Customizing the coregistration

By default, dcmri uses free-form deformation implemented in the package itk.elastix, which default settings for all configuration parameters.

This example shows how thse defaults can be modified, and how coregistration can be done by another package, skimage.

Import packages and data

Import packages

import numpy as np
import mdreg

Load test data

data = mdreg.fetch('MOLLI')
array = data['array'][:,:,0,:]

# Throughout this example we use the same signal model:
molli = {
    'func': mdreg.abs_exp_recovery_2p,
    'TI': np.array(data['TI'])/1000,
}

# Visualise the data
mdreg.animation(array, vmin=0, vmax=1e4, show=True)

Once Loop Reflect

<matplotlib.animation.ArtistAnimation object at 0x0000021EB9CE6110>

Customizing the elastix coregistration

By default fit performs image coregistration using the package elastix with bspline deformations, and default settings for all parameters. If the result is suboptimal, one way to improve is to customize the coregistration model in elastix.

A critical parameter in elastix bspline deformation is the grid spacing, which determines the level of detail in the deformation field. The default coregistration uses a grid spacing of 16mm. Let’s try what happens if we allow finer deformations of 5mm. In order to do that, we need to provide the actual pixel spacing of the data, and modify the default elastix parameter:

deform5mm = {

    # Pixel spacing in the images (in mm)
    'spacing': data['pixel_spacing'],

    # Provide custom grid spacing (in mm)
    'FinalGridSpacingInPhysicalUnits': 5.0,
}

We run mdreg using the 5mm coregistration:

coreg, defo, fit, pars = mdreg.fit(
    array, fit_image=molli, fit_coreg=deform5mm)

# Visualise the results
mdreg.plot_series(array, fit, coreg, vmin=0, vmax=1e4, show=True)

Once Loop Reflect

<matplotlib.animation.ArtistAnimation object at 0x0000021EBC273AC0>

Compared to results in other examples using the default grid spacing of 16mm, this shows smaller deformations in the upper pole of the right kidney that are not present in the original moving images. The example illustrates that the grid spacing should be chosen carefully to reflect the scale of the expected deformations.

Any coregistration method available in elastix can be applied in the same way by providing a custom set of elastix parameters.

Coregistration with skimage

While elastix is the default package for coregistration, mdreg also has an option to use the optical flow method optical_flow_tvl1() from the package skimage.

For this illustration we run skimage coregistration with default parameters, except for the attachment which is increased to 30 (default=15) to allow for finer deformations:

coreg_skimg = {

    # The package needs to be defined if it is not elastix
    'package': 'skimage',

    # Provide a custom attachment value
    'attachment': 30,
}

Run mdreg again, now using the skimage coregistration instead of elastix:

# Perform model-driven coregistration
coreg, defo, fit, pars = mdreg.fit(
    array, fit_image=molli, fit_coreg=coreg_skimg)

# Visualise the results
mdreg.plot_series(array, fit, coreg, vmin=0, vmax=1e4, show=True)

Once Loop Reflect

<matplotlib.animation.ArtistAnimation object at 0x0000021EBBF21B40>

This result shows good coregistration results, nicely preserving even fine grained features such as the small kidney cysts.