[Insight-users] BSpline Transformation and binary images
Luis Ibanez
luis.ibanez at kitware.com
Mon Oct 26 14:29:15 EDT 2009
Hi Serena,
Your last email is a bit confusing.
1) The Metric that offers the API to set the number of Samples
is the MattesMutualInformation one, no the MatchCardinality one.
So I assume that when you mention the 500 samples,
you are referring to the itkMattes MI metric.
In that case, please note that 500 is actually
a very small number of samples.
The typical advice is to use between 5% and 20%
of the number of pixels in your images.
How big are your images ?
(in number of pixels )
2) Have you made sure that you are compiling your
application for "Release" ?
It is well known that the computation time ratio between "Release"
and "Debug" builds in ITK applications can easily be of 1:20.
BTW: Please remind us what is your platform:
Windows/Linux/Mac, compiler version... etc.
Thanks
Luis
-------------------------------------------------
On Mon, Oct 26, 2009 at 12:46 PM, Serena Fabbri <fabbri at u.washington.edu> wrote:
>
> Hi Luis,
> I observe a long computation time when I use
> MatchCardinalityImageToImageMetric.
>
> I have checked again the documentation and only now i have realized that I
> can set the Number of samples with SetNumberOfSpatialSamples hence I used
> the default number (500) in my code.
>
> Do you think 500 samples are too much for my Registration task?
>
> Thank you.
>
> Serena.
>
>
>
>
>
>
>
> On Sat, 24 Oct 2009, Luis Ibanez wrote:
>
>> Hi Serena,
>>
>> You seem to have forgotten to mention the:
>>
>> Number of samples used in the Mattes Mutual Information metric.
>>
>> At this point, that's the main suspect of the
>> long computation time that you are observing.
>>
>>
>> Please let us know,
>>
>>
>> Thanks
>>
>>
>> Luis
>>
>>
>>
>> ---------------------------------------------------------------------------
>> On Thu, Oct 22, 2009 at 2:44 PM, Serena Fabbri <fabbri at u.washington.edu>
>> wrote:
>>>
>>> Hi Luise,
>>>
>>> This is great, I will try the correct MMI!!
>>>
>>> Regarding the Observer: good idea, thank you for the suggestion.
>>>
>>> For the registration task:
>>> I am using:
>>> Cardinality Metric with Amoeba optmizer, BSpline Transformation and
>>> NearestNeighbor Interpolator
>>> My binary images are (80;80;101) pixelsize and (4;4;1) mm spacing and
>>> (13;13;13) nodes Bspline grid.
>>>
>>>
>>> Thank you again for any advice.
>>> Serena.
>>>
>>>
>>>
>>>
>>> On Thu, 22 Oct 2009, Luis Ibanez wrote:
>>>
>>>> Hi Serena,
>>>>
>>>>
>>>> The code in
>>>>
>>>> Insight/Code/Algorithms
>>>>
>>>> has already been fixed.
>>>>
>>>> --
>>>>
>>>> I'm still working on patching the one in
>>>>
>>>> Insight/Code/Review
>>>>
>>>> This is described in :
>>>>
>>>> http://public.kitware.com/Bug/view.php?id=9745
>>>>
>>>> ---
>>>>
>>>>
>>>> Regarding the Observer:
>>>>
>>>> The number of iterations, can actually be provided
>>>> by introducing a counter in the observer itself and
>>>> counting events.
>>>>
>>>> ---
>>>>
>>>>
>>>> The time that you report is unusually long for registering
>>>> an image of that size.
>>>>
>>>> How many samples are you using in MattesMI ?
>>>>
>>>>
>>>> Please let us know,
>>>>
>>>>
>>>> Thanks
>>>>
>>>>
>>>> Luis
>>>>
>>>>
>>>>
>>>>
>>>> -----------------------------------------------------------------------------
>
> ---
>>>
>>> -------------
>>>>
>>>> On Thu, Oct 22, 2009 at 1:12 PM, Serena Fabbri <fabbri at u.washington.edu>
>>>> wrote:
>>>>>
>>>>>
>>>>> Hi Luise,
>>>>>
>>>>> I am glad that my suggestions were useful for you.
>>>>> When the MMI code will be fixed, do you write an email?
>>>>>
>>>>> Yes, I get output from optimizer->GetCachedValue() This instruction is
>>>>> inside the observer code. The registration runs but it is very slow, my
>>>>> computer is new......so what could it be wrong in my Registration task?
>>>>> I am using (13;13;13) nodes Bspline grid.
>>>>>
>>>>> I have read the the Amoeba optimizer documentation again and I have not
>>>>> found a function to obtain the number of iteration. Am I wrong?
>>>>>
>>>>>
>>>>> Thank again for your help.
>>>>>
>>>>>
>>>>> Serena.
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> On Tue, 20 Oct 2009, Luis Ibanez wrote:
>>>>>
>>>>>> Hi Serena,
>>>>>>
>>>>>> Thanks a lot for tracking this down.
>>>>>>
>>>>>> Your findings are quite interesting.
>>>>>>
>>>>>> The three lines of code that you have indicated are indeed performing
>>>>>> incorrect computations, in the case where of number of bins is == 2.
>>>>>>
>>>>>> I'm preparing a patch for this class,
>>>>>>
>>>>>> Insight/Code/Algorithms/
>>>>>> itkMattesMutualInformationImageToImageMetric.txx
>>>>>>
>>>>>> and its multi-threaded version:
>>>>>>
>>>>>> Insight/Code/Review/
>>>>>> itkOptMattesMutualInformationImageToImageMetric.txx
>>>>>>
>>>>>> that will cast properly the variable m_NumberOfHistogramBins
>>>>>> to the type OffsetValueType (which is signed).
>>>>>>
>>>>>> We need to run this in a couple of experimental builds first,
>>>>>> and then we should put together a test that exercises this
>>>>>> metric on a binary image.
>>>>>>
>>>>>>
>>>>>> Again, Thanks for tracking this issue,
>>>>>> and pointing to the location of the problem in the code.
>>>>>>
>>>>>>
>>>>>> About your second question:
>>>>>>
>>>>>> A typical registration process with images that are of size 100^3
>>>>>> pixels, should take more than 30 minutes in a modern computer.
>>>>>>
>>>>>> Are you seeing any output from the Command observer ?
>>>>>> Something is certainly wrong with the process.
>>>>>>
>>>>>> Also, please not that the call
>>>>>>
>>>>>> get_num_evaluations()
>>>>>>
>>>>>> is not quite equivalent, to number of Iterations.
>>>>>>
>>>>>> In the case of the Amoeba optimizer, what
>>>>>> happens is that the Metric may be computed multiple
>>>>>> times in the process of computing a single Iteration
>>>>>> of the optimizer. The method get_num_evaluations
>>>>>> tells you how many times the metric was evaluated
>>>>>> in the last iteration of the optimizer, not how many
>>>>>> iterations have been used by the optimizer.
>>>>>>
>>>>>>
>>>>>> Regards,
>>>>>>
>>>>>>
>>>>>> Luis
>>>>>>
>>>>>>
>>>>>> -------------------------------------------------------------------
>>>>>> On Thu, Oct 15, 2009 at 2:49 PM, Serena Fabbri
>>>>>> <fabbri at u.washington.edu>
>>>>>> wrote:
>>>>>>>
>>>>>>> Hi Luis,
>>>>>>>
>>>>>>> I have taken a look to itkMattesMutualInformationImageToImageMetric,
>>>>>>> and
>>>>>>> the
>>>>>>> problems seem to be:
>>>>>>> -line 241 (Initialize): m_NumberOfHistogramBins is unsigned long and
>>>>>>> the
>>>>>>> value of brackets expression is (-2). It can not be assigned.
>>>>>>> -line 677 (ComputeFixedImageParzenWindowIndices): same problem and
>>>>>>> besides
>>>>>>> pindex is unsigned int and it can not be (-1) the correct value.
>>>>>>> -line 758 (GetValue): same problem and again
>>>>>>> movingImageParzenWindowindex
>>>>>>> is
>>>>>>> unsigned int and it can not be (-1).
>>>>>>>
>>>>>>> I am not sure that to do a cast is ok because after the cast (line
>>>>>>> 682
>>>>>>> ComputeFixedImageParzenWindowIndices),
>>>>>>> (*iter).FixedimageParzenWindowIndex
>>>>>>> will be 1 and 2 and they (at least 2) can not be indexes for
>>>>>>> m_FixedImageMarginalPDF (line 767 GetValue) because the dimension of
>>>>>>> this
>>>>>>> array is the number of the bin (=2).
>>>>>>>
>>>>>>> Do you agree with me?
>>>>>>>
>>>>>>>
>>>>>>> Now I am running Cardinality Metric with Amoeba optmizer and BSpline
>>>>>>> Transformation.
>>>>>>> My image are (80;80;101) pixelsize and (4;4;1)mm spacing.
>>>>>>>
>>>>>>> I have set
>>>>>>> simplexDelta.Fill( 5.0 );
>>>>>>> optimizer->AutomaticInitialSimplexOff();
>>>>>>> optimizer->SetInitialSimplexDelta( simplexDelta );
>>>>>>> optimizer->SetParametersConvergenceTolerance( 2 );
>>>>>>> optimizer->SetFunctionConvergenceTolerance(1);
>>>>>>> optimizer->SetMaximumNumberOfIterations( 200 );
>>>>>>>
>>>>>>> I'd like to ask you how long the registration task can be, because it
>>>>>>> started 2 days ago and it is still running.
>>>>>>> Besides I have put optimizer->GetOptimizer()->get_num_evaluations()
>>>>>>> in
>>>>>>> observer code but it doesn't work. How can I do to visualize the
>>>>>>> number
>>>>>>> of
>>>>>>> iteration during the process?
>>>>>>>
>>>>>>>
>>>>>>> Thank you for any info.
>>>>>>>
>>>>>>> Serena.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> On Mon, 12 Oct 2009, Luis Ibanez wrote:
>>>>>>>
>>>>>>>> Hi Serena,
>>>>>>>>
>>>>>>>> I believe Hans Johnson reported a while ago that this
>>>>>>>> metric implementation had a bug that make it unsuitable
>>>>>>>> for registering binary images.
>>>>>>>>
>>>>>>>> (My mistake for suggesting to use this metric for this
>>>>>>>> case without having tried).
>>>>>>>>
>>>>>>>> Your options at this point are:
>>>>>>>>
>>>>>>>>
>>>>>>>> A) Help us debug the metric :-)
>>>>>>>>
>>>>>>>>
>>>>>>>> B) Use the Cardinality Metric instead.
>>>>>>>> itkMatchCardinalityImageToImageMetric.h
>>>>>>>>
>>>>>>>>
>>>>>>>> Please let us know what would be your preference,
>>>>>>>>
>>>>>>>>
>>>>>>>> Thanks
>>>>>>>>
>>>>>>>>
>>>>>>>> Luis
>>>>>>>>
>>>>>>>>
>>>>>>>> -----------------------------------------------------------------
>>>>>>>> On Mon, Oct 12, 2009 at 7:23 PM, Serena Fabbri
>>>>>>>> <fabbri at u.washington.edu>
>>>>>>>> wrote:
>>>>>>>>>
>>>>>>>>> Hi All,
>>>>>>>>>
>>>>>>>>> I am trying to register binary images of the head using Non Rigid
>>>>>>>>> Registration. They are (80;80;101) pixelsize and (4;4;1)mm spacing.
>>>>>>>>>
>>>>>>>>> I use Mattes-MI, LBFGSB, NearestNeighbor interpolator and BSpline
>>>>>>>>> Transformation.
>>>>>>>>> After reading the email
>>>>>>>>>
>>>>>>>>> http://www.itk.org/pipermail/insight-users/2007-November/024178.html
>>>>>>>>>
>>>>>>>>> I set 2 bin and 10% of statistic for MI and I have got this
>>>>>>>>> message:
>>>>>>>>>
>>>>>>>>> Starting Registration
>>>>>>>>> ExceptionObject caught !
>>>>>>>>>
>>>>>>>>> itk::ExceptionObject (0x1900fb0)
>>>>>>>>> Location: "void
>>>>>>>>> itk::MattesMutualInformationImageToImageMetric<TFixedImage,
>>>>>>>>> TMovingImage>::GetValueAndDerivative(typename
>>>>>>>>> itk::ImageToImageMetric<TFixedImage,
>>>>>>>>> TMovingImage>::ParametersType&,
>>>>>>>>> typename itk::ImageToImageMetric<TFixedImage,
>>>>>>>>> TMovingImage>::MeasureType&,
>>>>>>>>> typename itk::ImageToImageMetric<TFixedImage,
>>>>>>>>> TMovingImage>::DerivativeType&) const [with TFixedImage = main(int,
>>>>>>>>> char**)::FixedImageType, TMovingImage = main(int,
>>>>>>>>> char**)::MovingImageType]"
>>>>>>>>> File: /Users/InsightToolkit-
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> 3.10.0/Code/Algorithms/itkMattesMutualInformationImageToImageMetric.txx
>>>>>>>>> Line: 1160
>>>>>>>>> Description: itk::ERROR:
>>>>>>>>> MattesMutualInformationImageToImageMetric(0x1900260): Fixed image
>>>>>>>>> marginal
>>>>>>>>> PDF summed to zero
>>>>>>>>>
>>>>>>>>> I have increased the statistic until 30% and I get the same error.
>>>>>>>>> Is it possible that BSpline Transformation is not suitable for
>>>>>>>>> binary
>>>>>>>>> images
>>>>>>>>> registration?
>>>>>>>>>
>>>>>>>>> Any suggestion will be appreciate.
>>>>>>>>>
>>>>>>>>> Thanks.
>>>>>>>>>
>>>>>>>>> Serena.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> _____________________________________
>>>>>>>>> Powered by www.kitware.com
>>>>>>>>>
>>>>>>>>> Visit other Kitware open-source projects at
>>>>>>>>> http://www.kitware.com/opensource/opensource.html
>>>>>>>>>
>>>>>>>>> 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://www.itk.org/mailman/listinfo/insight-users
>>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>
>>>
>>>
>>>
>>
>
>
>
>
More information about the Insight-users
mailing list