algorithms.registration.similarity_measures¶
Module: algorithms.registration.similarity_measures
¶
Inheritance diagram for nipy.algorithms.registration.similarity_measures
:

Classes¶
CorrelationCoefficient
¶
-
class
nipy.algorithms.registration.similarity_measures.
CorrelationCoefficient
(shape, renormalize=False, dist=None)¶ Bases:
nipy.algorithms.registration.similarity_measures.SimilarityMeasure
Use a bivariate Gaussian as a distribution model
Methods
__call__
(H)loss
(H)npoints
(H)-
__init__
(shape, renormalize=False, dist=None)¶
-
loss
(H)¶
-
npoints
(H)¶
-
CorrelationRatio
¶
-
class
nipy.algorithms.registration.similarity_measures.
CorrelationRatio
(shape, renormalize=False, dist=None)¶ Bases:
nipy.algorithms.registration.similarity_measures.SimilarityMeasure
Use a nonlinear regression model with Gaussian errors as a distribution model
Methods
__call__
(H)loss
(H)npoints
(H)-
__init__
(shape, renormalize=False, dist=None)¶
-
loss
(H)¶
-
npoints
(H)¶
-
CorrelationRatioL1
¶
-
class
nipy.algorithms.registration.similarity_measures.
CorrelationRatioL1
(shape, renormalize=False, dist=None)¶ Bases:
nipy.algorithms.registration.similarity_measures.SimilarityMeasure
Use a nonlinear regression model with Laplace distributed errors as a distribution model
Methods
__call__
(H)loss
(H)npoints
(H)-
__init__
(shape, renormalize=False, dist=None)¶
-
loss
(H)¶
-
npoints
(H)¶
-
DiscreteParzenMutualInformation
¶
-
class
nipy.algorithms.registration.similarity_measures.
DiscreteParzenMutualInformation
(shape, renormalize=False, dist=None)¶ Bases:
nipy.algorithms.registration.similarity_measures.SimilarityMeasure
Use Parzen windowing in the discrete case to estimate the distribution model
Methods
__call__
(H)loss
(H)npoints
(H)-
__init__
(shape, renormalize=False, dist=None)¶
-
loss
(H)¶
-
npoints
(H)¶
-
MutualInformation
¶
-
class
nipy.algorithms.registration.similarity_measures.
MutualInformation
(shape, renormalize=False, dist=None)¶ Bases:
nipy.algorithms.registration.similarity_measures.SimilarityMeasure
Use the normalized joint histogram as a distribution model
Methods
__call__
(H)loss
(H)npoints
(H)-
__init__
(shape, renormalize=False, dist=None)¶
-
loss
(H)¶
-
npoints
(H)¶
-
NormalizedMutualInformation
¶
-
class
nipy.algorithms.registration.similarity_measures.
NormalizedMutualInformation
(shape, renormalize=False, dist=None)¶ Bases:
nipy.algorithms.registration.similarity_measures.SimilarityMeasure
- NMI = 2*(1 - H(I,J)/[H(I)+H(J)])
- = 2*MI/[H(I)+H(J)])
Methods
__call__
(H)loss
(H)npoints
(H)-
__init__
(shape, renormalize=False, dist=None)¶
-
loss
(H)¶
-
npoints
(H)¶
ParzenMutualInformation
¶
-
class
nipy.algorithms.registration.similarity_measures.
ParzenMutualInformation
(shape, renormalize=False, dist=None)¶ Bases:
nipy.algorithms.registration.similarity_measures.SimilarityMeasure
Use Parzen windowing to estimate the distribution model
Methods
__call__
(H)loss
(H)npoints
(H)-
__init__
(shape, renormalize=False, dist=None)¶
-
loss
(H)¶
-
npoints
(H)¶
-
SupervisedLikelihoodRatio
¶
-
class
nipy.algorithms.registration.similarity_measures.
SupervisedLikelihoodRatio
(shape, renormalize=False, dist=None)¶ Bases:
nipy.algorithms.registration.similarity_measures.SimilarityMeasure
Assume a joint intensity distribution model is given by self.dist
Methods
__call__
(H)loss
(H)npoints
(H)-
__init__
(shape, renormalize=False, dist=None)¶
-
loss
(H)¶
-
npoints
(H)¶
-
Functions¶
-
nipy.algorithms.registration.similarity_measures.
correlation2loglikelihood
(rho2, npts)¶ Re-normalize correlation.
Convert a squared normalized correlation to a proper log-likelihood associated with a registration problem. The result is a function of both the input correlation and the number of points in the image overlap.
See: Roche, medical image registration through statistical inference, 2001.
Parameters: rho2: float :
Squared correlation measure
npts: int :
Number of points involved in computing rho2
Returns: ll: float :
Log-likelihood re-normalized rho2
-
nipy.algorithms.registration.similarity_measures.
dist2loss
(q, qI=None, qJ=None)¶ Convert a joint distribution model q(i,j) into a pointwise loss:
L(i,j) = - log q(i,j)/(q(i)q(j))
where q(i) = sum_j q(i,j) and q(j) = sum_i q(i,j)
See: Roche, medical image registration through statistical inference, 2001.