>>>> > Biologie du Développement et de la Reproduction
>>>> > INRA de Jouy-en-Josas (France)
>>>> > Web: http://voxel.jouy.inra.fr
>>>> >
>>> _______________________________________________
>>> Insight-users mailing list
>>> Insight-users at itk.org
>>> http://www.itk.org/mailman/listinfo/insight-users
>>
>>
>>
>
>
>



-- 
Gaetan Lehmann <gaetan.lehmann at jouy.inra.fr>
Tel: +33 1 34 65 29 66
Biologie du Développement et de la Reproduction
INRA de Jouy-en-Josas (France)
Web: http://voxel.jouy.inra.fr

More information about the Insight-users mailing list