#include <ImageSimilarity.h>

Inheritance diagram for mirtk::ImageSimilarity:

Collaboration diagram for mirtk::ImageSimilarity:

Public Types
enum	ForegroundRegion { FG_Domain, FG_Mask, FG_Target, FG_Overlap, FG_Union }

typedef GenericImage< GradientType >	GradientImageType
	Type of similarity gradient image.

typedef double	GradientType
	Type of similarity gradient components.

typedef RegisteredImage::VoxelType	VoxelType
	Voxel type of registered images.

Public Types inherited from mirtk::EnergyTerm
typedef ObjectFactory< enum EnergyMeasure, EnergyTerm >	FactoryType
	Type of energy term factory.

Public Member Functions
void	CopyAttributes (const ImageSimilarity &)
	Copy attributes of this class from another instance.

virtual void	Exclude (const blocked_range3d< int > &)

virtual void	Include (const blocked_range3d< int > &)

virtual void	Initialize ()
	Initialize similarity measure once input and parameters have been set.

bool	IsForeground (int) const
	Whether to evaluate similarity at specified voxel.

bool	IsForeground (int, int, int) const
	Whether to evaluate similarity at specified voxel.

	mirtkAttributeMacro (bool, InitialUpdate)
	Whether Update has not been called since initialization.

	mirtkComponentMacro (GradientImageType, GradientWrtTarget)
	Memory for (non-parametric) similarity gradient w.r.t target transformation.

	mirtkComponentMacro (GradientImageType, GradientWrtSource)
	Memory for (non-parametric) similarity gradient w.r.t source transformation.

	mirtkComponentMacro (double, Gradient)
	Memory for (parametric) similarity gradient.

	mirtkLooseComponentMacro (RegisteredImage, Target)
	(Transformed) Target image

	mirtkLooseComponentMacro (RegisteredImage, Source)
	(Transformed) Source image

	mirtkPublicAggregateMacro (BinaryImage, Mask)

	mirtkPublicAttributeMacro (ImageAttributes, Domain)
	Finite regular domain on which to resample images and evaluate similarity.

	mirtkPublicAttributeMacro (ForegroundRegion, Foreground)
	Set operation used to define common foreground region of co-registered images.

	mirtkPublicAttributeMacro (int, NumberOfVoxels)
	Number of voxels per registered image.

	mirtkPublicAttributeMacro (bool, NormalizeImageGradient)
	Divide transformed image gradient by input intensity range.

	mirtkPublicAttributeMacro (bool, UseApproximateGradient)

	mirtkPublicAttributeMacro (double, VoxelWisePreconditioning)

	mirtkPublicAttributeMacro (double, NodeBasedPreconditioning)

	mirtkPublicAttributeMacro (bool, SkipTargetInitialization)
	Skip initialization of target image.

	mirtkPublicAttributeMacro (bool, SkipSourceInitialization)
	Skip initialization of source image.

virtual ParameterList	Parameter () const
	Get parameter key/value as string map.

virtual void	Print (Indent=0) const
	Print debug information.

void	ReleaseSource ()
	Release input source image.

void	ReleaseTarget ()
	Release input target image.

virtual void	Update (bool=true)
	Update moving input image(s) and internal state of similarity measure.

virtual void	WriteDataSets (const char , const char , bool=true) const
	Write input of data fidelity term.

virtual void	WriteGradient (const char , const char ) const
	Write gradient of data fidelity term w.r.t each transformed input.

virtual	~ImageSimilarity ()
	Destructor.

Public Member Functions inherited from mirtk::DataFidelity
virtual	~DataFidelity ()
	Destructor.

Public Member Functions inherited from mirtk::EnergyTerm
virtual enum EnergyMeasure	EnergyMeasure () const =0
	Energy measure implemented by this term.

void	Gradient (double *gradient, double step)

virtual void	GradientStep (const double *gradient, double &min, double &max) const

double	InitialValue ()
	Returns initial value of energy term.

void	NormalizedGradient (double *gradient, double step)

string	Prefix (const char *=NULL) const
	Prefix to be used for debug output files.

virtual double	RawValue (double) const

double	RawValue ()

void	ResetInitialValue ()
	Reset initial value of energy term.

void	ResetValue ()
	Reset cached value of energy term.

virtual bool	Upgrade ()
	Update energy term after convergence.

double	Value ()
	Evaluate energy term.

virtual	~EnergyTerm ()
	Destructor.

Public Member Functions inherited from mirtk::Configurable
string	DefaultName () const

virtual bool	Set (const char , const char )
	Set parameter value from string.

virtual	~Configurable ()
	Destructor.

Public Member Functions inherited from mirtk::Observable
void	AddObserver (Observer &)
	Add observer.

void	Broadcast (Event, const void *=NULL)
	Broadcast event to observers.

void	ClearObservers ()
	Delete all observers.

void	DeleteObserver (Observer &)
	Delete observer.

void	NotifyObservers (Event, const void *=NULL)
	Notify all observers about given event if this object has changed.

int	NumberOfObservers () const
	Number of current observers.

virtual	~Observable ()
	Destructor.

Public Member Functions inherited from mirtk::Object
virtual const char *	NameOfClass () const =0
	Get name of class, which this object is an instance of.

bool	Parameter (const ParameterList &)
	Set parameters from name/value pairs.

virtual	~Object ()
	Destructor.

Static Public Member Functions
static ImageSimilarity *	New (SimilarityMeasure, const char *="", double=1.0)
	Instantiate specified similarity measure.

Static Public Member Functions inherited from mirtk::EnergyTerm
static FactoryType &	Factory ()
	Get global energy term factory instance.

static EnergyTerm *	New (EnergyMeasure, const char *="", double=1.0)
	Construct new energy term.

static EnergyTerm *	TryNew (EnergyMeasure, const char *="", double=1.0)
	Construct new energy term or return nullptr if term not available.

Static Public Member Functions inherited from mirtk::Object
static const char *	NameOfType ()
	Get name of this class type.

Protected Member Functions
void	ApproximateGradient (RegisteredImage image, FreeFormTransformation ffd, double *gradient, double step, double weight)

void	ApproximateGradient (RegisteredImage image, double gradient, double step, double weight)

virtual void	EvaluateGradient (RegisteredImage image, GradientImageType &np_gradient, double *gradient, double step, double weight)

virtual void	EvaluateGradient (double *gradient, double step, double weight)

	ImageSimilarity (const char *="", double=1.0)
	Constructor.

	ImageSimilarity (const ImageSimilarity &)
	Copy constructor.

virtual void	InitializeInput (const ImageAttributes &domain)

void	MultiplyByImageGradient (const RegisteredImage image, GradientImageType gradient)

virtual bool	NonParametricGradient (const RegisteredImage image, GradientImageType gradient)

virtual void	NormalizeGradient (GradientImageType *gradient)

virtual void	NormalizeGradient (const RegisteredImage image, double gradient)

ImageSimilarity &	operator= (const ImageSimilarity &)
	Assignment operator.

virtual void	ParametricGradient (const RegisteredImage image, GradientImageType np_gradient, double *gradient, double weight)

virtual bool	SetWithoutPrefix (const char , const char )
	Set parameter value from string.

Protected Member Functions inherited from mirtk::DataFidelity
	DataFidelity (const char *="", double=1.0)
	Constructor.

	DataFidelity (const DataFidelity &)
	Copy constructor.

DataFidelity &	operator= (const DataFidelity &)
	Assignment operator.

virtual bool	SetWithPrefix (const char , const char )
	Set parameter value from string.

Protected Member Functions inherited from mirtk::EnergyTerm
	EnergyTerm (const char *="", double=1.0)
	Constructor.

	EnergyTerm (const EnergyTerm &)
	Copy constructor.

virtual double	Evaluate ()=0
	Evaluate unweighted energy term.

EnergyTerm &	operator= (const EnergyTerm &)
	Assignment operator.

Protected Member Functions inherited from mirtk::Configurable
	Configurable (const char *="")
	Constructor.

	Configurable (const Configurable &)
	Copy constructor.

string	DefaultPrefix () const
	Get default object name prefix (if any)

bool	HasName () const
	Whether this object has an explicit name.

bool	HasPrefix () const
	Whether this object has either an explicit name or default prefix.

template<class T >
bool	InsertWithPrefix (ParameterList &, string, T) const
	Insert parameter into name/value list with object name prefix.

bool	InsertWithPrefix (ParameterList &, const ParameterList &) const
	Insert parameters into name/value list with object name prefix.

Configurable &	operator= (const Configurable &)
	Assignment operator.

string	ParameterNameWithoutPrefix (const char *) const
	Get name of parameter without object name prefix.

string	ParameterNameWithPrefix (const string &) const
	Get name of parameter with default object name prefix.

string	ParameterNameWithPrefix (const char *) const
	Get name of parameter with default object name prefix.

Protected Member Functions inherited from mirtk::Observable
	Observable ()
	Default constructor.

	Observable (const Observable &)
	Copy constructor.

Observable &	operator= (const Observable &)
	Assignment operator.

Protected Member Functions inherited from mirtk::Object
template<typename... Args>
void	Throw (ErrorType err, const char *func, Args... args) const

Additional Inherited Members
Static Protected Member Functions inherited from mirtk::Object
template<typename... Args>
static void	ThrowStatic (ErrorType err, const char cls, const char func, Args... args)

Detailed Description

Base class for image similarity measures

The lower the value of the similarity measure, the more similar the images are. It may therefore more precisely be referred to as dissimilarity measure, but both terms are for historic reasons used interchangeably in this framework.

If the transformed image has a foreground region defined, we assume that this corresponds to the region which needs to be matched with the untransformed image. Therefore we make use of the entire transformed foreground region to evaluate similarity and in particular the gradient forces defined within this region. Note that outside this source foreground region, the gradient is always zero and thus cannot be used to drive the image registration.

Otherwise, the region of interest is generally defined by the foreground of the untransformed image. For each such target voxel we want to find a suitable corresponding voxel in the other image. Therefore, we restrict the similarity evaluation to this region. In case of a symmetric transformation of both images, the union of the foreground of both images defines the region for which similarity is evaluated and forces are computed.

Subclasses which implement a particular similarity measure should call the IsForeground member function to decide whether or not to consider a given voxel of the grid on which the registered images are defined.

Note: The similarity measure owns the registered input images and may thus modify these to improve the runtime of each Update step.

Definition at line 61 of file ImageSimilarity.h.

Member Enumeration Documentation

§ ForegroundRegion

enum mirtk::ImageSimilarity::ForegroundRegion

Enumeration of available set operations to define region within which to evaluate the image similarity given the two foreground regions of the two co-registered input images. The resulting foreground region is further intersected with the specified binary mask. When no mask is given, a mask with constant value 1 is assumed.

Enumerator
FG_Domain	Evaluate similarity for all voxels in image domain, ignore mask.
FG_Mask	Evaluate similarity for all voxels in domain or with non-zero mask value.
FG_Target	Evaluate similarity for foreground of untransformed image.
FG_Overlap	Evaluate similarity for intersection of foreground regions.
FG_Union	Evaluate similarity for union of foreground regions.

Definition at line 83 of file ImageSimilarity.h.

Member Function Documentation

§ ApproximateGradient() [1/2]

void mirtk::ImageSimilarity::ApproximateGradient	(	RegisteredImage *	image,
		FreeFormTransformation *	ffd,
		double *	gradient,
		double	step,
		double	weight
	)

protected

Approximate similarity gradient using finite differences

If the image similarity does not provide an implementation of the NonParametricGradient function, the similarity gradient is approximated instead using finite differences.

Parameters

[in]	image	Transformed image.
[in]	ffd	Free-form deformation.
[in,out]	gradient	Gradient to which the computed parametric gradient is added, after multiplication by the given `weight`.
[in]	step	Step size to use for finite differences.
[in]	weight	Weight of image similarity.

§ ApproximateGradient() [2/2]

void mirtk::ImageSimilarity::ApproximateGradient	(	RegisteredImage *	image,
		double *	gradient,
		double	step,
		double	weight
	)

protected

Approximate similarity gradient using finite differences

If the image similarity does not provide an implementation of the NonParametricGradient function, the similarity gradient is approximated instead using finite differences.

Parameters

[in]	image	Transformed image.
[in,out]	gradient	Gradient to which the computed parametric gradient is added, after multiplication by the given `weight`.
[in]	step	Step size to use for finite differences.
[in]	weight	Weight of image similarity.

§ EvaluateGradient() [1/2]

virtual void mirtk::ImageSimilarity::EvaluateGradient	(	RegisteredImage *	image,
		GradientImageType *&	np_gradient,
		double *	gradient,
		double	step,
		double	weight
	)

protectedvirtual

Evaluate similarity gradient

This function calls the virtual NonParametricGradient function to be implemented by subclasses for each transformed input image to obtain the voxel-wise similarity gradient. It then converts this gradient into a gradient w.r.t the transformation parameters using the ParametricGradient.

If both target and source are transformed by different transformations, the resulting gradient vector contains first the derivative values w.r.t the parameters of the target transformation followed by those computed w.r.t the parameters of the source transformation. If both images are transformed by the same transformation, the sum of the derivative values is added to the resulting gradient vector. This is in particular the case for a velocity based transformation model which is applied to deform both images "mid-way". Otherwise, only one input image is transformed (usually the source) and the derivative values of only the respective transformation parameters added to the gradient vector.

See also: NonParametricGradient, ParametricGradient

Parameters

[in]	image	Transformed image.
[in,out]	np_gradient	Memory for voxel-wise non-parametric gradient.
[in,out]	gradient	Gradient to which the computed gradient of the image similarity is added after multiplying by the given similarity `weight`.
[in]	step	Step size to use for finite differences.
[in]	weight	Weight of image similarity.

§ EvaluateGradient() [2/2]

virtual void mirtk::ImageSimilarity::EvaluateGradient	(	double *	gradient,
		double	step,
		double	weight
	)

protectedvirtual

Evaluate similarity gradient

This function calls the virtual NonParametricGradient function to be implemented by subclasses for each transformed input image to obtain the voxel-wise similarity gradient. It then converts this gradient into a gradient w.r.t the transformation parameters using the ParametricGradient.

If both target and source are transformed by different transformations, the resulting gradient vector contains first the derivative values w.r.t the parameters of the target transformation followed by those computed w.r.t the parameters of the source transformation. If both images are transformed by the same transformation, the sum of the derivative values is added to the resulting gradient vector. This is in particular the case for a velocity based transformation model which is applied to deform both images "mid-way". Otherwise, only one input image is transformed (usually the source) and the derivative values of only the respective transformation parameters added to the gradient vector.

See also: NonParametricGradient, ParametricGradient

Parameters

[in,out]	gradient	Gradient to which the computed gradient of the image similarity is added after multiplying by the given similarity `weight`.
[in]	step	Step size to use for finite differences.
[in]	weight	Weight of image similarity.

Implements mirtk::EnergyTerm.

§ Exclude()

virtual void mirtk::ImageSimilarity::Exclude ( const blocked_range3d< int > & )

virtual

Exclude region from similarity evaluation

Called by ApproximateGradient before the registered image region of the transformed image is updated.

Override in subclass for more efficient ApproximateGradient evaluation If the analytic derivation is possible and thus the NonParametericGradient function is overriden instead, this function is not used. Otherwise, if this function is not overriden to update the similarity measure, the Evaluate function must re-evaluate the similarity for all voxels.

See also: ApproximateGradient, Include

Reimplemented in mirtk::NormalizedIntensityCrossCorrelation, mirtk::HistogramImageSimilarity, and mirtk::SumOfSquaredIntensityDifferences.

§ Include()

virtual void mirtk::ImageSimilarity::Include ( const blocked_range3d< int > & )

virtual

Include region in similarity evaluation

Called by ApproximateGradient after the registered image region of the transformed image is updated.

Override in subclass for more efficient ApproximateGradient evaluation If the analytic derivation is possible and thus the NonParametericGradient function is overriden instead, this function is not used. Otherwise, if this function is not overriden to update the similarity measure, the Evaluate function must re-evaluate the similarity for all voxels.

See also: ApproximateGradient, Exclude

Reimplemented in mirtk::NormalizedIntensityCrossCorrelation, mirtk::HistogramImageSimilarity, and mirtk::SumOfSquaredIntensityDifferences.

§ InitializeInput()

virtual void mirtk::ImageSimilarity::InitializeInput ( const ImageAttributes & domain )

protectedvirtual

Initialize similarity measure once input and parameters have been set

Parameters

[in] domain Image domain on which the similarity is evaluated.

Reimplemented in mirtk::GradientFieldSimilarity.

§ mirtkPublicAggregateMacro()

mirtk::ImageSimilarity::mirtkPublicAggregateMacro	(	BinaryImage	,
		Mask
	)

Mask which defines arbitrary domain on which the similarity is evaluated

Intensities outside the mask (i.e., mask value is zero) are excluded from the similarity comparison. The foreground domain of the registered image is the intersection of the domain defined by non-zero mask entries with the foreground domain used when no mask is set.

§ mirtkPublicAttributeMacro() [1/3]

mirtk::ImageSimilarity::mirtkPublicAttributeMacro	(	bool	,
		UseApproximateGradient
	)

Approximate gradient using finite differences even if the similarity measure implements the NonParametricGradient function

§ mirtkPublicAttributeMacro() [2/3]

mirtk::ImageSimilarity::mirtkPublicAttributeMacro	(	double	,
		VoxelWisePreconditioning
	)

Voxel-wise gradient preconditioning sigma used to supress noise. A non-positive value disables the voxel-wise preconditioning all together.

Zikic, D., Baust, M., Kamen, A., & Navab, N. A General Preconditioning Scheme for Difference Measures in Deformable Registration. In ICCV 2011.

§ mirtkPublicAttributeMacro() [3/3]

mirtk::ImageSimilarity::mirtkPublicAttributeMacro	(	double	,
		NodeBasedPreconditioning
	)

Node-based (M)FFD control point gradient preconditioning sigma used to supress noise. A non-positive value disables the node-based preconditioning all together.

Zikic, D., Baust, M., Kamen, A., & Navab, N. A General Preconditioning Scheme for Difference Measures in Deformable Registration. In ICCV 2011.

§ MultiplyByImageGradient()

void mirtk::ImageSimilarity::MultiplyByImageGradient	(	const RegisteredImage *	image,
		GradientImageType *	gradient
	)

protected

Multiply voxel-wise similarity gradient by transformed image gradient

This function is intended for use by subclass implementations to compute the NonParametericGradient. It applies the chain rule to compute \(\frac{dSimilarity}{dy} = \frac{dSimilarity}{dI} * \frac{dI}{dy}\), given \(\frac{dSimilarity}{dI}\) as input, where \(y = T(x)\).

Parameters

[in]	image	Transformed image.
[in,out]	gradient	Input must be the gradient of the image similarity w.r.t. the transformed `image` in x. Output is the voxel-wise gradient of the similarity w.r.t. T(x).

§ NonParametricGradient()

virtual bool mirtk::ImageSimilarity::NonParametricGradient	(	const RegisteredImage *	image,
		GradientImageType *	gradient
	)

protectedvirtual

Compute voxel-wise non-parametric similarity gradient w.r.t the given image

Must be implemented by subclasses to compute the similarity gradient. The base class implementation can be used to convert the similarity gradient computed w.r.t transformed image (i.e., dSimilarity/dI) to a voxel-wise non-parametric gradient (i.e., dSimilarity/dT). Note that the input must be a scalar field only. The base class copies the x component of the input gradient image to the y and z components before applying the chain rule.

Parameters

[in]	image	Transformed image
[out]	gradient	Non-parametric similarity gradient.

Returns: Whether voxel-wise similarity gradient has been computed.

Reimplemented in mirtk::NormalizedIntensityCrossCorrelation, mirtk::SumOfSquaredIntensityDifferences, mirtk::NormalizedMutualImageInformation, and mirtk::CosineOfNormalizedGradientField.

§ NormalizeGradient() [1/2]

virtual void mirtk::ImageSimilarity::NormalizeGradient ( GradientImageType * gradient )

protectedvirtual

Normalize voxel-wise non-parametric similarity gradient

Zikic, D., Baust, M., Kamen, A., & Navab, N. A General Preconditioning Scheme for Difference Measures in Deformable Registration. In ICCV 2011.

Parameters

[in,out] gradient Non-parametric similarity gradient.

§ NormalizeGradient() [2/2]

virtual void mirtk::ImageSimilarity::NormalizeGradient	(	const RegisteredImage *	image,
		double *	gradient
	)

protectedvirtual

Normalize node-based similarity gradient

Zikic, D., Baust, M., Kamen, A., & Navab, N. A General Preconditioning Scheme for Difference Measures in Deformable Registration. In ICCV 2011.

This function applies the normalization for FFD transformations to the gradient vectors of the control point coefficients. It does nothing for non-FFD transformations. In case of a multi-level FFD with more than one active level, it furthermore normalizes the gradient vectors across levels.

Parameters

[in]	image	Transformed image.
[in,out]	gradient	Parametric similarity gradient.

§ ParametricGradient()

virtual void mirtk::ImageSimilarity::ParametricGradient	(	const RegisteredImage *	image,
		GradientImageType *	np_gradient,
		double *	gradient,
		double	weight
	)

protectedvirtual

Convert non-parametric similarity gradient into gradient w.r.t transformation parameters

This function calls Transformation::ParametricGradient of the transformation to apply the chain rule in order to obtain the similarity gradient w.r.t the transformation parameters. It adds the weighted gradient to the final registration energy gradient.

Parameters

[in]	image	Transformed image.
[in]	np_gradient	Voxel-wise non-parametric gradient.
[in,out]	gradient	Gradient to which the computed parametric gradient is added, after multiplication by the given `weight`.
[in]	weight	Weight of image similarity.

Reimplemented in mirtk::GradientFieldSimilarity.

The documentation for this class was generated from the following file:

ImageSimilarity.h

Public Types

Public Member Functions

Static Public Member Functions

Protected Member Functions

Additional Inherited Members

Detailed Description

Member Enumeration Documentation

§ ForegroundRegion

Member Function Documentation

§ ApproximateGradient() [1/2]

§ ApproximateGradient() [2/2]

§ EvaluateGradient() [1/2]

§ EvaluateGradient() [2/2]

§ Exclude()

§ Include()

§ InitializeInput()

§ mirtkPublicAggregateMacro()

§ mirtkPublicAttributeMacro() [1/3]

§ mirtkPublicAttributeMacro() [2/3]

§ mirtkPublicAttributeMacro() [3/3]

§ MultiplyByImageGradient()

§ NonParametricGradient()

§ NormalizeGradient() [1/2]

§ NormalizeGradient() [2/2]

§ ParametricGradient()