[Insight-users] weird behavior of ActiveShapeModelGradientSearchMethod related with number of iterations

[Sergio Vera]

Well, after a couple of days analyzing the code I cannot find the cause of
the problem.

At the training stage and using default values, the
ActiveShapeModelCalculator does not generate a valid model (0
eigenvectors/eigenvalues are extracted from the model). This can be solved
by changing the ActiveShapeModelCalculator::SetUpperThresholdDistance() from
the default 1 to 0.8. Changing this value, 3 eigenvectors are extracted.

However, at the searching stage, the algorithm seems to diverge:
ActiveShapeModelGradientSearchMethod starts to move far from the target
region, as iterations increase, escaping from the image. I've done many
tests, even with sintentic images but so far I don't know why this happens.

Could someone try the code with the supplied example images and see if they
experiment the same behavior?
Many thanks

// CODE--------------------------------------------------------
#include <conio.h>
#include "itkImageFileReader.h"
#include "itkActiveShapeModelCalculator.h" // .h and .txx files needed
#include "itkActiveShapeModelGradientSearchMethod.h" // .h and .txx files
needed

int itkActiveShapeModelCalculatorTest(int, char*[]);
int
main(int argc, char* argv[])
{
    exit(itkActiveShapeModelCalculatorTest(argc,argv));
}

int itkActiveShapeModelCalculatorTest( int argc, char * argv[] )
{
    if( argc < 3 )
    {
        std::cerr << "Usage: " << std::endl;
        std::cerr << argv[0] << " Input3DImageBinaryFile Input2DImageFile"
<< std::endl;
        return -1;
    }
// ------ALGORITHM PARAMETERS------
  const unsigned int Dimension             = 3;
  const float m_Tolerance                  = 2.0;
  const float m_Thickness                     = 0.8; // devault value (1)
does not generate a valid model (0 eigenvectors)!!!!
  const unsigned int m_LenghtOfProfile            = 3;
  const unsigned int m_NumberOfIteration          = 2; // increasing this
value ends with a sliding shape and an app crash!!!!!!


  typedef float      PixelType;
  typedef vnl_vector<double>
VectorOfDoubleType;
  typedef vnl_matrix<double>
MatrixOfDoubleType;
  typedef itk::Image< PixelType, Dimension >    Image3DType;
  typedef itk::Image< unsigned char, 2 >    Image2DType;

  typedef itk::ImageFileReader< Image3DType  >  ReaderType;
  typedef itk::ActiveShapeModelCalculator< Image3DType >
ActiveShapeModelCalculatorType;
  typedef itk::ActiveShapeModelGradientSearchMethod< Image2DType>
ImageSearchType;
  ActiveShapeModelCalculatorType::Pointer
    applyActiveShapeModelCalculator = ActiveShapeModelCalculatorType::New();
  ImageSearchType::Pointer
    ImageSearch = ImageSearchType::New();
  ReaderType::Pointer reader = ReaderType::New();
  Image3DType::ConstPointer inputImage;

  const char * inputFilename  = argv[1];
  reader->SetFileName( inputFilename  );
  inputImage = reader->GetOutput();
  reader->Update( );
  applyActiveShapeModelCalculator->SetImage( inputImage );
  //----------------------------------------------------------------------
  //Set the parameters for ActiveShapeModelCalculator
  //----------------------------------------------------------------------
  applyActiveShapeModelCalculator->SetTolerance( m_Tolerance );

  applyActiveShapeModelCalculator->SetUpperThresholdDistance(m_Thickness);

  applyActiveShapeModelCalculator->GenerateData();

  //Test the printself function to increase coverage
  applyActiveShapeModelCalculator->Print(std::cout);

  //Exercise TypeMacro in superclass
  typedef ActiveShapeModelCalculatorType::Superclass GenericEstimatorType;
  std::cout <<
applyActiveShapeModelCalculator->GenericEstimatorType::GetNameOfClass() <<
std::endl;

  //Print out the number of training images and the number of principal
  //components
  std::cout << "The number of training images are: " <<
    applyActiveShapeModelCalculator->GetNumberOfTrainingImages() <<
std::endl;

  std::cout << "The mean shape: " <<
    applyActiveShapeModelCalculator->GetMeanShape() << std::endl;

  vnl_vector<double> eigenValues  =
applyActiveShapeModelCalculator->GetEigenvalues();
  vnl_matrix<double> eigenVectors =
applyActiveShapeModelCalculator->GetEigenvector();

  //Print the eigen values and eigen vectors
  std::cout << "The " << eigenValues.size() << " first values of the eigen
vectors are:" << std::endl;
   for(unsigned int i = 0; i < eigenVectors.rows(); i++)
    {
    std::cout<< eigenVectors.get_row(i)<<" ";
    }
  std::cout <<""<<std::endl;
  std::cout << "The " << eigenValues.size() << " largest eigen values are:"
<< std::endl;
  std::cout << eigenValues << std::endl;

  typedef itk::ImageFileReader< Image2DType  >  Reader2DType;
  Reader2DType::Pointer reader1 = Reader2DType::New();
  const char * input2DFilename  = argv[2];
  reader1->SetFileName( input2DFilename  );
  reader1->Update( );

  ImageSearch->SetImage( reader1->GetOutput() );

 //----------------------------------------------------------------------
  //Set the parameters ActiveShapeModelSearchingImageFilter
  //----------------------------------------------------------------------
  ImageSearch->SetLenghtOfProfile( m_LenghtOfProfile );
  ImageSearch->SetNumberOfIteration( m_NumberOfIteration );
  ImageSearch->SetMeanShape( applyActiveShapeModelCalculator->GetMeanShape()
+ 0.5);
  ImageSearch->SetEigenValues(
applyActiveShapeModelCalculator->GetEigenvalues() );
  ImageSearch->SetEigenVectors(
applyActiveShapeModelCalculator->GetEigenvector() );

  ImageSearch->GenerateData();

  //Test the printself function to increase coverage
  ImageSearch->Print(std::cout);

  //Exercise TypeMacro in superclass
  typedef ImageSearchType::Superclass GenericEstimator2Type;
  std::cout << ImageSearch->GenericEstimator2Type::GetNameOfClass() <<
std::endl;

  std::cout << "The new shape: " <<
  ImageSearch->GetNewShape() << std::endl;

  // Software Guide : BeginCodeSnippet
  // Software Guide : EndCodeSnippet
  return 0;
}

On Thu, Feb 11, 2010 at 5:55 PM, Sergio Vera <sergio.vera at alma3d.com> wrote:

> Hello!
>
> I've wrapped the example of Active Shape Models (ASM) found in the example
> "itkActiveShapeModelCalculatorTest.cxx" with a small gui to be able to play
> with the tolerance, iteration and profilelenght values.
> Using the supplied model.mha and feet.mha images, the program runs smoothly
> and generates a reasonably correct mean shape:
>
> However, with the 2 iterations shown in the example cxx file, in the
> ActiveShapeModelGradientSearchMethod  the process does not segment correctly
> the image.  When incrementing the iteration steps, the result seems strange
> to me.
> The resulting polygon (that starts reasonably close to the target bone),
> does not seems to deformate or adapt to the input image at all, and only
> seems to move out of the image as iterations increase.
>
> At 20 iterations the polygon is starting to be outside the image, with a
> couple of vertex with negative x values, and at 21 iterations the
> coordinates of the resulting polygon have skyrocketed to the order of
> hundreds of millions of pixels.
> At 22 iterations the application crashes.
>
> Does any of you has experienced a similar behavior with this example?
> Any clue for what may be happening?
>
> Regards
>
> Sergio
>
>


-- 
Sergio Vera

Alma IT Systems
C/ Vilana, 4B, 4º 1ª
08022 Barcelona
T. (+34) 932 380 592
www.alma3d.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.itk.org/pipermail/insight-users/attachments/20100217/00f4267c/attachment-0001.htm>