The STAPLE filter implements the Simultaneous Truth and Performance Level Estimation algorithm for generating ground truth volumes from a set of binary expert segmentations. More...

#include <itkSTAPLEImageFilter.h>

Inheritance diagram for itk::STAPLEImageFilter< TInputImage, TOutputImage >:

Collaboration diagram for itk::STAPLEImageFilter< TInputImage, TOutputImage >:

List of all members.

Public Types
typedef SmartPointer< const Self >	ConstPointer
typedef InputImageType::Pointer	InputImagePointer
typedef TInputImage	InputImageType
typedef TInputImage::PixelType	InputPixelType
typedef OutputImageType::Pointer	OutputImagePointer
typedef Superclass::OutputImageRegionType	OutputImageRegionType
typedef TOutputImage	OutputImageType
typedef TOutputImage::PixelType	OutputPixelType
typedef SmartPointer< Self >	Pointer
typedef NumericTraits < InputPixelType >::RealType	RealType
typedef STAPLEImageFilter	Self
typedef ImageToImageFilter < TInputImage, TOutputImage >	Superclass
Public Member Functions
virtual ::itk::LightObject::Pointer	CreateAnother (void) const
virtual unsigned int	GetElapsedIterations () const
virtual const char *	GetNameOfClass () const
const std::vector< double > &	GetSensitivity () const
const std::vector< double > &	GetSpecificity () const
	typedef (Concept::HasNumericTraits< InputPixelType >) InputHasNumericTraitsCheck

virtual void	SetForegroundValue (InputPixelType _arg)
virtual InputPixelType	GetForegroundValue () const

double	GetSensitivity (unsigned int i)

double	GetSpecificity (unsigned int i)

virtual void	SetMaximumIterations (unsigned int _arg)
virtual unsigned int	GetMaximumIterations () const

virtual void	SetConfidenceWeight (double _arg)
virtual double	GetConfidenceWeight () const
Static Public Member Functions
static Pointer	New ()
Static Public Attributes
static const unsigned int	ImageDimension = TOutputImage::ImageDimension
Protected Member Functions
void	GenerateData ()
void	PrintSelf (std::ostream &, Indent) const
virtual	~STAPLEImageFilter ()

	STAPLEImageFilter ()
Private Member Functions
void	operator= (const Self &)
	STAPLEImageFilter (const Self &)
Private Attributes
double	m_ConfidenceWeight
unsigned int	m_ElapsedIterations
InputPixelType	m_ForegroundValue
unsigned int	m_MaximumIterations
std::vector< double >	m_Sensitivity
std::vector< double >	m_Specificity

Detailed Description

template<typename TInputImage, typename TOutputImage>
class itk::STAPLEImageFilter< TInputImage, TOutputImage >

The STAPLE filter implements the Simultaneous Truth and Performance Level Estimation algorithm for generating ground truth volumes from a set of binary expert segmentations.

The STAPLE algorithm treats segmentation as a pixelwise classification, which leads to an averaging scheme that accounts for systematic biases in the behavior of experts in order to generate a fuzzy ground truth volume and simultaneous accuracy assessment of each expert. The ground truth volumes produced by this filter are floating point volumes of values between zero and one that indicate probability of each pixel being in the object targeted by the segmentation.

The STAPLE algorithm is described in

S. Warfield, K. Zou, W. Wells, "Validation of image segmentation and expert quality with an expectation-maximization algorithm" in MICCAI 2002: Fifth International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer-Verlag, Heidelberg, Germany, 2002, pp. 298-306

INPUTS: Input volumes to the STAPLE filter must be binary segmentations of an image, that is, there must be a single foreground value that represents positively classified pixels (pixels that are considered to belong inside the segmentation). Any number of background pixel values may be present in the input images. You can, for example, input volumes with many different labels as long as the structure you are interested in creating ground truth for is consistently labeled among all input volumes. Pixel type of the input volumes does not matter. Specify the label value for positively classified pixels using SetForegroundValue. All other labels will be considered to be negatively classified pixels (background).

Input volumes must all contain the same size RequestedRegions.

OUTPUTS: The STAPLE filter produces a single output volume with a range of floating point values from zero to one. IT IS VERY IMPORTANT TO INSTANTIATE THIS FILTER WITH A FLOATING POINT OUTPUT TYPE (floats or doubles). You may threshold the output above some probability threshold if you wish to produce a binary ground truth.

PARAMETERS: The STAPLE algorithm requires a number of inputs. You may specify any number of input volumes using the SetInput(i, p_i) method, where i ranges from zero to N-1, N is the total number of input segmentations, and p_i is the SmartPointer to the i-th segmentation.

The SetConfidenceWeight parameter is a modifier for the prior probability that any pixel would be classified as inside the target object. This implementation of the STAPLE algorithm automatically calculates prior positive classification probability as the average fraction of the image volume filled by the target object in each input segmentation. The ConfidenceWeight parameter allows for scaling the of this default prior probability: if g_t is the prior probability that a pixel would be classified inside the target object, then g_t is set to g_t * ConfidenceWeight before iterating on the solution. In general ConfidenceWeight should be left to the default of 1.0.

You must provide a foreground value using SetForegroundValue that the STAPLE algorithm will use to identify positively classified pixels in the the input images. All other values in the image will be treated as background values. For example, if your input segmentations consist of 1's everywhere inside the segmented region, then use SetForegroundValue(1).

The STAPLE algorithm is an iterative E-M algorithm and will converge on a solution after some number of iterations that cannot be known a priori. After updating the filter, the total elapsed iterations taken to converge on the solution can be queried through GetElapsedIterations(). You may also specify a MaximumNumberOfIterations, after which the algorithm will stop iterating regardless of whether or not it has converged. This implementation of the STAPLE algorithm will find the solution to within seven digits of precision unless it is stopped early.

Once updated, the Sensitivity (true positive fraction, q) and Specificity (true negative fraction, q) for each expert input volume can be queried using GetSensitivity(i) and GetSpecificity(i), where i is the i-th input volume.

REQUIRED PARAMETERS: The only required parameters for this filter are the ForegroundValue and the input volumes. All other parameters may be safely left to their default values. Please see the paper cited above for more information on the STAPLE algorithm and its parameters. A proper understanding of the algorithm is important for interpreting the results that it produces.

EVENTS: This filter invokes IterationEvent() at each iteration of the E-M algorithm. Setting the AbortGenerateData() flag will cause the algorithm to halt after the current iteration and produce results just as if it had converged. The algorithm makes no attempt to report its progress since the number of iterations needed cannot be known in advance.

Definition at line 121 of file itkSTAPLEImageFilter.h.