Copy headers / metadata between segmentations of different resolutions

Algorithms
, , ,
1

I have a coarse segmentation that I smooth, and the smoothing techniques overwrites the header. I want to retrieve the original one but I have had no success so far. The headers are

_, coarse_header            = nrrd.read(path_to_coarse_image)
smooth_image, smooth_header = nrrd.read(path_to_smooth_image)

coarse_header:

OrderedDict([('type', 'short'),
             ('dimension', 3),
             ('space', 'left-posterior-superior'),
             ('sizes', array([748, 749, 749])),
             ('space directions',
              array([[ 0.26375667, -0.09178596, -0.1095799 ],
                     [ 0.10153998,  0.2821788 ,  0.00804705],
                     [ 0.10060839, -0.04416401,  0.27915496]])),
             ('kinds', ['domain', 'domain', 'domain']),
             ('endian', 'little'),
             ('encoding', 'raw'),
             ('space origin',
              array([-148.47837839,  -45.81777566,  -68.16309759]))])

smooth_header:

OrderedDict([('type', 'uint8'),
             ('dimension', 3),
             ('sizes', array([1314,  884,  900])),
             ('encoding', 'raw'),
             ('spacings', array([0.15, 0.15, 0.15])),
             ('axis mins',
              array([-139.14450073,   -3.77869463,  -26.14629364]))])

Now I don’t care about the information in the smooth_header. The main difference is that it has a higher resolution now, from a voxel spacing of 0.3 now is 0.15.

I figured that the spacing information of the coarse_header is encoded in the norm of the vectors of space_directions, so I copied the the header as follows:

new_smooth_header = dict(coarse_header)

#Update the fields in the new header to match the coarse header, except for the spacing
new_smooth_header['sizes'] = smooth_header['sizes']

smooth_spacing = np.array(smooth_header['spacings'])  # Spacing from smooth_header
coarse_space_directions = np.array(coarse_header['space directions'])  # Space directions from coarse_header

# Calculate new space directions for smooth header
new_smooth_space_directions = np.copy(coarse_space_directions)
for i in range(len(smooth_spacing)):
    norm_coarse = np.linalg.norm(coarse_space_directions[i])
    new_smooth_space_directions[i] = smooth_spacing[i] * (coarse_space_directions[i] / norm_coarse)


new_smooth_header['space directions'] = new_smooth_space_directions

# # Assign the new header to the smooth image
# smooth_image_with_coarse_header = (smooth_image, new_smooth_header)

ordered_new_smooth_header = OrderedDict(new_smooth_header.items())

nrrd.write(path_to_save_image, smooth_image, header=ordered_new_smooth_header)

Now the new header is

OrderedDict([('type', 'uint8'),
             ('dimension', 3),
             ('space', 'left-posterior-superior'),
             ('sizes', array([1314,  884,  900])),
             ('space directions',
              array([[ 0.13187834, -0.04589298, -0.05478995],
                     [ 0.05076999,  0.1410894 ,  0.00402352],
                     [ 0.0503042 , -0.02208201,  0.13957748]])),
             ('kinds', ['domain', 'domain', 'domain']),
             ('encoding', 'raw'),
             ('space origin',
              array([-148.47837839,  -45.81777566,  -68.16309759]))])

So in theory they should be in the same space, as far as I understand. They’re not. Size seems fine, as well as orientation, it looks that it just need a translation, but I don’t know which one.

header_problems
header_problems1495×922 105 KB

2

How does the smooth image look like if you don’t fiddle with its header? You seem to be ignoring the origin in the smooth image (axis mins field), maybe you shouldn’t, or you need to use it somehow? The smooth image also has identity orientation, maybe you need to preserve that?

What is lacking is information how your smooth image was produced. That information is needed, to account for its spatial discrepancy (if any).