3D translation: image to mask

This example illustrates coregistration by 3D translation. The moving volume is a 3D magnetization transfer map and the static volume is a 3D mask covering both kidneys.

Coregistration is performed by brute force optimization using a mutual information metric.

Setup

import numpy as np
import vreg
import vreg.plot as plt

Get data

# Static oblique volumes
lk = vreg.fetch('left_kidney')
rk = vreg.fetch('right_kidney')

# Moving volume
mtr = vreg.fetch('MTR')

# Geometrical reference
dixon = vreg.fetch('Dixon_water')

Format data

Create a mask containing both kidneys (bk) with the geometry of the complete DIXON series

bk = lk.slice_like(dixon).add(rk)

Bounding box

Extract a bounding box to reduce the size of the volume. This is not necessary but it speeds up the calculation a little as the volume is smaller.

bk = bk.bounding_box()

Overlay data before registration

If we overlay the mask on the volume, we clearly see the misalignment due to different breath holding positions:

plt.overlay_2d(mtr, bk, title='Before 3D translation',
               vmin=np.percentile(mtr.values, 10),
               vmax=np.percentile(mtr.values, 99))

Coregister

We are coregistering using a 3D translation in the reference frame of the moving volume. We are using a brute force optimization which is slow but robust. We allow for translations between [-20, 20] mm in-slice, and [-5, 5] mm through-slice, in steps of 2mm.

optimizer = {
    'method': 'brute',
    'grid': (
        [-20, 20, 20],
        [-20, 20, 20],
        [-5, 5, 5],
    ),
}
params = mtr.find_translate_to(bk, optimizer=optimizer, coords='volume')
mtr = mtr.translate(params, coords='volume')

Overlay data after registration

If we overlay the mask on the new volume, we can see that the misalignment is significantly reduced:

plt.overlay_2d(mtr, bk, title='After 3D translation',
               vmin=np.percentile(mtr.values, 10),
               vmax=np.percentile(mtr.values, 99))