Affine optimization

AffineOptimizer

class valis.affine_optimizer.AffineOptimizer(nlevels=1, nbins=256, optimization='Powell', transformation='EuclideanTransform')[source]

Class that optimizes ridid registration

nlevels

Number of levels in the Gaussian pyramid

Type

int

nbins

Number of bins to have in histograms used to estimate mutual information

Type

int

optimization

Optimization method. Can be any method from scipy.optimize “FuzzyPSO” for Fuzzy Self-Tuning PSO in the fst-pso package (https://pypi.org/project/fst-pso/) “gp_minimize”, “forest_minimize”, “gbrt_minimize” from scikit-opt

Type

str

transformation

Type of transformation, “EuclideanTransform” or “SimilarityTransform”

Type

str

current_level

Current level of the Guassian pyramid that is being registered

Type

int

accepts_xy

Bool declaring whether or not the optimizer will use corresponding points to optimize the registration

Type

bool

setup(moving, fixed, mask, initial_M=None)[source]

Gets images ready for alignment

cost_fxn(fixed_image, transformed, mask)[source]

Calculates metric that is to be minimized

align(moving, fixed, mask, initial_M=None, moving_xy=None, fixed_xy=None)[source]

Align images by minimizing cost_fxn

Notes

All AffineOptimizer subclasses need to have the method align(moving, fixed, mask, initial_M, moving_xy, fixed_xy) that returns the aligned image, optimal_M, cost_list

AffineOptimizer subclasses must also have a cost_fxn(fixed_image, transformed, mask) method that returns the registration metric value

If one wants to use the same optimization methods, but a different cost function, then the subclass only needs to have a new cost_fxn method. See AffineOptimizerDisplacement for an example implementing a new cost function

Major overhauls are possible too. See AffineOptimizerMattesMI for an example on using SimpleITK’s optimization methods inside of an AffineOptimizer subclass

If the optimizer uses corressponding points, then the class attribute accepts_xy needs to be set to True. The default is False.

__init__(nlevels=1, nbins=256, optimization='Powell', transformation='EuclideanTransform')[source]

AffineOptimizer registers moving and fixed images by minimizing a cost function

Parameters
  • nlevels (int) – Number of levels in the Gaussian pyramid

  • nbins (int) – Number of bins to have in histograms used to estimate mutual information

  • optimization (str) – Optimization method. Can be any method from scipy.optimize

  • transformation (str) – Type of transformation, “EuclideanTransform” or “SimilarityTransform”

align(moving, fixed, mask, initial_M=None, moving_xy=None, fixed_xy=None)[source]

Align images by minimizing self.cost_fxn. Aligns each level of the Gaussian pyramid, and uses previous transform as the initial guess in the next round of optimization. Also uses other “good” estimates to define the parameter boundaries.

Parameters
  • moving (ndarray) – Image to warp to align with fixed

  • fixed (ndarray) – Image moving is warped to align with

  • mask (ndarray) – 2D array having non-zero pixel values, where values of 0 are ignnored during registration

  • initial_M ((3x3) array) – Initial transformation matrix

  • moving_xy (ndarray, optional) – (N, 2) array containing points in the moving image that correspond to those in the fixed image

  • fixed_xy (ndarray, optional) – (N, 2) array containing points in the fixed image that correspond to those in the moving image

Returns

  • aligned ((N,M) array) – Moving image warped to align with the fixed image

  • M ((3,3) array) – Optimal transformation matrix

  • cost_list (list) – list containing the minimized cost for each level in the pyramid

AffineOptimizerMattesMI

class valis.affine_optimizer.AffineOptimizerMattesMI(nlevels=4.0, nbins=32, optimization='AdaptiveStochasticGradientDescent', transform='EuclideanTransform')[source]

Bases: AffineOptimizer

Optimize rigid registration using Simple ITK

AffineOptimizerMattesMI is an AffineOptimizer subclass that uses simple ITK’s AdvancedMattesMutualInformation. If moving_xy and fixed_xy are also provided, then Mattes mutual information will be maximized, while the distance between moving_xy and fixed_xy will be minimized (the CorrespondingPointsEuclideanDistanceMetric in Simple ITK).

nlevels

Number of levels in the Gaussian pyramid

Type

int

nbins

Number of bins to have in histograms used to estimate mutual information

Type

int

transformation

Type of transformation, “EuclideanTransform” or “SimilarityTransform”

Type

str

Reg

sitk.ElastixImageFilter object that will perform the optimization

Type

sitk.ElastixImageFilter

fixed_kp_fname

Name of file where to fixed_xy will be temporarily be written. Eventually deleted

Type

str

moving_kp_fname

Name of file where to moving_xy will be temporarily be written. Eventually deleted

Type

str

setup(moving, fixed, mask, initial_M=None, moving_xy=None, fixed_xy=None)[source]

Create parameter map and initialize Reg

calc_cost(p)[source]

Inherited but not used, returns None

write_elastix_kp(kp, fname)[source]

Temporarily write fixed_xy and moving_xy to file

align(moving, fixed, mask, initial_M=None, moving_xy=None, fixed_xy=None)[source]

Align images by minimizing cost_fxn

__init__(nlevels=4.0, nbins=32, optimization='AdaptiveStochasticGradientDescent', transform='EuclideanTransform')[source]

AffineOptimizer registers moving and fixed images by minimizing a cost function

Parameters
  • nlevels (int) – Number of levels in the Gaussian pyramid

  • nbins (int) – Number of bins to have in histograms used to estimate mutual information

  • optimization (str) – Optimization method. Can be any method from scipy.optimize

  • transformation (str) – Type of transformation, “EuclideanTransform” or “SimilarityTransform”

align(moving, fixed, mask, initial_M=None, moving_xy=None, fixed_xy=None)[source]

Optimize rigid registration

Parameters
  • moving (ndarray) – Image to warp to align with fixed

  • fixed (ndarray) – Image moving is warped to align with

  • mask (ndarray) – 2D array having non-zero pixel values, where values of 0 are ignnored during registration

  • initial_M ((3x3) array) – Initial transformation matrix

  • moving_xy (ndarray, optional) – (N, 2) array containing points in the moving image that correspond to those in the fixed image

  • fixed_xy (ndarray, optional) – (N, 2) array containing points in the fixed image that correspond to those in the moving image

Returns

  • aligned ((N,M) array) – Moving image warped to align with the fixed image

  • M ((3,3) array) – Optimal transformation matrix

  • cost_list (None) – None is returned because costs are not recorded