def pre_processing(fname_target,
fname_sc_seg,
fname_level=None,
fname_manual_gmseg=None,
new_res=0.3,
square_size_size_mm=22.5,
denoising=True,
verbose=1,
rm_tmp=True,
for_model=False):
printv('\nPre-process data...', verbose, 'normal')
tmp_dir = sct.tmp_create()
sct.copy(fname_target, tmp_dir)
fname_target = ''.join(extract_fname(fname_target)[1:])
sct.copy(fname_sc_seg, tmp_dir)
fname_sc_seg = ''.join(extract_fname(fname_sc_seg)[1:])
curdir = os.getcwd()
os.chdir(tmp_dir)
original_info = {
'orientation': None,
'im_sc_seg_rpi': None,
'interpolated_images': []
}
im_target = Image(fname_target).copy()
im_sc_seg = Image(fname_sc_seg).copy()
# get original orientation
printv(' Reorient...', verbose, 'normal')
original_info['orientation'] = im_target.orientation
# assert images are in the same orientation
assert im_target.orientation == im_sc_seg.orientation, "ERROR: the image to segment and it's SC segmentation are not in the same orientation"
im_target_rpi = im_target.copy().change_orientation(
'RPI', generate_path=True).save()
im_sc_seg_rpi = im_sc_seg.copy().change_orientation(
'RPI', generate_path=True).save()
original_info['im_sc_seg_rpi'] = im_sc_seg_rpi.copy(
) # target image in RPI will be used to post-process segmentations
# denoise using P. Coupe non local means algorithm (see [Manjon et al. JMRI 2010]) implemented in dipy
if denoising:
printv(' Denoise...', verbose, 'normal')
# crop image before denoising to fasten denoising
nx, ny, nz, nt, px, py, pz, pt = im_target_rpi.dim
size_x, size_y = (square_size_size_mm + 1) / px, (square_size_size_mm +
1) / py
size = int(np.ceil(max(size_x, size_y)))
# create mask
fname_mask = 'mask_pre_crop.nii.gz'
sct_create_mask.main([
'-i', im_target_rpi.absolutepath, '-p',
'centerline,' + im_sc_seg_rpi.absolutepath, '-f', 'box', '-size',
str(size), '-o', fname_mask
])
# crop image
cropper = ImageCropper(im_target_rpi)
cropper.get_bbox_from_mask(Image(fname_mask))
im_target_rpi_crop = cropper.crop()
# crop segmentation
cropper = ImageCropper(im_sc_seg_rpi)
cropper.get_bbox_from_mask(Image(fname_mask))
im_sc_seg_rpi_crop = cropper.crop()
# denoising
from sct_maths import denoise_nlmeans
block_radius = 3
block_radius = int(
im_target_rpi_crop.data.shape[2] /
2) if im_target_rpi_crop.data.shape[2] < (block_radius *
2) else block_radius
patch_radius = block_radius - 1
data_denoised = denoise_nlmeans(im_target_rpi_crop.data,
block_radius=block_radius,
patch_radius=patch_radius)
im_target_rpi_crop.data = data_denoised
im_target_rpi = im_target_rpi_crop
im_sc_seg_rpi = im_sc_seg_rpi_crop
else:
fname_mask = None
# interpolate image to reference square image (resample and square crop centered on SC)
printv(' Interpolate data to the model space...', verbose, 'normal')
list_im_slices = interpolate_im_to_ref(im_target_rpi,
im_sc_seg_rpi,
new_res=new_res,
sq_size_size_mm=square_size_size_mm)
original_info[
'interpolated_images'] = list_im_slices # list of images (not Slice() objects)
printv(' Mask data using the spinal cord segmentation...', verbose,
'normal')
list_sc_seg_slices = interpolate_im_to_ref(
im_sc_seg_rpi,
im_sc_seg_rpi,
new_res=new_res,
sq_size_size_mm=square_size_size_mm,
interpolation_mode=1)
for i in range(len(list_im_slices)):
# list_im_slices[i].data[list_sc_seg_slices[i].data == 0] = 0
list_sc_seg_slices[i] = binarize(list_sc_seg_slices[i],
thr_min=0.5,
thr_max=1)
list_im_slices[
i].data = list_im_slices[i].data * list_sc_seg_slices[i].data
printv(' Split along rostro-caudal direction...', verbose, 'normal')
list_slices_target = [
Slice(slice_id=i, im=im_slice.data, gm_seg=[], wm_seg=[])
for i, im_slice in enumerate(list_im_slices)
]
# load vertebral levels
if fname_level is not None:
printv(' Load vertebral levels...', verbose, 'normal')
# copy level file to tmp dir
os.chdir(curdir)
sct.copy(fname_level, tmp_dir)
os.chdir(tmp_dir)
# change fname level to only file name (path = tmp dir now)
fname_level = ''.join(extract_fname(fname_level)[1:])
# load levels
list_slices_target = load_level(list_slices_target, fname_level)
os.chdir(curdir)
# load manual gmseg if there is one (model data)
if fname_manual_gmseg is not None:
printv('\n\tLoad manual GM segmentation(s) ...', verbose, 'normal')
list_slices_target = load_manual_gmseg(list_slices_target,
fname_manual_gmseg,
tmp_dir,
im_sc_seg_rpi,
new_res,
square_size_size_mm,
for_model=for_model,
fname_mask=fname_mask)
if rm_tmp:
# remove tmp folder
sct.rmtree(tmp_dir)
return list_slices_target, original_info
def pre_processing(fname_target, fname_sc_seg, fname_level=None, fname_manual_gmseg=None, new_res=0.3, square_size_size_mm=22.5, denoising=True, verbose=1, rm_tmp=True, for_model=False):
printv('\nPre-process data...', verbose, 'normal')
tmp_dir = sct.tmp_create()
sct.copy(fname_target, tmp_dir)
fname_target = ''.join(extract_fname(fname_target)[1:])
sct.copy(fname_sc_seg, tmp_dir)
fname_sc_seg = ''.join(extract_fname(fname_sc_seg)[1:])
curdir = os.getcwd()
os.chdir(tmp_dir)
original_info = {'orientation': None, 'im_sc_seg_rpi': None, 'interpolated_images': []}
im_target = Image(fname_target).copy()
im_sc_seg = Image(fname_sc_seg).copy()
# get original orientation
printv(' Reorient...', verbose, 'normal')
original_info['orientation'] = im_target.orientation
# assert images are in the same orientation
assert im_target.orientation == im_sc_seg.orientation, "ERROR: the image to segment and it's SC segmentation are not in the same orientation"
im_target_rpi = im_target.copy().change_orientation('RPI', generate_path=True).save()
im_sc_seg_rpi = im_sc_seg.copy().change_orientation('RPI', generate_path=True).save()
original_info['im_sc_seg_rpi'] = im_sc_seg_rpi.copy() # target image in RPI will be used to post-process segmentations
# denoise using P. Coupe non local means algorithm (see [Manjon et al. JMRI 2010]) implemented in dipy
if denoising:
printv(' Denoise...', verbose, 'normal')
# crop image before denoising to fasten denoising
nx, ny, nz, nt, px, py, pz, pt = im_target_rpi.dim
size_x, size_y = (square_size_size_mm + 1) / px, (square_size_size_mm + 1) / py
size = int(np.ceil(max(size_x, size_y)))
# create mask
fname_mask = 'mask_pre_crop.nii.gz'
sct_create_mask.main(['-i', im_target_rpi.absolutepath, '-p', 'centerline,' + im_sc_seg_rpi.absolutepath, '-f', 'box', '-size', str(size), '-o', fname_mask])
# crop image
fname_target_crop = add_suffix(im_target_rpi.absolutepath, '_pre_crop')
crop_im = ImageCropper(input_file=im_target_rpi.absolutepath, output_file=fname_target_crop, mask=fname_mask)
im_target_rpi_crop = crop_im.crop()
# crop segmentation
fname_sc_seg_crop = add_suffix(im_sc_seg_rpi.absolutepath, '_pre_crop')
crop_sc_seg = ImageCropper(input_file=im_sc_seg_rpi.absolutepath, output_file=fname_sc_seg_crop, mask=fname_mask)
im_sc_seg_rpi_crop = crop_sc_seg.crop()
# denoising
from sct_maths import denoise_nlmeans
block_radius = 3
block_radius = int(im_target_rpi_crop.data.shape[2] / 2) if im_target_rpi_crop.data.shape[2] < (block_radius*2) else block_radius
patch_radius = block_radius -1
data_denoised = denoise_nlmeans(im_target_rpi_crop.data, block_radius=block_radius, patch_radius=patch_radius)
im_target_rpi_crop.data = data_denoised
im_target_rpi = im_target_rpi_crop
im_sc_seg_rpi = im_sc_seg_rpi_crop
else:
fname_mask = None
# interpolate image to reference square image (resample and square crop centered on SC)
printv(' Interpolate data to the model space...', verbose, 'normal')
list_im_slices = interpolate_im_to_ref(im_target_rpi, im_sc_seg_rpi, new_res=new_res, sq_size_size_mm=square_size_size_mm)
original_info['interpolated_images'] = list_im_slices # list of images (not Slice() objects)
printv(' Mask data using the spinal cord segmentation...', verbose, 'normal')
list_sc_seg_slices = interpolate_im_to_ref(im_sc_seg_rpi, im_sc_seg_rpi, new_res=new_res, sq_size_size_mm=square_size_size_mm, interpolation_mode=1)
for i in range(len(list_im_slices)):
# list_im_slices[i].data[list_sc_seg_slices[i].data == 0] = 0
list_sc_seg_slices[i] = binarize(list_sc_seg_slices[i], thr_min=0.5, thr_max=1)
list_im_slices[i].data = list_im_slices[i].data * list_sc_seg_slices[i].data
printv(' Split along rostro-caudal direction...', verbose, 'normal')
list_slices_target = [Slice(slice_id=i, im=im_slice.data, gm_seg=[], wm_seg=[]) for i, im_slice in enumerate(list_im_slices)]
# load vertebral levels
if fname_level is not None:
printv(' Load vertebral levels...', verbose, 'normal')
# copy level file to tmp dir
os.chdir(curdir)
sct.copy(fname_level, tmp_dir)
os.chdir(tmp_dir)
# change fname level to only file name (path = tmp dir now)
fname_level = ''.join(extract_fname(fname_level)[1:])
# load levels
list_slices_target = load_level(list_slices_target, fname_level)
os.chdir(curdir)
# load manual gmseg if there is one (model data)
if fname_manual_gmseg is not None:
printv('\n\tLoad manual GM segmentation(s) ...', verbose, 'normal')
list_slices_target = load_manual_gmseg(list_slices_target, fname_manual_gmseg, tmp_dir, im_sc_seg_rpi, new_res, square_size_size_mm, for_model=for_model, fname_mask=fname_mask)
if rm_tmp:
# remove tmp folder
sct.rmtree(tmp_dir)
return list_slices_target, original_info