Merge pull request #24 from BrainLesion/default_from_original

Updated default parameters from original repo

data/default_rigid.yaml

@@ -0,0 +1,80 @@
+# Whether to use bias correction before registration (helpful to generate more accurate transforms at the cost of time). Default: false
+bias: false
+
+# The registration metric. Options: ["mattes_mutual_information", "ants_neighborhood_correlation", "correlation", "demons", "joint_histogram_mutual_information", "mean_squares"]
+# this can be a list as well, which will then become a multi-metric registration using composite transforms [https://simpleitk.org/doxygen/latest/html/classitk_1_1simple_1_1CompositeTransform.html]
+metric: mean_squares
+
+## metric-specific parameters
+metric_parameters: {
+  histogram_bins: 50, # mattes_mutual_information, joint_histogram_mutual_information
+  radius: 5, # ants_neighborhood_correlation
+  intensityDifferenceThreshold: 0.001, # demons
+  varianceForJointPDFSmoothing: 1.5, # joint_histogram_mutual_information
+}
+
+# Optimizer. Options: ["gradient_descent", "regular_step_gradient_descent", "amoeba", "conjugate", "exhaustive", "gadient_descent_line_search", "lbfgsb", "lbfgsb2", "one_plus_one_evolutionary", "powell"]
+optimizer: regular_step_gradient_descent
+
+## optimizer-specific parameters
+optimizer_parameters: {
+  min_step: 1e-4, # regular_step_gradient_descent
+  max_step: 1.0, # gradient_descent, regular_step_gradient_descent
+  iterations: 200, # regular_step_gradient_descent, gradient_descent_line_search, gradient_descent, conjugate, lbfgsb, lbfgsb2
+  relaxation: 0.1, # regular_step_gradient_descent
+  learningrate: 1.0, # gradient_descent, gradient_descent_line_search
+  tolerance: 1e-6, # gradient_descent, gradient_descent_line_search
+  convergence_minimum: 1e-6, # gradient_descent, gradient_descent_line_search
+  convergence_window_size: 10, # gradient_descent, gradient_descent_line_search
+  line_search_lower_limit: 0.0, # gradient_descent_line_search
+  line_search_upper_limit: 5.0, # gradient_descent_line_search
+  line_search_epsilon: 0.01, # gradient_descent_line_search
+  step_length: 0.1, # conjugate, exhaustive, powell
+  simplex_delta: 0.1, # amoeba
+  maximum_number_of_corrections: 5, # lbfgsb, lbfgsb2
+  maximum_number_of_function_evaluations: 2000, # lbfgsb, lbfgsb2
+  solution_accuracy: 1e-5, # lbfgsb2
+  hessian_approximate_accuracy: 1e-5, # lbfgsb2
+  delta_convergence_distance: 1e-5, # lbfgsb2
+  delta_convergence_tolerance: 1e-5, # lbfgsb2
+  line_search_maximum_evaluations: 50, # lbfgsb2
+  line_search_minimum_step: 1e-20, # lbfgsb2
+  line_search_accuracy: 1e-4, # lbfgsb2
+  epsilon: 1e-8, # one_plus_one_evolutionary
+  initial_radius: 1.0, # one_plus_one_evolutionary
+  growth_factor: -1.0, # one_plus_one_evolutionary
+  shrink_factor: -1.0, # one_plus_one_evolutionary
+  maximum_line_iterations: 100, # powell
+  step_tolerance: 1e-6, # powell
+  value_tolerance: 1e-6, # powell
+}
+
+# The registration transform. Options: ["translation", "versor", "versor_rigid", "euler", "similarity", "scale", "scale_versor", "scale_skew_versor", "affine", "bspline", "displacement"]
+transform: versorrigid
+
+# Composite transform
+composite_transform: false
+
+# Previous transforms saved to disk: only used if composite_transform is true
+previous_transforms: []
+
+# Transform initialization. Options: ["moments", "geometry", "selfmoments", "selfgeometry"]
+initialization: geometry
+
+# Interpolator. Options: ["linear", "bspline", "nearestneighbor", "gaussian", "labelgaussian"]
+interpolator: linear
+
+# Sampling strategy. Options: ["regular", "random", "none"]
+sampling_strategy: "random"
+
+# Sampling percentage. Can be a list of percentages with the same length as the number of levels.
+sampling_percentage: 0.01
+
+# Shrink factor at each level for pyramid registration.
+shrink_factors: [4, 2, 1]
+
+# Smoothing sigma at each level for pyramid registration.
+smoothing_sigmas: [2, 1, 0]
+
+# The number of attempts to try to find a good registration (useful when using random sampling)
+attempts: 5

data/sample_config.yaml

@@ -20,9 +20,10 @@ optimizer: regular_step_gradient_descent
 optimizer_parameters: {
   min_step: 1e-6, # regular_step_gradient_descent
   max_step: 1.0, # gradient_descent, regular_step_gradient_descent
-  maximumStepSizeInPhysicalUnits: 1.0, # regular_step_gradient_descent, gradient_descent_line_search, gradient_descent,
   iterations: 1000, # regular_step_gradient_descent, gradient_descent_line_search, gradient_descent, conjugate, lbfgsb, lbfgsb2
+  relaxation: 0.5, # regular_step_gradient_descent
   learningrate: 1.0, # gradient_descent, gradient_descent_line_search
+  tolerance: 1e-4, # gradient_descent, gradient_descent_line_search
   convergence_minimum: 1e-6, # gradient_descent, gradient_descent_line_search
   convergence_window_size: 10, # gradient_descent, gradient_descent_line_search
   line_search_lower_limit: 0.0, # gradient_descent_line_search
@@ -69,18 +70,6 @@ sampling_strategy: none
 # Sampling percentage. Can be a list of percentages with the same length as the number of levels.
 sampling_percentage: 0.5
 
-# Registration relaxation factor.
-relaxation: 0.5
-
-# Registration gradient tolerance
-tolerance: 1e-4
-
-# Maximum number of iterations at each level.
-max_step: 5.0
-
-# Minimum step size at each level.
-min_step: 1e-5
-
 # Shrink factor at each level for pyramid registration.
 shrink_factors: [2, 1]
 
@@ -89,6 +78,3 @@ smoothing_sigmas: [1, 0]
 
 # The number of attempts to try to find a good registration (useful when using random sampling)
 attempts: 5
-
-# Total number of iterations.
-iterations: 1000

ereg/registration.py

@@ -123,11 +123,6 @@ def update_parameters(self, config_file: Union[str, dict], **kwargs):
         self.parameters["initialization"] = self.parameters.get(
             "initialization", "geometry"
         ).lower()
-        self.parameters["max_step"] = self.parameters.get("max_step", 5.0)
-        self.parameters["min_step"] = self.parameters.get("min_step", 0.01)
-        self.parameters["iterations"] = self.parameters.get("iterations", 200)
-        self.parameters["relaxation"] = self.parameters.get("relaxation", 0.5)
-        self.parameters["tolerance"] = self.parameters.get("tolerance", 1e-4)
         self.parameters["bias_correct"] = self.parameters.get(
             "bias_correct", self.parameters.get("bias", False)
         )
@@ -738,14 +733,15 @@ def _register_image_and_get_transform(
             )
 
         # R.SetOptimizerScalesFromJacobian()
-        # R.SetOptimizerScalesFromPhysicalShift()
+        R.SetOptimizerScalesFromPhysicalShift()
 
         R.SetShrinkFactorsPerLevel(self.parameters["shrink_factors"])
         R.SetSmoothingSigmasPerLevel(self.parameters["smoothing_sigmas"])
         R.SmoothingSigmasAreSpecifiedInPhysicalUnitsOn()
         transform_function = self._get_transform_wrapper(
             self.parameters["transform"], dimension
         )
+        ## todo: evaluate the viability of having default options for "rigid", "affine", and "deformable" registrations
         # rigid_registration = False
         # # euler transforms need special processing
         # if isinstance(transform_function, sitk.Euler3DTransform) or isinstance(
@@ -792,6 +788,7 @@ def _register_image_and_get_transform(
             )
         R.SetInitialTransform(final_transform, inPlace=False)
         ## set the interpolator - all options: https://simpleitk.org/doxygen/latest/html/namespaceitk_1_1simple.html#a7cb1ef8bd02c669c02ea2f9f5aa374e5
+        # this should be linear to optimize results and computational efficacy
         R.SetInterpolator(sitk.sitkLinear)
 
         # R.AddCommand(sitk.sitkIterationEvent, lambda: R)
@@ -811,25 +808,24 @@ def _register_image_and_get_transform(
             raise RuntimeError("Registration failed.")
 
         registration_transform_sitk = output_transform
-        if "rigid_registration" in self.parameters:
-            if self.parameters["rigid_registration"]:
-                try:
-                    # Euler Transform used:
-                    registration_transform_sitk = eval(
-                        "sitk.Euler%dDTransform(registration_transform_sitk)"
-                        % (dimension)
-                    )
-                except:
-                    # VersorRigid used: Transform from VersorRigid to Euler
-                    registration_transform_sitk = eval(
-                        "sitk.VersorRigid%dDTransform(registration_transform_sitk)"
-                        % (dimension)
-                    )
-                    tmp = eval("sitk.Euler%dDTransform()" % (dimension))
-                    tmp.SetMatrix(registration_transform_sitk.GetMatrix())
-                    tmp.SetTranslation(registration_transform_sitk.GetTranslation())
-                    tmp.SetCenter(registration_transform_sitk.GetCenter())
-                    registration_transform_sitk = tmp
+        # if user is requesting a rigid registration, convert the transform to a rigid transform
+        if self.parameters["transform"] in ["euler", "versorrigid"]:
+            try:
+                # Euler Transform used:
+                registration_transform_sitk = eval(
+                    "sitk.Euler%dDTransform(registration_transform_sitk)" % (dimension)
+                )
+            except:
+                # VersorRigid used: Transform from VersorRigid to Euler
+                registration_transform_sitk = eval(
+                    "sitk.VersorRigid%dDTransform(registration_transform_sitk)"
+                    % (dimension)
+                )
+                tmp = eval("sitk.Euler%dDTransform()" % (dimension))
+                tmp.SetMatrix(registration_transform_sitk.GetMatrix())
+                tmp.SetTranslation(registration_transform_sitk.GetTranslation())
+                tmp.SetCenter(registration_transform_sitk.GetCenter())
+                registration_transform_sitk = tmp
         ## additional information
         # print("Metric: ", R.MetricEvaluate(target_image, moving_image), flush=True)
         # print(