def process(image, brain_mask, args, logger):
img = io.open_nii(image)
mask = io.open_nii(brain_mask)
dirname, base, ext = io.split_filename(image)
if args.output_dir is not None:
dirname = args.output_dir
if not os.path.exists(dirname):
logger.info('Making output directory: {}'.format(dirname))
os.mkdir(dirname)
if args.find_background_mask:
bg_mask = background_mask(img)
bgfile = os.path.join(dirname, base + '_bgmask' + ext)
io.save_nii(bg_mask, bgfile, is_nii=True)
if args.wm_peak is not None:
logger.info('Loading WM peak: ', args.wm_peak)
peak = float(np.load(args.wm_peak))
else:
peak = gmm_class_mask(img, brain_mask=mask, contrast=args.contrast)
if args.save_wm_peak:
np.save(os.path.join(dirname, base + '_wmpeak.npy'), peak)
normalized = gmm.gmm_normalize(img, mask, args.norm_value, args.contrast,
args.background_mask, peak)
outfile = os.path.join(dirname, base + '_gmm' + ext)
logger.info('Normalized image saved: {}'.format(outfile))
io.save_nii(normalized, outfile, is_nii=True)
def run_intensity_fcm(infile, outfolder):
from intensity_normalization.normalize import fcm
from intensity_normalization.utilities import io
try:
if not exists(join(outfolder, "Robex")):
makedirs(join(outfolder, "Robex"))
if not exists(join(outfolder, "wm_masks")):
makedirs(join(outfolder, "wm_masks"))
filename = infile.split(sep)[-1].split(".")[0]
i = io.open_nii(infile)
b_mask = io.open_nii(
join(outfolder, "robex_masks", filename + "_mask.nii.gz"))
wm_mask = fcm.find_tissue_mask(i, b_mask)
normalized = fcm.fcm_normalize(i, wm_mask)
io.save_nii(wm_mask,
join(outfolder, "wm_masks", filename + "_wmmask.nii.gz"))
io.save_nii(normalized, join(outfolder, filename + "_fcm.nii.gz"))
shutil.move(join(infile), join(outfolder, "Robex"))
except:
e = sys.exc_info()
print("Error: ", str(e[0]))
print("Error: ", str(e[1]))
print("Error: executing fcm method")
sys.exit(2)
def setUp(self):
wd = os.path.dirname(os.path.abspath(__file__))
self.data_dir = os.path.join(wd, 'test_data', 'images')
self.mask_dir = os.path.join(wd, 'test_data', 'masks')
self.img = io.open_nii(os.path.join(self.data_dir, 'test.nii.gz'))
self.brain_mask = io.open_nii(
os.path.join(self.mask_dir, 'mask.nii.gz'))
def train(img_fns, mask_fns=None, i_min=1, i_max=99, i_s_min=1, i_s_max=100, l_percentile=10, u_percentile=90, step=10):
"""
determine the standard scale for the set of images
Args:
img_fns (list): set of NifTI MR image paths which are to be normalized
mask_fns (list): set of corresponding masks (if not provided, estimated)
i_min (float): minimum percentile to consider in the images
i_max (float): maximum percentile to consider in the images
i_s_min (float): minimum percentile on the standard scale
i_s_max (float): maximum percentile on the standard scale
l_percentile (int): middle percentile lower bound (e.g., for deciles 10)
u_percentile (int): middle percentile upper bound (e.g., for deciles 90)
step (int): step for middle percentiles (e.g., for deciles 10)
Returns:
standard_scale (np.ndarray): average landmark intensity for images
percs (np.ndarray): array of all percentiles used
"""
mask_fns = [None] * len(img_fns) if mask_fns is None else mask_fns
percs = np.concatenate(([i_min], np.arange(l_percentile, u_percentile+1, step), [i_max]))
standard_scale = np.zeros(len(percs))
for i, (img_fn, mask_fn) in enumerate(zip(img_fns, mask_fns)):
img_data = io.open_nii(img_fn).get_data()
mask = io.open_nii(mask_fn) if mask_fn is not None else None
mask_data = img_data > img_data.mean() if mask is None else mask.get_data()
masked = img_data[mask_data > 0]
landmarks = get_landmarks(masked, percs)
min_p = np.percentile(masked, i_min)
max_p = np.percentile(masked, i_max)
f = interp1d([min_p, max_p], [i_s_min, i_s_max])
landmarks = np.array(f(landmarks))
standard_scale += landmarks
standard_scale = standard_scale / len(img_fns)
return standard_scale, percs
def nyul_normalize(img_dir, mask_dir=None, output_dir=None, standard_hist=None, write_to_disk=True):
"""
Use Nyul and Udupa method ([1,2]) to normalize the intensities of a set of MR images
Args:
img_dir (str): directory containing MR images
img_dir (str): directory containing masks for MR images
output_dir (str): directory to save images if you do not want them saved in
same directory as data_dir
standard_hist (str): path to output or use standard histogram landmarks
write_to_disk (bool): write the normalized data to disk or nah
Returns:
normalized (np.ndarray): last normalized image from img_dir
References:
[1] N. Laszlo G and J. K. Udupa, “On Standardizing the MR Image
Intensity Scale,” Magn. Reson. Med., vol. 42, pp. 1072–1081,
1999.
[2] M. Shah, Y. Xiao, N. Subbanna, S. Francis, D. L. Arnold,
D. L. Collins, and T. Arbel, “Evaluating intensity
normalization on MRIs of human brain with multiple sclerosis,”
Med. Image Anal., vol. 15, no. 2, pp. 267–282, 2011.
"""
input_files = io.glob_nii(img_dir)
if output_dir is None:
out_fns = [None] * len(input_files)
else:
out_fns = []
for fn in input_files:
_, base, ext = io.split_filename(fn)
out_fns.append(os.path.join(output_dir, base + '_hm' + ext))
if not os.path.exists(output_dir):
os.mkdir(output_dir)
mask_files = [None] * len(input_files) if mask_dir is None else io.glob_nii(mask_dir)
if standard_hist is None:
logger.info('Learning standard scale for the set of images')
standard_scale, percs = train(input_files, mask_files)
elif not os.path.isfile(standard_hist):
logger.info('Learning standard scale for the set of images')
standard_scale, percs = train(input_files, mask_files)
np.save(standard_hist, np.vstack((standard_scale, percs)))
else:
logger.info('Loading standard scale ({}) for the set of images'.format(standard_hist))
standard_scale, percs = np.load(standard_hist)
for i, (img_fn, mask_fn, out_fn) in enumerate(zip(input_files, mask_files, out_fns)):
_, base, _ = io.split_filename(img_fn)
logger.info('Transforming image {} to standard scale ({:d}/{:d})'.format(base, i+1, len(input_files)))
img = io.open_nii(img_fn)
mask = io.open_nii(mask_fn) if mask_fn is not None else None
normalized = do_hist_norm(img, percs, standard_scale, mask)
if write_to_disk:
io.save_nii(normalized, out_fn, is_nii=True)
return normalized
def setUp(self):
wd = os.path.dirname(os.path.abspath(__file__))
self.data_dir = os.path.join(wd, 'test_data', 'images')
self.mask_dir = os.path.join(wd, 'test_data', 'masks')
self.img = io.open_nii(os.path.join(self.data_dir, 'test.nii.gz'))
self.brain_mask = io.open_nii(
os.path.join(self.mask_dir, 'mask.nii.gz'))
self.template_mask = os.path.join(self.mask_dir, 'mask.nii.gz')
self.wm_mask = fcm.find_wm_mask(self.img, self.brain_mask)
self.norm_val = 1000
def lsq_normalize(img_dir, mask_dir=None, output_dir=None, write_to_disk=True):
"""
normalize intensities of a set of MR images by minimizing the squared distance
between CSF, GM, and WM means within the set
Args:
img_dir (str): directory containing MR images
mask_dir (str): directory containing masks for MR images
output_dir (str): directory to save images if you do not want them saved in
same directory as data_dir
write_to_disk (bool): write the normalized data to disk or nah
Returns:
normalized (np.ndarray): last normalized image from img_dir
"""
input_files = io.glob_nii(img_dir)
if output_dir is None:
out_fns = [None] * len(input_files)
else:
out_fns = []
for fn in input_files:
_, base, ext = io.split_filename(fn)
out_fns.append(os.path.join(output_dir, base + '_lsq' + ext))
if not os.path.exists(output_dir):
os.mkdir(output_dir)
mask_files = [None] * len(
input_files) if mask_dir is None else io.glob_nii(mask_dir)
standard_tissue_means = None
normalized = None
for i, (img_fn, mask_fn,
out_fn) in enumerate(zip(input_files, mask_files, out_fns)):
_, base, _ = io.split_filename(img_fn)
logger.info(
'Transforming image {} to standard scale ({:d}/{:d})'.format(
base, i + 1, len(input_files)))
img = io.open_nii(img_fn)
mask = io.open_nii(mask_fn) if mask_fn is not None else None
tissue_mem = mask_util.fcm_class_mask(img, mask)
if standard_tissue_means is None:
csf_tissue_mask = find_tissue_mask(img, mask, tissue_type='csf')
csf_normed_data = fcm_normalize(img, csf_tissue_mask).get_fdata()
standard_tissue_means = calc_tissue_means(csf_normed_data,
tissue_mem)
del csf_tissue_mask, csf_normed_data
img_data = img.get_fdata()
tissue_means = calc_tissue_means(img_data, tissue_mem)
sf = find_scaling_factor(tissue_means, standard_tissue_means)
logger.debug('Scaling factor for {}: {:0.3e}'.format(base, sf))
normalized = nib.Nifti1Image(sf * img_data, img.affine, img.header)
if write_to_disk:
io.save_nii(normalized, out_fn, is_nii=True)
return normalized
def main(args=None):
args = arg_parser().parse_args(args)
if args.verbosity == 1:
level = logging.getLevelName('INFO')
elif args.verbosity >= 2:
level = logging.getLevelName('DEBUG')
else:
level = logging.getLevelName('WARNING')
logging.basicConfig(
format='%(asctime)s - %(name)s - %(levelname)s - %(message)s',
level=level)
logger = logging.getLogger(__name__)
try:
if not os.path.isdir(args.img_dir):
raise ValueError(
'(-i / --img-dir) argument needs to be a directory of NIfTI images.'
)
if args.mask_dir is not None:
if not os.path.isdir(args.mask_dir):
raise ValueError(
'(-m / --mask-dir) argument needs to be a directory of NIfTI images.'
)
img_fns = io.glob_nii(args.img_dir)
if args.mask_dir is not None:
mask_fns = io.glob_nii(args.mask_dir)
else:
mask_fns = [None] * len(img_fns)
if not os.path.exists(args.output_dir):
logger.info('Making Output Directory: {}'.format(args.output_dir))
os.mkdir(args.output_dir)
hard_seg = not args.memberships
for i, (img_fn, mask_fn) in enumerate(zip(img_fns, mask_fns), 1):
_, base, _ = io.split_filename(img_fn)
logger.info('Creating Mask for Image: {}, ({:d}/{:d})'.format(
base, i, len(img_fns)))
img = io.open_nii(img_fn)
mask = io.open_nii(mask_fn)
tm = fcm_class_mask(img, mask,
hard_seg) if not args.gmm else gmm_class_mask(
img, mask, 't1', False, hard_seg)
tissue_mask = os.path.join(args.output_dir, base + '_tm')
if args.memberships:
classes = ('csf', 'gm', 'wm')
for j, c in enumerate(classes):
io.save_nii(img, tissue_mask + '_' + c + '.nii.gz', tm[...,
j])
else:
io.save_nii(img, tissue_mask + '.nii.gz', tm)
return 0
except Exception as e:
logger.exception(e)
return 1
def process(image_fn, brain_mask_fn, args, logger):
img = io.open_nii(image_fn)
if args.brain_mask is not None:
mask = io.open_nii(brain_mask_fn)
else:
mask = None
dirname, base, _ = io.split_filename(image_fn)
if args.output_dir is not None:
dirname = args.output_dir
if not os.path.exists(dirname):
logger.info('Making output directory: {}'.format(dirname))
os.mkdir(dirname)
normalized = kde.kde_normalize(img, mask, args.contrast, args.norm_value)
outfile = os.path.join(dirname, base + '_kde.nii.gz')
logger.info('Normalized image saved: {}'.format(outfile))
io.save_nii(normalized, outfile, is_nii=True)
def match_histograms(reference_path, apply_to_path):
print('Start')
for name in ['InPhase', 'OutPhase']:
print(f'Working on: {name}')
ref_files = list()
for folder in os.listdir(reference_path):
ref_files.append(f'{reference_path}/{folder}/{name}.nii')
print('Training')
mask_files = [None] * len(ref_files)
standard_scale, percs = nyul.train(ref_files, mask_files)
input_files, output_files = list(), list()
for folder in os.listdir(apply_to_path):
input_files.append(f'{apply_to_path}/{folder}/{name}.nii')
output_files.append(f'{apply_to_path}/{folder}/{name}_HM.nii')
print('Normalizing')
for img_fn, out_fn in zip(input_files, output_files):
print(img_fn, '->', out_fn)
_, base, _ = io.split_filename(img_fn)
img = io.open_nii(img_fn)
normalized = nyul.do_hist_norm(img,
percs,
standard_scale,
mask=None)
io.save_nii(normalized, out_fn, is_nii=True)
print('Done')
def process(image_fn, brain_mask_fn, output_dir, logger):
img = io.open_nii(image_fn)
dirname, base, _ = io.split_filename(image_fn)
if output_dir is not None:
dirname = output_dir
if not os.path.exists(dirname):
logger.info('Making output directory: {}'.format(dirname))
os.mkdir(dirname)
if brain_mask_fn is None:
mask = None
else:
if brain_mask_fn == 'nomask':
mask = 'nomask'
else:
mask = io.open_nii(brain_mask_fn)
normalized = zscore.zscore_normalize(img, mask)
outfile = os.path.join(dirname, base + '_zscore.nii.gz')
logger.info('Normalized image saved: {}'.format(outfile))
io.save_nii(normalized, outfile, is_nii=True)
def process(image_fn, brain_mask_fn, wm_mask_fn, output_dir, args, logger):
img = io.open_nii(image_fn)
dirname, base, _ = io.split_filename(image_fn)
if output_dir is not None:
dirname = output_dir
if not os.path.exists(dirname):
logger.info('Making output directory: {}'.format(dirname))
os.mkdir(dirname)
if brain_mask_fn is not None:
mask = io.open_nii(brain_mask_fn)
wm_mask = fcm.find_wm_mask(img, mask)
outfile = os.path.join(dirname, base + '_wmmask.nii.gz')
io.save_nii(wm_mask, outfile, is_nii=True)
if wm_mask_fn is not None:
wm_mask = io.open_nii(wm_mask_fn)
normalized = fcm.fcm_normalize(img, wm_mask, args.norm_value)
outfile = os.path.join(dirname, base + '_fcm.nii.gz')
logger.info('Normalized image saved: {}'.format(outfile))
io.save_nii(normalized, outfile, is_nii=True)
def run_intensity_gmm(infile, outfolder):
from intensity_normalization.normalize import gmm
from intensity_normalization.utilities import io
try:
if not exists(join(outfolder, "Robex")):
makedirs(join(outfolder, "Robex"))
filename = infile.split(sep)[-1].split(".")[0]
i = io.open_nii(join(outfolder, infile))
b_mask = io.open_nii(
join(outfolder, "robex_masks", filename + "_mask.nii.gz"))
normalized = gmm.gmm_normalize(i, b_mask)
io.save_nii(normalized, join(outfolder, filename + "_gmm.nii.gz"))
shutil.move(join(outfolder, infile), join(outfolder, "Robex"))
except:
e = sys.exc_info()
print("Error: ", str(e[0]))
print("Error: ", str(e[1]))
print("Error: executing gmm method")
sys.exit(2)
def run_intensity_ws(infile, outfolder):
from intensity_normalization.normalize import whitestripe
from intensity_normalization.utilities import io
try:
filename = infile.split(sep)[-1].split(".")[0]
if not exists(join(outfolder, "Robex")):
makedirs(join(outfolder, "Robex"))
print('running intensity white stripe...')
mask = whitestripe.whitestripe(io.open_nii(infile), "T1")
normalized = whitestripe.whitestripe_norm(io.open_nii(infile), mask)
io.save_nii(normalized, join(outfolder, filename + "_ws.nii.gz"))
shutil.move(infile, join(outfolder, "Robex"))
except:
e = sys.exc_info()
print("Error: ", str(e[0]))
print("Error: ", str(e[1]))
print("Error: executing white stripe method")
sys.exit(2)
def run_intensity_zscore(infile, outfolder):
from intensity_normalization.normalize import zscore
from intensity_normalization.utilities import io
try:
filename = infile.split(sep)[-1].split(".")[0]
i = io.open_nii(infile)
b_mask = io.open_nii(
join(outfolder, "robex_masks", filename + "_mask.nii.gz"))
if not exists(join(outfolder, "Robex")):
makedirs(join(outfolder, "Robex"))
print('running intensity zscore...')
normalized = zscore.zscore_normalize(i, b_mask)
io.save_nii(normalized, join(outfolder, filename + "_zscore.nii.gz"))
shutil.move(join(infile), join(outfolder, "Robex"))
except:
e = sys.exc_info()
print("Error: ", str(e[0]))
print("Error: ", str(e[1]))
print("Error: executing z-score method")
sys.exit(2)
def image_matrix_to_images(V, imgs):
"""
convert an image matrix to a list of the correctly formated nifti images
Args:
V (np.ndarray): image matrix (rows are voxels, columns are images)
imgs (list): list of paths to corresponding MR images in V
Returns:
img_list (list): list of nifti images extracted from V
"""
img_list = []
for i, img_fn in enumerate(imgs):
img = io.open_nii(img_fn)
nimg = nib.Nifti1Image(V[:, i].reshape(img.get_data().shape),
img.affine, img.header)
img_list.append(nimg)
return img_list
def ws_normalize(img_dir,
contrast,
mask_dir=None,
output_dir=None,
write_to_disk=True):
"""
Use WhiteStripe normalization method ([1]) to normalize the intensities of
a set of MR images by normalizing an area around the white matter peak of the histogram
Args:
img_dir (str): directory containing MR images to be normalized
contrast (str): contrast of MR images to be normalized (T1, T2, or FLAIR)
mask_dir (str): if images are not skull-stripped, then provide brain mask
output_dir (str): directory to save images if you do not want them saved in
same directory as img_dir
write_to_disk (bool): write the normalized data to disk or nah
Returns:
normalized (np.ndarray): last normalized image data from img_dir
I know this is an odd behavior, but yolo
References:
[1] R. T. Shinohara, E. M. Sweeney, J. Goldsmith, N. Shiee,
F. J. Mateen, P. A. Calabresi, S. Jarso, D. L. Pham,
D. S. Reich, and C. M. Crainiceanu, “Statistical normalization
techniques for magnetic resonance imaging,” NeuroImage Clin.,
vol. 6, pp. 9–19, 2014.
"""
# grab the file names for the images of interest
data = io.glob_nii(img_dir)
# define and get the brain masks for the images, if defined
if mask_dir is None:
masks = [None] * len(data)
else:
masks = io.glob_nii(mask_dir)
if len(data) != len(masks):
raise NormalizationError(
'Number of images and masks must be equal, Images: {}, Masks: {}'
.format(len(data), len(masks)))
# define the output directory and corresponding output file names
if output_dir is None:
output_files = [None] * len(data)
else:
output_files = []
for fn in data:
_, base, ext = io.split_filename(fn)
output_files.append(os.path.join(output_dir, base + '_ws' + ext))
if not os.path.exists(output_dir):
os.mkdir(output_dir)
# do whitestripe normalization and save the results
for i, (img_fn, mask_fn,
output_fn) in enumerate(zip(data, masks, output_files), 1):
logger.info('Normalizing image: {} ({:d}/{:d})'.format(
img_fn, i, len(data)))
img = io.open_nii(img_fn)
mask = io.open_nii(mask_fn) if mask_fn is not None else None
indices = whitestripe(img, contrast, mask=mask)
normalized = whitestripe_norm(img, indices)
if write_to_disk:
logger.info('Saving normalized image: {} ({:d}/{:d})'.format(
output_fn, i, len(data)))
io.save_nii(normalized, output_fn)
# output the last normalized image (mostly for testing purposes)
return normalized
def run_intensity_ravel(outfolder):
from intensity_normalization.normalize import ravel
from intensity_normalization.utilities import io, csf
try:
images = []
brainMasks = []
csfMasks = []
_, _, filenames = next(walk(outfolder))
for f in filenames:
filename = f.split(sep)[-1].split(".")[0]
images.append(io.open_nii(join(outfolder, f.split(sep)[-1])))
brainMasks.append(
io.open_nii(
join(outfolder, 'robex_masks', filename + "_mask.nii.gz")))
if not exists(join(outfolder, "Robex")):
makedirs(join(outfolder, "Robex"))
if not exists(join(outfolder, "csf_masks")):
makedirs(join(outfolder, "csf_masks"))
print("creating csf masks...")
for image, brainMask, f in zip(images, brainMasks, filenames):
filename = f.split(sep)[-1].split(".")[0]
csfMask = csf.csf_mask(image,
brainMask,
contrast='T1',
csf_thresh=0.9,
return_prob=False,
mrf=0.25,
use_fcm=False)
output = nib.Nifti1Image(csfMask, None)
io.save_nii(
output,
join(outfolder, 'csf_masks', filename + "_csfmask.nii.gz"))
shutil.move(join(outfolder,
f.split(sep)[-1]), join(outfolder, "Robex"))
print('running intensity ravel...')
ravel.ravel_normalize(join(outfolder, 'Robex'),
join(outfolder, 'csf_masks'),
'T1',
output_dir=outfolder,
write_to_disk=True,
do_whitestripe=True,
b=1,
membership_thresh=0.99,
segmentation_smoothness=0.25,
do_registration=False,
use_fcm=True,
sparse_svd=False,
csf_masks=True)
for i in filenames:
rename(
join(outfolder,
i.split(sep)[-1]),
join(outfolder,
i.split(sep)[-1].split(".")[0] + "_RAVEL.nii.gz"))
except:
e = sys.exc_info()
print("Error: ", str(e[0]))
print("Error: ", str(e[1]))
print("Error: executing ravel method")
sys.exit(2)
def ravel_normalize(img_dir,
mask_dir,
contrast,
output_dir=None,
write_to_disk=False,
do_whitestripe=True,
b=1,
membership_thresh=0.99,
segmentation_smoothness=0.25,
do_registration=False,
use_fcm=True):
"""
Use RAVEL [1] to normalize the intensities of a set of MR images to eliminate
unwanted technical variation in images (but, hopefully, preserve biological variation)
this function has an option that is modified from [1] in where no registration is done,
the control mask is defined dynamically by finding a tissue segmentation of the brain and
thresholding the membership at a very high level (this seems to work well and is *much* faster)
but there seems to be some more inconsistency in the results
Args:
img_dir (str): directory containing MR images to be normalized
mask_dir (str): brain masks for imgs
contrast (str): contrast of MR images to be normalized (T1, T2, or FLAIR)
output_dir (str): directory to save images if you do not want them saved in
same directory as data_dir
write_to_disk (bool): write the normalized data to disk or nah
do_whitestripe (bool): whitestripe normalize the images before applying RAVEL correction
b (int): number of unwanted factors to estimate
membership_thresh (float): threshold of membership for control voxels
segmentation_smoothness (float): segmentation smoothness parameter for atropos ANTsPy
segmentation scheme (i.e., mrf parameter)
do_registration (bool): deformably register images to find control mask
use_fcm (bool): use FCM for segmentation instead of atropos (may be less accurate)
Returns:
Z (np.ndarray): unwanted factors (used in ravel correction)
normalized (np.ndarray): set of normalized images from data_dir
References:
[1] J. P. Fortin, E. M. Sweeney, J. Muschelli, C. M. Crainiceanu,
and R. T. Shinohara, “Removing inter-subject technical variability
in magnetic resonance imaging studies,” Neuroimage, vol. 132,
pp. 198–212, 2016.
"""
img_fns = io.glob_nii(img_dir)
mask_fns = io.glob_nii(mask_dir)
if output_dir is None or not write_to_disk:
out_fns = None
else:
out_fns = []
for fn in img_fns:
_, base, ext = io.split_filename(fn)
out_fns.append(os.path.join(output_dir, base + ext))
if not os.path.exists(output_dir):
os.mkdir(output_dir)
# get parameters necessary and setup the V array
V, Vc = image_matrix(img_fns,
contrast,
masks=mask_fns,
do_whitestripe=do_whitestripe,
return_ctrl_matrix=True,
membership_thresh=membership_thresh,
do_registration=do_registration,
smoothness=segmentation_smoothness,
use_fcm=use_fcm)
# estimate the unwanted factors Z
_, _, vh = np.linalg.svd(Vc)
Z = vh.T[:, 0:b]
# perform the ravel correction
V_norm = ravel_correction(V, Z)
# save the results to disk if desired
if write_to_disk:
for i, (img_fn, out_fn) in enumerate(zip(img_fns, out_fns)):
img = io.open_nii(img_fn)
norm = V_norm[:, i].reshape(img.get_data().shape)
io.save_nii(img, out_fn, data=norm)
return Z, V_norm
def image_matrix(imgs,
contrast,
masks=None,
do_whitestripe=True,
return_ctrl_matrix=False,
membership_thresh=0.99,
smoothness=0.25,
max_ctrl_vox=10000,
do_registration=False,
ctrl_prob=1,
use_fcm=False):
"""
creates an matrix of images where the rows correspond the the voxels of
each image and the columns are the images
Args:
imgs (list): list of paths to MR images of interest
contrast (str): contrast of the set of imgs (e.g., T1)
masks (list or str): list of corresponding brain masks or just one (template) mask
do_whitestripe (bool): do whitestripe on the images before storing in matrix or nah
return_ctrl_matrix (bool): return control matrix for imgs (i.e., a subset of V's rows)
membership_thresh (float): threshold of membership for control voxels (want this very high)
this option is only used if the registration is turned off
smoothness (float): smoothness parameter for segmentation for control voxels
this option is only used if the registration is turned off
max_ctrl_vox (int): maximum number of control voxels (if too high, everything
crashes depending on available memory) only used if do_registration is false
do_registration (bool): register the images together and take the intersection of the csf
masks (as done in the original paper, note that this takes much longer)
ctrl_prob (float): given all data, proportion of data labeled as csf to be
used for intersection (i.e., when do_registration is true)
use_fcm (bool): use FCM for segmentation instead of atropos (may be less accurate)
Returns:
V (np.ndarray): image matrix (rows are voxels, columns are images)
Vc (np.ndarray): image matrix of control voxels (rows are voxels, columns are images)
Vc only returned if return_ctrl_matrix is True
"""
img_shape = io.open_nii(imgs[0]).get_data().shape
V = np.zeros((int(np.prod(img_shape)), len(imgs)))
if return_ctrl_matrix:
ctrl_vox = []
if masks is None and return_ctrl_matrix:
raise NormalizationError(
'Brain masks must be provided if returning control memberships')
if masks is None:
masks = [None] * len(imgs)
for i, (img_fn, mask_fn) in enumerate(zip(imgs, masks)):
_, base, _ = io.split_filename(img_fn)
img = io.open_nii(img_fn)
mask = io.open_nii(mask_fn) if mask_fn is not None else None
# do whitestripe on the image before applying RAVEL (if desired)
if do_whitestripe:
logger.info('Applying WhiteStripe to image {} ({:d}/{:d})'.format(
base, i + 1, len(imgs)))
inds = whitestripe(img, contrast, mask)
img = whitestripe_norm(img, inds)
img_data = img.get_data()
if img_data.shape != img_shape:
raise NormalizationError(
'Cannot normalize because image {} needs to have same dimension '
'as all other images ({} != {})'.format(
base, img_data.shape, img_shape))
V[:, i] = img_data.flatten()
if return_ctrl_matrix:
if do_registration and i == 0:
logger.info(
'Creating control mask for image {} ({:d}/{:d})'.format(
base, i + 1, len(imgs)))
verbose = True if logger.getEffectiveLevel(
) == logging.getLevelName('DEBUG') else False
ctrl_masks = []
reg_imgs = []
reg_imgs.append(csf.nibabel_to_ants(img))
ctrl_masks.append(
csf.csf_mask(img,
mask,
contrast=contrast,
csf_thresh=membership_thresh,
mrf=smoothness,
use_fcm=use_fcm))
elif do_registration and i != 0:
template = ants.image_read(imgs[0])
tmask = ants.image_read(masks[0])
img = csf.nibabel_to_ants(img)
logger.info(
'Starting registration for image {} ({:d}/{:d})'.format(
base, i + 1, len(imgs)))
reg_result = ants.registration(template,
img,
type_of_transform='SyN',
mask=tmask,
verbose=verbose)
img = reg_result['warpedmovout']
mask = csf.nibabel_to_ants(mask)
reg_imgs.append(img)
logger.info(
'Creating control mask for image {} ({:d}/{:d})'.format(
base, i + 1, len(imgs)))
ctrl_masks.append(
csf.csf_mask(img,
mask,
contrast=contrast,
csf_thresh=membership_thresh,
mrf=smoothness,
use_fcm=use_fcm))
else:
logger.info(
'Finding control voxels for image {} ({:d}/{:d})'.format(
base, i + 1, len(imgs)))
ctrl_mask = csf.csf_mask(img,
mask,
contrast=contrast,
csf_thresh=membership_thresh,
mrf=smoothness,
use_fcm=use_fcm)
if np.sum(ctrl_mask) == 0:
raise NormalizationError(
'No control voxels found for image ({}) at threshold ({})'
.format(base, membership_thresh))
elif np.sum(ctrl_mask) < 100:
logger.warning(
'Few control voxels found ({:d}) (potentially a problematic image ({}) or '
'threshold ({}) too high)'.format(
int(np.sum(ctrl_mask)), base, membership_thresh))
ctrl_vox.append(img_data[ctrl_mask == 1].flatten())
if return_ctrl_matrix and not do_registration:
min_len = min(min(map(len, ctrl_vox)), max_ctrl_vox)
logger.info('Using {:d} control voxels'.format(min_len))
Vc = np.zeros((min_len, len(imgs)))
for i in range(len(imgs)):
ctrl_voxs = ctrl_vox[i][:min_len]
logger.info(
'Image {:d} control voxel stats - mean: {:.3f}, std: {:.3f}'.
format(i + 1, np.mean(ctrl_voxs), np.std(ctrl_voxs)))
Vc[:, i] = ctrl_voxs
elif return_ctrl_matrix and do_registration:
ctrl_sum = reduce(
add,
ctrl_masks) # need to use reduce instead of sum b/c data structure
intersection = np.zeros(ctrl_sum.shape)
intersection[ctrl_sum >= np.floor(len(ctrl_masks) * ctrl_prob)] = 1
num_ctrl_vox = int(np.sum(intersection))
Vc = np.zeros((num_ctrl_vox, len(imgs)))
for i, img in enumerate(reg_imgs):
ctrl_voxs = img.numpy()[intersection == 1]
logger.info(
'Image {:d} control voxel stats - mean: {:.3f}, std: {:.3f}'.
format(i + 1, np.mean(ctrl_voxs), np.std(ctrl_voxs)))
Vc[:, i] = ctrl_voxs
del ctrl_masks, reg_imgs
import gc
gc.collect(
) # force a garbage collection, since we just used the majority of the system memory
return V if not return_ctrl_matrix else (V, Vc)