def read_all_channels(im_paths, vox_size=None):
im, im_meta, im_names = [], [], []
for path in im_paths:
filename_pieces = basename(path).split('.')
if 'nii' not in filename_pieces and vox_size is None:
print('ERROR: For non-Nifti images you must specify vox size')
sys.exit()
data, meta = fileio.read_image(abspath(path))
def_meta = fileio.get_default_nifti_header(shape=data.shape,
dtype=data.dtype)
meta = {**def_meta, **meta}
meta['dim'][1:4] = data.shape
if vox_size is not None:
meta['pixdim'][1:4] = [float(x) for x in vox_size.split('x')]
im.append(data)
im_meta.append(meta)
im_names.append(filename_pieces[0])
return im, im_meta, im_names
def read_image(source, *args, **pargs):
return fileio.read_image(source, *args, **pargs)
def read_image( source, *args, **pargs ):
return fileio.read_image( source, *args, **pargs )
def preprocess(args):
"""Also initializing some helpful variables for later"""
im1, im1_vs = args.image1, args.im1_vox_size
im2, im2_vs = args.image2, args.im2_vox_size
im1_list, im1_meta_list, im1_names = read_all_channels(im1, im1_vs)
im2_list, im2_meta_list, im2_names = read_all_channels(im2, im2_vs)
dim = len(im1_list[0].shape) # image dimension
v1 = im1_meta_list[0]['pixdim'][1:dim + 1] # im1 voxel size
v2 = im2_meta_list[0]['pixdim'][1:dim + 1]
im1_brain_mask, im2_brain_mask = None, None
if args.im1_mask is not None:
im1_brain_mask, mask_meta = fileio.read_image(args.im1_mask)
if args.im2_mask is not None:
im2_brain_mask, mask_meta = fileio.read_image(args.im2_mask)
"""All outputs will be at this voxel size, so we can overwrite inputs here
Update meta data voxel size, important to keep track of all spatial changes"""
min_size = np.array([
0.25,
] * dim)
if args.min_vox_size:
min_size = np.array([float(x) for x in args.min_vox_size.split('x')])
for i, (im1, im2) in enumerate(zip(im1_list, im2_list)):
a, b, newvox = pp.normalize_voxelsize(im1,
v1,
im2,
v2,
min_size=min_size)
im1_list[i], im2_list[i], = a, b
if im1_brain_mask is not None:
im1_brain_mask, b, nv = pp.normalize_voxelsize(im1_brain_mask,
v1,
im1_list[0],
newvox,
min_size=min_size)
if im2_brain_mask is not None:
im2_brain_mask, b, nv = pp.normalize_voxelsize(im2_brain_mask,
v2,
im2_list[0],
newvox,
min_size=min_size)
# TODO: normalize_intensity has 'sigmoid' mode, experiment with that
# TODO: consider better options here (sum projection in foreground, mean
# projection in background - using a quick heuristic foreground/
# background threshold)
# Use the very nice two-channel image I stitched for Yu/Sujatha
# to test alternatives. Do a dual channel alignment to the Z-brain.
if args.im1_mask is None:
"""This norm is only for preprocessing purposes, don't overwrite im lists"""
im_list_norm = [pp.normalize_intensity(im) for im in im1_list]
im_sum = np.sum(np.array(im_list_norm), axis=0)
im1_brain_mask = pp.brain_detection(im_sum,
v1,
foreground_mask=None,
iterations=[60, 25, 5],
lambda2=args.im1_lambda)
if args.im2_mask is None:
im_list_norm = [pp.normalize_intensity(im) for im in im2_list]
im_sum = np.sum(np.array(im_list_norm), axis=0)
im2_brain_mask = pp.brain_detection(im_sum,
v2,
foreground_mask=None,
iterations=[60, 25, 5],
lambda2=args.im2_lambda)
"""Padding is after brain detection, so zeros can't affect masking
Masks will be used for registration, and pads are not within mask
so pads shouldn't affect registration either"""
im1_box = pp.minimal_bounding_box(im1_brain_mask)
im2_box = pp.minimal_bounding_box(im2_brain_mask)
im1_box_dims = np.array([a.stop - a.start for a in im1_box])
im2_box_dims = np.array([a.stop - a.start for a in im2_box])
final_dims = np.maximum(im1_box_dims, im2_box_dims) + 2 * args.pad_size
diff1, diff2 = (final_dims - im1_box_dims) / 2., (final_dims -
im2_box_dims) / 2.
pad1 = np.array([(np.ceil(d), np.floor(d)) for d in diff1]).astype(int)
pad2 = np.array([(np.ceil(d), np.floor(d)) for d in diff2]).astype(int)
for i, (im1, im2) in enumerate(zip(im1_list, im2_list)):
im1_list[i] = np.pad(im1[im1_box], pad1, mode='constant')
im2_list[i] = np.pad(im2[im2_box], pad2, mode='constant')
im1_brain_mask = np.pad(im1_brain_mask[im1_box], pad1, mode='constant')
im2_brain_mask = np.pad(im2_brain_mask[im2_box], pad2, mode='constant')
"""Preprocessing changes spatial relationship with raw image,
which will be recorded in sform; ITK uses qform, and this
which must be zeros for registration to execute accurately"""
xyz = ['x', 'y', 'z']
im1_origin = np.array(
[a.start - np.ceil(b) for a, b in zip(im1_box, diff1)]) * newvox
im2_origin = np.array(
[a.start - np.ceil(b) for a, b in zip(im2_box, diff2)]) * newvox
for meta1, meta2 in zip(im1_meta_list, im2_meta_list):
meta1['sform_code'], meta1['qform_code'] = 2, 1
meta2['sform_code'], meta2['qform_code'] = 2, 1
meta1['pixdim'][1:dim + 1], meta1['dim'][1:dim +
1] = newvox, final_dims
meta2['pixdim'][1:dim + 1], meta2['dim'][1:dim +
1] = newvox, final_dims
for i, (a, b, c) in enumerate(zip(xyz, im1_origin, im2_origin)):
meta1['srow_' + a][i], meta2['srow_' + a][i] = newvox[i], newvox[i]
meta1['srow_' + a][-1], meta2['srow_' + a][-1] = b, c
meta1['qoffset_' + a], meta2['qoffset_' + a] = 0, 0
if all([meta1['quatern_' + a] == 0 for a in ['b', 'c', 'd']]):
meta1['quatern_d'] = 1
if all([meta2['quatern_' + a] == 0 for a in ['b', 'c', 'd']]):
meta2['quatern_d'] = 1
sfx = '_regiprep_mask.nii.gz'
fileio.write_image(args.outdir + '/' + im1_names[0] + sfx, im1_brain_mask,
im1_meta_list[0])
fileio.write_image(args.outdir + '/' + im2_names[0] + sfx, im2_brain_mask,
im2_meta_list[0])
sfx = '_regiprep_pp.nii.gz'
for i in range(len(im1_list)):
fileio.write_image(args.outdir + '/' + im1_names[i] + sfx, im1_list[i],
im1_meta_list[i])
fileio.write_image(args.outdir + '/' + im2_names[i] + sfx, im2_list[i],
im2_meta_list[i])