Optimizer's stopping condition is rounding error. Registration framework

Algorithms
, ,
1

Hi everyone,

I’m doing B-Spline registration of CT vs CBCT images.

The Registration method Im using is:

 registration_method = sitk.ImageRegistrationMethod()
  
 grid_physical_spacing = [15.0]*3 
 image_physical_size = [size*spacing for size,spacing in zip(fixed_image.GetSize(), fixed_image.GetSpacing())]
 mesh_size = [int(image_size/grid_spacing + 0.5) \
              for image_size,grid_spacing in zip(image_physical_size,grid_physical_spacing)]

 initial_transform = sitk.BSplineTransformInitializer(image1 = fixed_image, 
                                                      transformDomainMeshSize = mesh_size, order=3)    

 registration_method.SetInitialTransformAsBSpline(initial_transform,
                                                  inPlace=False)

 registration_method.SetMetricAsMattesMutualInformation(numberOfHistogramBins=150)


 registration_method.SetMetricSamplingStrategy(registration_method.RANDOM)
 registration_method.SetMetricSamplingPercentage(0.001)



 registration_method.SetInterpolator(sitk.sitkLinear)
 registration_method.SetOptimizerAsLBFGS2(solutionAccuracy=1e-10, numberOfIterations=150, deltaConvergenceTolerance=1e-10)

The Optimizer’s stopping condition is: A rounding error occurred

I would like to understand why the optimizer always stop for a rounding error and how to stop this to happen.

Thanks in advance,
Luciano

2

This is very strict. The usual values are close to the image spacing, .e.g 0.01 or 0.001. That might be the reason.

3

To the image dimension in which dimension, meters?

For which parameter is the suggestion? solutionAccuracy and deltaConvergenceTolerance?

Thanks,
Luciano

4

ITK is unitless. ITK uses whichever unit your image spacing is expressed in. That is usually millimeters in medicine, meters in earth observation satellite images, milli-arc-seconds in astronomy (I think).

Probably both parameters need to be much larger.

5

I have the similar issue in multiscale mode. Following the instructions here:
https://insightsoftwareconsortium.github.io/SimpleITK-Notebooks/Python_html/65_Registration_FFD.html#:~:text=registration_method.SetOptimizerAsLBFGS2(solutionAccuracy%3D1e-2%2C%20numberOfIterations%3D100%2C%20deltaConvergenceTolerance%3D0.01)

registration_method.SetOptimizerAsLBFGS2(
            solutionAccuracy=1e-2,
            # hessianApproximateAccuracy=60,
            numberOfIterations=1000,
            deltaConvergenceTolerance=10,
            # lineSearchMaximumEvaluations=40,
            # lineSearchMaximumStep=1e20,
            # lineSearchMinimumStep=1e-20,
            # lineSearchAccuracy=1e-4
        )

The fixed and moving image sizeare: (96, 72) (57, 86)
The metric curve looks like:
image