def dice_evaluation(gt_dir,
seg_dir,
path_label_list,
path_result_dice_array):
"""Computes Dice scores for all labels contained in path_segmentation_label_list. Files in gt_folder and seg_folder
are matched by sorting order.
:param gt_dir: folder containing ground truth files.
:param seg_dir: folder containing evaluation files.
:param path_label_list: path of numpy vector containing all labels to compute the Dice for.
:param path_result_dice_array: path where the resulting Dice will be writen as numpy array.
:return: numpy array containing all dice scores (labels in rows, subjects in columns).
"""
# create result folder
if not os.path.exists(os.path.dirname(path_result_dice_array)):
os.mkdir(os.path.dirname(path_result_dice_array))
# get list label maps to compare
path_gt_labels = utils.list_images_in_folder(gt_dir)
path_segs = utils.list_images_in_folder(seg_dir)
if len(path_gt_labels) != len(path_segs):
print('different number of files in data folders, had {} and {}'.format(len(path_gt_labels), len(path_segs)))
# load labels list
label_list, neutral_labels = utils.get_list_labels(label_list=path_label_list, FS_sort=True, labels_dir=gt_dir)
label_list_sorted = np.sort(label_list)
# initialise result matrix
dice_coefs = np.zeros((label_list.shape[0], len(path_segs)))
# loop over segmentations
for idx, (path_gt, path_seg) in enumerate(zip(path_gt_labels, path_segs)):
utils.print_loop_info(idx, len(path_segs), 10)
# load gt labels and segmentation
gt_labels = utils.load_volume(path_gt, dtype='int')
seg = utils.load_volume(path_seg, dtype='int')
# crop images
gt_labels, cropping = edit_volumes.crop_volume_around_region(gt_labels, margin=10)
seg = edit_volumes.crop_volume_with_idx(seg, cropping)
# compute dice scores
tmp_dice = fast_dice(gt_labels, seg, label_list_sorted)
dice_coefs[:, idx] = tmp_dice[np.searchsorted(label_list_sorted, label_list)]
# write dice results
np.save(path_result_dice_array, dice_coefs)
return dice_coefs
def dilate_lesions(labels_dir, result_dir, recompute=True):
utils.mkdir(result_dir)
path_labels = utils.list_images_in_folder(labels_dir)
for path_label in path_labels:
path_result_label = os.path.join(result_dir,
os.path.basename(path_label))
if (not os.path.isfile(path_result_label)) | recompute:
label, aff, h = utils.load_volume(path_label, im_only=False)
# define lesion, WM, and LV masks
WM = (label == 2) | (label == 41)
lesion = label == 77
LV_and_lesion = (label == 4) | lesion
# morphological operations to bridge the gaps between lesions and LV
morph_struct = utils.build_binary_structure(2, len(WM.shape))
LV_and_lesion = binary_dilation(LV_and_lesion, morph_struct)
LV_and_lesion = binary_erosion(LV_and_lesion, morph_struct)
lesion = (LV_and_lesion & WM) | lesion
label[lesion] = 77
# save new label maps
utils.save_volume(label, aff, h, path_result_label)
def paste_lesions_on_buckner(lesion_dir,
buckner_dir,
result_dir,
dilate=2,
recompute=False):
path_lesions = utils.list_images_in_folder(lesion_dir)
path_buckners = utils.list_images_in_folder(buckner_dir)
utils.mkdir(result_dir)
# loop over buckner label maps
loop_info = utils.LoopInfo(len(path_buckners), 1, 'processing', True)
for idx_buckner, path_buckner in enumerate(path_buckners):
loop_info.update(idx_buckner)
buckner_name = os.path.basename(path_buckner).replace(
'_seg', '').replace('.nii.gz', '')
buckner = utils.load_volume(path_buckner)
WM = (buckner == 2) | (buckner == 7) | (buckner == 16) | (
buckner == 41) | (buckner == 46)
# loop over challenge data
for path_lesion in path_lesions:
lesion_name = os.path.basename(path_lesion).replace(
'.samseg_and_lesions.nii.gz', '')
path_result = os.path.join(
result_dir, buckner_name + '_' + lesion_name + '.nii.gz')
if (not os.path.isfile(path_result)) | recompute:
lesion = utils.load_volume(path_lesion)
assert lesion.shape == buckner.shape, 'lesions should have same shape as buckner labels'
# define lesion, WM, and LV masks
lesion = (lesion == 77) & WM
LV_and_lesion = (buckner == 4) | lesion
# morphological operations to bridge the gaps between lesions and LV
morph_struct = utils.build_binary_structure(
dilate, len(lesion.shape))
lesion = binary_dilation(LV_and_lesion, morph_struct)
lesion = binary_erosion(lesion, morph_struct)
lesion = lesion & WM
buckner_lesions = np.where(lesion, 77, buckner)
# save map
utils.save_volume(buckner_lesions, None, None, path_result)
def cross_validate_posteriors_threshold(list_seg_dir,
list_posteriors_dir,
list_gt_dir,
list_thresholds,
recompute=True):
for fold_idx, (seg_dir, posteriors_dir, gt_dir) in enumerate(
zip(list_seg_dir, list_posteriors_dir, list_gt_dir)):
path_dice = os.path.join(os.path.dirname(seg_dir),
'dice_lesions_for_thresholds.npy')
path_dice_means = os.path.join(
os.path.dirname(seg_dir), 'dice_lesions_means_for_thresholds.npy')
if (not os.path.isfile(path_dice)) | (
not os.path.isfile(path_dice_means)) | recompute:
path_segs = [path for path in utils.list_images_in_folder(seg_dir)]
path_posteriors = [
path for path in utils.list_images_in_folder(posteriors_dir)
]
path_gts = [path for path in utils.list_images_in_folder(gt_dir)]
dice = np.zeros((len(list_thresholds), len(path_gts)))
for subject_idx, (path_seg, path_post, path_gt) in enumerate(
zip(path_segs, path_posteriors, path_gts)):
seg = utils.load_volume(path_seg)
posteriors = utils.load_volume(path_post)
gt = utils.load_volume(path_gt)
seg[seg == 77] = 2
for idx, threshold in enumerate(list_thresholds):
tmp_seg = deepcopy(seg)
lesion_mask = posteriors > threshold
tmp_seg[lesion_mask] = 77
dice[idx, subject_idx] = fast_dice(gt, tmp_seg, [77])
np.save(path_dice, dice)
np.save(path_dice_means, np.mean(dice, axis=1))
dice_means = np.load(path_dice_means)
max_threshold = list_thresholds[np.argmax(dice_means)]
print('max threshold for fold {0}: {1:.2f}'.format(
fold_idx, max_threshold))
def estimate_t2_cropping(image_dir, result_dir=None, dilation=5):
"""This function takes all the hippocampus images (with 2 channels) within the specified directory, and estimates
the cropping dimensions around the hippocampus in the t2 channel.
It returns the mean and sts deviation for the minimal and maximal croppings, proportional to image size.
:param image_dir: path of the folder containing hippocampus images
:param result_dir: if not None, path of the folder where to write the computed statistics.
:param dilation: dilation coefficient used to extract full brain mask. Default is 5.
:returns t2_cropping_stats: numpy vector of size 4 [mean min crop, std min crop, mean max crop, std max crop]
"""
# create result dir
if result_dir is not None:
if not os.path.exists(result_dir):
os.mkdir(result_dir)
# loop through images
list_image_paths = utils.list_images_in_folder(image_dir)
max_cropping_proportions = np.zeros(len(list_image_paths))
min_cropping_proportions = np.zeros(len(list_image_paths))
for im_idx, image_path in enumerate(list_image_paths):
utils.print_loop_info(im_idx, len(list_image_paths), 10)
# load t2 channel
im = utils.load_volume(image_path)
t2 = im[..., 1]
shape = t2.shape
hdim = int(np.argmax(shape))
# mask image
_, mask = edit_volumes.mask_volume(t2,
threshold=0,
dilate=dilation,
return_mask=True)
# find cropping indices
indices = np.nonzero(mask)[hdim]
min_cropping_proportions[im_idx] = np.maximum(
np.min(indices) + int(dilation / 2), 0) / shape[hdim]
max_cropping_proportions[im_idx] = np.minimum(
np.max(indices) - int(dilation / 2), shape[hdim]) / shape[hdim]
# compute and save stats
t2_cropping_stats = np.array([
np.mean(min_cropping_proportions),
np.std(min_cropping_proportions),
np.mean(max_cropping_proportions),
np.std(max_cropping_proportions)
])
# save stats if necessary
if result_dir is not None:
np.save(os.path.join(result_dir, 't2_cropping_stats.npy'),
t2_cropping_stats)
return t2_cropping_stats
def supervised_training(image_dir,
labels_dir,
model_dir,
path_segmentation_labels=None,
batchsize=1,
output_shape=None,
flipping=True,
scaling_bounds=.15,
rotation_bounds=15,
shearing_bounds=.012,
translation_bounds=False,
nonlin_std=3.,
nonlin_shape_factor=.04,
bias_field_std=.5,
bias_shape_factor=.025,
n_levels=5,
nb_conv_per_level=2,
conv_size=3,
unet_feat_count=24,
feat_multiplier=2,
dropout=0,
activation='elu',
lr=1e-4,
lr_decay=0,
wl2_epochs=5,
dice_epochs=100,
steps_per_epoch=1000,
checkpoint=None,
reinitialise_momentum=False,
freeze_layers=False):
# check epochs
assert (wl2_epochs > 0) | (dice_epochs > 0), \
'either wl2_epochs or dice_epochs must be positive, had {0} and {1}'.format(wl2_epochs, dice_epochs)
# prepare data files
path_images = utils.list_images_in_folder(image_dir)
path_labels = utils.list_images_in_folder(labels_dir)
assert len(path_images) == len(path_labels), "There should be as many images as label maps."
# get label lists
label_list, n_neutral_labels = utils.get_list_labels(label_list=path_segmentation_labels, labels_dir=labels_dir,
FS_sort=True)
n_labels = np.size(label_list)
# create augmentation model and input generator
im_shape, _, _, n_channels, _, _ = utils.get_volume_info(path_images[0], aff_ref=np.eye(4))
augmentation_model = build_augmentation_model(im_shape,
n_channels,
label_list,
n_neutral_labels,
output_shape=output_shape,
output_div_by_n=2 ** n_levels,
flipping=flipping,
aff=np.eye(4),
scaling_bounds=scaling_bounds,
rotation_bounds=rotation_bounds,
shearing_bounds=shearing_bounds,
translation_bounds=translation_bounds,
nonlin_std=nonlin_std,
nonlin_shape_factor=nonlin_shape_factor,
bias_field_std=bias_field_std,
bias_shape_factor=bias_shape_factor)
unet_input_shape = augmentation_model.output[0].get_shape().as_list()[1:]
# prepare the segmentation model
unet_model = nrn_models.unet(nb_features=unet_feat_count,
input_shape=unet_input_shape,
nb_levels=n_levels,
conv_size=conv_size,
nb_labels=n_labels,
feat_mult=feat_multiplier,
nb_conv_per_level=nb_conv_per_level,
conv_dropout=dropout,
batch_norm=-1,
activation=activation,
input_model=augmentation_model)
# input generator
input_generator = utils.build_training_generator(build_model_inputs(path_images, path_labels, batchsize), batchsize)
# pre-training with weighted L2, input is fit to the softmax rather than the probabilities
if wl2_epochs > 0:
wl2_model = models.Model(unet_model.inputs, [unet_model.get_layer('unet_likelihood').output])
wl2_model = metrics.metrics_model(wl2_model, label_list, 'wl2')
train_model(wl2_model, input_generator, lr, lr_decay, wl2_epochs, steps_per_epoch, model_dir, 'wl2', checkpoint)
checkpoint = os.path.join(model_dir, 'wl2_%03d.h5' % wl2_epochs)
# freeze all layers but last if necessary (use -2 because the very last layer only applies softmax activation)
if freeze_layers:
for layer in unet_model.layers[:-2]:
layer.trainable = False
# fine-tuning with dice metric
dice_model = metrics.metrics_model(unet_model, label_list, 'dice')
train_model(dice_model, input_generator, lr, lr_decay, dice_epochs, steps_per_epoch, model_dir, 'dice', checkpoint,
reinitialise_momentum=reinitialise_momentum)
def prepare_output_files(path_images, out_seg, out_posteriors, out_volumes, recompute):
assert out_seg or out_posteriors, "output segmentation (or posteriors) is required"
# convert path to absolute paths
path_images = os.path.abspath(path_images)
basename = os.path.basename(path_images)
out_seg = os.path.abspath(out_seg) if (out_seg is not None) else out_seg
out_posteriors = os.path.abspath(out_posteriors) if (out_posteriors is not None) else out_posteriors
out_volumes = os.path.abspath(out_volumes) if (out_volumes is not None) else out_volumes
# prepare input/output volumes
if ('.nii.gz' not in basename) & ('.nii' not in basename) & ('.mgz' not in basename) & ('.npz' not in basename):
if os.path.isfile(path_images):
raise Exception('extension not supported for %s, only use: nii.gz, .nii, .mgz, or .npz' % path_images)
images_to_segment = utils.list_images_in_folder(path_images)
if out_seg is not None:
utils.mkdir(out_seg)
out_seg = [os.path.join(out_seg, os.path.basename(image)).replace('.nii', '_seg.nii') for image in
images_to_segment]
out_seg = [seg_path.replace('.mgz', '_seg.mgz') for seg_path in out_seg]
out_seg = [seg_path.replace('.npz', '_seg.npz') for seg_path in out_seg]
recompute_seg = [not os.path.isfile(path_seg) for path_seg in out_seg]
else:
out_seg = [out_seg] * len(images_to_segment)
recompute_seg = [False] * len(images_to_segment)
if out_posteriors is not None:
utils.mkdir(out_posteriors)
out_posteriors = [os.path.join(out_posteriors, os.path.basename(image)).replace('.nii',
'_posteriors.nii') for image in images_to_segment]
out_posteriors = [posteriors_path.replace('.mgz', '_posteriors.mgz') for posteriors_path in out_posteriors]
out_posteriors = [posteriors_path.replace('.npz', '_posteriors.npz') for posteriors_path in out_posteriors]
recompute_post = [not os.path.isfile(path_post) for path_post in out_posteriors]
else:
out_posteriors = [out_posteriors] * len(images_to_segment)
recompute_post = [out_volumes is not None] * len(images_to_segment)
else:
assert os.path.isfile(path_images), "files does not exist: %s " \
"\nplease make sure the path and the extension are correct" % path_images
images_to_segment = [path_images]
if out_seg is not None:
if ('.nii.gz' not in out_seg) & ('.nii' not in out_seg) & ('.mgz' not in out_seg) & ('.npz' not in out_seg):
utils.mkdir(out_seg)
filename = os.path.basename(path_images).replace('.nii', '_seg.nii')
filename = filename.replace('mgz', '_seg.mgz')
filename = filename.replace('.npz', '_seg.npz')
out_seg = [os.path.join(out_seg, filename)]
else:
utils.mkdir(os.path.dirname(out_seg))
out_seg = [out_seg]
recompute_seg = [not os.path.isfile(out_seg[0])]
else:
out_seg = [out_seg]
recompute_seg = [False]
if out_posteriors is not None:
if ('.nii.gz' not in out_posteriors) & ('.nii' not in out_posteriors) & ('.mgz' not in out_posteriors) & \
('.npz' not in out_posteriors):
utils.mkdir(out_posteriors)
filename = os.path.basename(path_images).replace('.nii', '_posteriors.nii')
filename = filename.replace('mgz', '_posteriors.mgz')
filename = filename.replace('.npz', '_posteriors.npz')
out_posteriors = [os.path.join(out_posteriors, filename)]
else:
utils.mkdir(os.path.dirname(out_posteriors))
out_posteriors = [out_posteriors]
recompute_post = [not os.path.isfile(out_posteriors[0])]
else:
out_posteriors = [out_posteriors]
recompute_post = [out_volumes is not None]
recompute_list = [recompute | re_seg | re_post for (re_seg, re_post) in zip(recompute_seg, recompute_post)]
if out_volumes is not None:
if out_volumes[-4:] != '.csv':
print('out_volumes provided without csv extension. Adding csv extension to output_volumes.')
out_volumes += '.csv'
utils.mkdir(os.path.dirname(out_volumes))
return images_to_segment, out_seg, out_posteriors, out_volumes, recompute_list
def inter_rater_reproducibility_cross_val_exp(manual_seg_dir,
ref_image_dir=None,
recompute=True):
# list subjects
list_subjects = utils.list_subfolders(manual_seg_dir)
# create result directories
if ref_image_dir is not None:
realigned_seg_dir = os.path.join(os.path.dirname(manual_seg_dir),
'registered_to_t1')
list_ref_subjects = utils.list_images_in_folder(ref_image_dir)
else:
realigned_seg_dir = os.path.join(os.path.dirname(manual_seg_dir),
'realigned')
list_ref_subjects = [None] * len(list_subjects)
utils.mkdir(realigned_seg_dir)
path_dice = os.path.join(realigned_seg_dir, 'dice.npy')
# loop over subjects
dice = list()
if (not os.path.isfile(path_dice)) | recompute:
for subject_dir, ref_subject in zip(list_subjects, list_ref_subjects):
# align all images to first image
if ref_subject is not None:
ref_image = ref_subject
else:
ref_image = utils.list_images_in_folder(subject_dir)[0]
result_dir = os.path.join(realigned_seg_dir,
os.path.basename(subject_dir))
edit_volumes.mri_convert_images_in_dir(subject_dir,
result_dir,
interpolation='nearest',
reference_dir=ref_image,
same_reference=True,
recompute=recompute)
# load all volumes and compute distance maps
list_segs = [
utils.load_volume(path)
for path in utils.list_images_in_folder(result_dir)
]
list_distance_maps = [
edit_volumes.compute_distance_map(labels, crop_margin=20)
for labels in list_segs
]
distance_maps = np.stack(list_distance_maps, axis=-1)
n_raters = len(list_segs)
# compare each segmentation to the consensus of all others
tmp_dice = list()
for i, seg in enumerate(list_segs):
tmp_distance_maps = distance_maps[...,
np.arange(n_raters) != i]
tmp_distance_maps = (np.mean(tmp_distance_maps, axis=-1) >
0) * 1
seg = (seg > 0) * 1
tmp_dice.append(2 * np.sum(tmp_distance_maps * seg) /
(np.sum(tmp_distance_maps) + np.sum(seg)))
dice.append(tmp_dice)
np.save(path_dice, np.array(dice))
def training(image_dir,
labels_dir,
cropping=None,
flipping=True,
scaling_range=0.07,
rotation_range=10,
shearing_range=0.01,
nonlin_std_dev=3,
nonlin_shape_fact=0.04,
crop_channel_2=None,
conv_size=3,
n_levels=5,
nb_conv_per_level=2,
feat_multiplier=2,
dropout=0,
unet_feat_count=24,
no_batch_norm=False,
lr=1e-4,
lr_decay=0,
batch_size=1,
wl2_epochs=50,
dice_epochs=500,
steps_per_epoch=100,
background_weight=1e-4,
include_background=False,
load_model_file=None,
initial_epoch_wl2=0,
initial_epoch_dice=0,
path_label_list=None,
model_dir=None):
# check epochs
assert (wl2_epochs > 0) | (dice_epochs > 0), \
'either wl2_epochs or dice_epochs must be positive, had {0} and {1}'.format(wl2_epochs, dice_epochs)
# prepare data files
image_paths = utils.list_images_in_folder(image_dir)
labels_paths = utils.list_images_in_folder(labels_dir)
assert len(image_paths) == len(labels_paths), "There should be as many images as label maps."
# get label and classes lists
rotation_range = utils.load_array_if_path(rotation_range)
scaling_range = utils.load_array_if_path(scaling_range)
crop_channel_2 = utils.load_array_if_path(crop_channel_2)
label_list, n_neutral_labels = utils.get_list_labels(label_list=path_label_list, FS_sort=True)
n_labels = np.size(label_list)
# prepare model folder
if not os.path.isdir(model_dir):
os.mkdir(model_dir)
# prepare log folder
log_dir = os.path.join(model_dir, 'logs')
if not os.path.isdir(log_dir):
os.mkdir(log_dir)
# create augmentation model and input generator
im_shape, aff, _, n_channels, _, _ = utils.get_volume_info(image_paths[0])
augmentation_model, unet_input_shape = labels_to_image_model(im_shape=im_shape,
n_channels=n_channels,
crop_shape=cropping,
label_list=label_list,
n_neutral_labels=n_neutral_labels,
vox2ras=aff,
nonlin_shape_factor=nonlin_shape_fact,
crop_channel2=crop_channel_2,
output_div_by_n=2 ** n_levels,
flipping=flipping)
model_input_generator = build_model_input_generator(images_paths=image_paths,
labels_paths=labels_paths,
n_channels=n_channels,
im_shape=im_shape,
scaling_range=scaling_range,
rotation_range=rotation_range,
shearing_range=shearing_range,
nonlin_shape_fact=nonlin_shape_fact,
nonlin_std_dev=nonlin_std_dev,
batch_size=batch_size)
training_generator = utils.build_training_generator(model_input_generator, batch_size)
# prepare the segmentation model
if no_batch_norm:
batch_norm_dim = None
else:
batch_norm_dim = -1
unet_model = nrn_models.unet(nb_features=unet_feat_count,
input_shape=unet_input_shape,
nb_levels=n_levels,
conv_size=conv_size,
nb_labels=n_labels,
feat_mult=feat_multiplier,
dilation_rate_mult=1,
nb_conv_per_level=nb_conv_per_level,
conv_dropout=dropout,
batch_norm=batch_norm_dim,
input_model=augmentation_model)
# pre-training with weighted L2, input is fit to the softmax rather than the probabilities
if wl2_epochs > 0:
wl2_model = Model(unet_model.inputs, [unet_model.get_layer('unet_likelihood').output])
wl2_model = metrics_model.metrics_model(input_shape=unet_input_shape[:-1] + [n_labels],
segmentation_label_list=label_list,
input_model=wl2_model,
metrics='weighted_l2',
weight_background=background_weight,
name='metrics_model')
if load_model_file is not None:
print('loading', load_model_file)
wl2_model.load_weights(load_model_file)
train_model(wl2_model, training_generator, lr, lr_decay, wl2_epochs, steps_per_epoch, model_dir, log_dir,
'wl2', initial_epoch_wl2)
# fine-tuning with dice metric
if dice_epochs > 0:
dice_model = metrics_model.metrics_model(input_shape=unet_input_shape[:-1] + [n_labels],
segmentation_label_list=label_list,
input_model=unet_model,
include_background=include_background,
name='metrics_model')
if wl2_epochs > 0:
last_wl2_model_name = os.path.join(model_dir, 'wl2_%03d.h5' % wl2_epochs)
dice_model.load_weights(last_wl2_model_name, by_name=True)
elif load_model_file is not None:
print('loading', load_model_file)
dice_model.load_weights(load_model_file)
train_model(dice_model, training_generator, lr, lr_decay, dice_epochs, steps_per_epoch, model_dir, log_dir,
'dice', initial_epoch_dice)
def sample_intensity_stats_from_single_dataset(image_dir, labels_dir, labels_list, classes_list=None, max_channel=3,
rescale=True):
"""This function aims at estimating the intensity distributions of K different structure types from a set of images.
The distribution of each structure type is modelled as a Gaussian, parametrised by a mean and a standard deviation.
Because the intensity distribution of structures can vary accross images, we additionally use Gausian priors for the
parameters of each Gaussian distribution. Therefore, the intensity distribution of each structure type is described
by 4 parameters: a mean/std for the mean intensity, and a mean/std for the std deviation.
This function uses a set of images along with corresponding segmentations to estimate the 4*K parameters.
Structures can share the same statistics by being regrouped into classes of similar structure types.
Images can be multi-modal (n_channels), in which case different statistics are estimated for each modality.
:param image_dir: path of directory with images to estimate the intensity distribution
:param labels_dir: path of directory with segmentation of input images.
They are matched with images by sorting order.
:param labels_list: list of labels for which to evaluate mean and std intensity.
Can be a sequence, a 1d numpy array, or the path to a 1d numpy array.
:param classes_list: (optional) enables to regroup structures into classes of similar intensity statistics.
Intenstites associated to regrouped labels will thus contribute to the same Gaussian during statistics estimation.
Can be a sequence, a 1d numpy array, or the path to a 1d numpy array.
It should have the same length as labels_list, and contain values between 0 and K-1, where K is the total number of
classes. Default is all labels have different classes (K=len(labels_list)).
:param max_channel: (optional) maximum number of channels to consider if the data is multispectral. Default is 3.
:param rescale: (optional) whether to rescale images between 0 and 255 before intensity estimation
:return: 2 numpy arrays of size (2*n_channels, K), one with the evaluated means/std for the mean
intensity, and one for the mean/std for the standard deviation.
Each block of two rows correspond to a different modality (channel). For each block of two rows, the first row
represents the mean, and the second represents the std.
"""
# list files
path_images = utils.list_images_in_folder(image_dir)
path_labels = utils.list_images_in_folder(labels_dir)
assert len(path_images) == len(path_labels), 'image and labels folders do not have the same number of files'
# reformat list labels and classes
labels_list = np.array(utils.reformat_to_list(labels_list, load_as_numpy=True, dtype='int'))
if classes_list is not None:
classes_list = np.array(utils.reformat_to_list(classes_list, load_as_numpy=True, dtype='int'))
else:
classes_list = np.arange(labels_list.shape[0])
assert len(classes_list) == len(labels_list), 'labels and classes lists should have the same length'
# get unique classes
unique_classes, unique_indices = np.unique(classes_list, return_index=True)
n_classes = len(unique_classes)
if not np.array_equal(unique_classes, np.arange(n_classes)):
raise ValueError('classes_list should only contain values between 0 and K-1, '
'where K is the total number of classes. Here K = %d' % n_classes)
# initialise result arrays
n_dims, n_channels = utils.get_dims(utils.load_volume(path_images[0]).shape, max_channels=max_channel)
means = np.zeros((len(path_images), n_classes, n_channels))
stds = np.zeros((len(path_images), n_classes, n_channels))
# loop over images
loop_info = utils.LoopInfo(len(path_images), 10, 'estimating', print_time=True)
for idx, (path_im, path_la) in enumerate(zip(path_images, path_labels)):
loop_info.update(idx)
# load image and label map
image = utils.load_volume(path_im)
la = utils.load_volume(path_la)
if n_channels == 1:
image = utils.add_axis(image, -1)
# loop over channels
for channel in range(n_channels):
im = image[..., channel]
if rescale:
im = edit_volumes.rescale_volume(im)
stats = sample_intensity_stats_from_image(im, la, labels_list, classes_list=classes_list)
means[idx, :, channel] = stats[0, :]
stds[idx, :, channel] = stats[1, :]
# compute prior parameters for mean/std
mean_means = np.mean(means, axis=0)
std_means = np.std(means, axis=0)
mean_stds = np.mean(stds, axis=0)
std_stds = np.std(stds, axis=0)
# regroup prior parameters in two different arrays: one for the mean and one for the std
prior_means = np.zeros((2 * n_channels, n_classes))
prior_stds = np.zeros((2 * n_channels, n_classes))
for channel in range(n_channels):
prior_means[2 * channel, :] = mean_means[:, channel]
prior_means[2 * channel + 1, :] = std_means[:, channel]
prior_stds[2 * channel, :] = mean_stds[:, channel]
prior_stds[2 * channel + 1, :] = std_stds[:, channel]
return prior_means, prior_stds
def training(image_dir,
labels_dir,
model_dir,
path_label_list=None,
save_label_list=None,
n_neutral_labels=1,
batchsize=1,
target_res=None,
output_shape=None,
flipping=True,
flip_rl_only=False,
scaling_bounds=0.15,
rotation_bounds=15,
enable_90_rotations=False,
shearing_bounds=.012,
translation_bounds=False,
nonlin_std=3.,
nonlin_shape_factor=.04,
bias_field_std=.3,
bias_shape_factor=.025,
same_bias_for_all_channels=False,
augment_intensitites=True,
noise_std=1.,
augment_channels_separately=True,
n_levels=5,
nb_conv_per_level=2,
conv_size=3,
unet_feat_count=24,
feat_multiplier=1,
dropout=0,
activation='elu',
lr=1e-4,
lr_decay=0,
wl2_epochs=5,
dice_epochs=200,
steps_per_epoch=1000,
checkpoint=None):
"""
This function trains a neural network with aggressively augmented images. The model is implemented on the GPU and
contains three sub-model: one for augmentation, one neural network (UNet), and one for computing the loss function.
The network is pre-trained with a weighted sum of square error, in order to bring the weights in a favorable
optimisation landscape. The training then continues with a soft dice loss function.
:param image_dir: path of folder with all input images, or to a single image (if only one training example)
:param labels_dir: path of folder with all input label maps, or to a single label map (if only one training example)
labels maps and images are likend by sorting order.
:param model_dir: path of a directory where the models will be saved during training.
#---------------------------------------------- Generation parameters ----------------------------------------------
# output-related parameters
:param path_label_list: (optional) path to a numpy array containing all the label values to be segmented.
By default, this is computed by taking all the label values in the training label maps.
:param save_label_list: (optional) path where to write the computed list of segmentation labels.
:param n_neutral_labels: (optional) number of non-sided labels in label_list. This is used for determining which
label values to swap when right/left flipping the training examples. Default is 1 (to account for the background).
:param batchsize: (optional) number of images per mini-batch. Default is 1.
:param target_res: (optional) target resolution at which to produce the segmentation label maps. The training data
will be resampled to this resolution before being run through the network. If None, no resampling is performed.
Can be a number (isotropic resolution), or the path to a 1d numpy array.
:param output_shape: (optional) desired shape of the output image, obtained by randomly cropping the generated image
Can be an integer (same size in all dimensions), a sequence, a 1d numpy array, or the path to a 1d numpy array.
# Augmentation parameters
:param flipping: (optional) whether to introduce random flipping. Default is True.
:param flip_rl_only: (optional) if flipping is True, whether to flip only in the right/left axis. Default is False.
:param scaling_bounds: (optional) if apply_linear_trans is True, the scaling factor for each dimension is
sampled from a uniform distribution of predefined bounds. scaling_bounds can either be:
1) a number, in which case the scaling factor is independently sampled from the uniform distribution of bounds
(1-scaling_bounds, 1+scaling_bounds) for each dimension.
2) the path to a numpy array of shape (2, n_dims), in which case the scaling factor is sampled from the uniform
distribution of bounds (scaling_bounds[0, i], scaling_bounds[1, i]) for the i-th dimension.
3) False, in which case scaling is completely turned off.
Default is scaling_bounds = 0.15
:param rotation_bounds: (optional) same as scaling bounds but for the rotation angle, except that for case 1 the
bounds are centred on 0 rather than 1, i.e. (0+rotation_bounds[i], 0-rotation_bounds[i]).
Default is rotation_bounds = 15.
:param enable_90_rotations: (optional) wheter to rotate the input by a random angle chosen in {0, 90, 180, 270}.
This is done regardless of the value of rotation_bounds. If true, a different value is sampled for each dimension.
:param shearing_bounds: (optional) same as scaling bounds. Default is shearing_bounds = 0.012.
:param translation_bounds: (optional) same as scaling bounds. Default is translation_bounds = False, but we
encourage using it when cropping is deactivated (i.e. when output_shape=None in BrainGenerator).
:param nonlin_std: (optional) maximum value for the standard deviation of the normal distribution from which we
sample the first tensor for synthesising the deformation field. Set to 0 to completely turn it off.
:param nonlin_shape_factor: (optional) ratio between the size of the input label maps and the size of the sampled
tensor for synthesising the elastic deformation field.
:param bias_field_std: (optional) If strictly positive, this triggers the corruption of images with a bias field.
It is obtained by sampling a first small tensor from a normal distribution, resizing it to full size, and rescaling
it to positive values by taking the voxel-wise exponential. bias_field_std designates the std dev of the normal
distribution from which we sample the first tensor. Set to 0 to deactivate biad field.
:param bias_shape_factor: (optional) If bias_field_std is not False, this designates the ratio between the size
of the input label maps and the size of the first sampled tensor for synthesising the bias field.
:param same_bias_for_all_channels: (optional) If bias_field_std is not False, whether to apply the same bias field
to all channels or not.
:param augment_intensitites: (optional) whether to augment the intensities of the images with gamma augmentation.
:param noise_std: (optional) if augment_intensities is true, maximum value for the standard deviation of the normal
distribution from which we sample a Gaussian white noise. Set to False to deactivate white noise augmentation.
Default value is 1.
:param augment_channels_separately: (optional) whether to augment the intensities of each channel indenpendently.
Only applied if augment_intensity is True, and the training images have several channels. Default is True.
# ------------------------------------------ UNet architecture parameters ------------------------------------------
:param n_levels: (optional) number of level for the Unet. Default is 5.
:param nb_conv_per_level: (optional) number of convolutional layers per level. Default is 2.
:param conv_size: (optional) size of the convolution kernels. Default is 2.
:param unet_feat_count: (optional) number of feature for the first layr of the Unet. Default is 24.
:param feat_multiplier: (optional) multiply the number of feature by this nummber at each new level. Default is 1.
:param dropout: (optional) probability of dropout for the Unet. Deafult is 0, where no dropout is applied.
:param activation: (optional) activation function. Can be 'elu', 'relu'.
# ----------------------------------------------- Training parameters ----------------------------------------------
:param lr: (optional) learning rate for the training. Default is 1e-4
:param lr_decay: (optional) learing rate decay. Default is 0, where no decay is applied.
:param wl2_epochs: (optional) number of epohs for which the network (except the soft-max layer) is trained with L2
norm loss function. Default is 5.
:param dice_epochs: (optional) number of epochs with the soft Dice loss function. default is 100.
:param steps_per_epoch: (optional) number of steps per epoch. Default is 1000. Since no online validation is
possible, this is equivalent to the frequency at which the models are saved.
:param checkpoint: (optional) path of an already saved model to load before starting the training.
"""
# check epochs
assert (wl2_epochs > 0) | (dice_epochs > 0), \
'either wl2_epochs or dice_epochs must be positive, had {0} and {1}'.format(wl2_epochs, dice_epochs)
# prepare data files
path_images = utils.list_images_in_folder(image_dir)
path_label_maps = utils.list_images_in_folder(labels_dir)
assert len(path_images) == len(path_label_maps), 'not the same number of training images and label maps.'
# read info from image and get label list
im_shape, _, _, n_channels, _, image_res = utils.get_volume_info(path_images[0], aff_ref=np.eye(4))
label_list, _ = utils.get_list_labels(path_label_list, labels_dir=labels_dir, save_label_list=save_label_list)
n_labels = np.size(label_list)
# prepare model folder
utils.mkdir(model_dir)
# transformation model
augmentation_model = build_augmentation_model(im_shape=im_shape,
n_channels=n_channels,
label_list=label_list,
n_neutral_labels=n_neutral_labels,
image_res=image_res,
target_res=target_res,
output_shape=output_shape,
output_div_by_n=2 ** n_levels,
flipping=flipping,
flip_rl_only=flip_rl_only,
aff=np.eye(4),
scaling_bounds=scaling_bounds,
rotation_bounds=rotation_bounds,
enable_90_rotations=enable_90_rotations,
shearing_bounds=shearing_bounds,
translation_bounds=translation_bounds,
nonlin_std=nonlin_std,
nonlin_shape_factor=nonlin_shape_factor,
bias_field_std=bias_field_std,
bias_shape_factor=bias_shape_factor,
same_bias_for_all_channels=same_bias_for_all_channels,
apply_intensity_augmentation=augment_intensitites,
noise_std=noise_std,
augment_channels_separately=augment_channels_separately)
unet_input_shape = augmentation_model.output[0].get_shape().as_list()[1:]
# prepare the segmentation model
unet_model = nrn_models.unet(nb_features=unet_feat_count,
input_shape=unet_input_shape,
nb_levels=n_levels,
conv_size=conv_size,
nb_labels=n_labels,
feat_mult=feat_multiplier,
nb_conv_per_level=nb_conv_per_level,
conv_dropout=dropout,
batch_norm=-1,
activation=activation,
input_model=augmentation_model)
# input generator
train_example_gen = image_seg_generator(path_images=path_images,
path_labels=path_label_maps,
batchsize=batchsize,
n_channels=n_channels)
input_generator = utils.build_training_generator(train_example_gen, batchsize)
# pre-training with weighted L2, input is fit to the softmax rather than the probabilities
if wl2_epochs > 0:
wl2_model = models.Model(unet_model.inputs, [unet_model.get_layer('unet_likelihood').output])
wl2_model = metrics.metrics_model(input_model=wl2_model, metrics='wl2')
train_model(wl2_model, input_generator, lr, lr_decay, wl2_epochs, steps_per_epoch, model_dir, 'wl2', checkpoint)
checkpoint = os.path.join(model_dir, 'wl2_%03d.h5' % wl2_epochs)
# fine-tuning with dice metric
dice_model = metrics.metrics_model(input_model=unet_model, metrics='dice')
train_model(dice_model, input_generator, lr, lr_decay, dice_epochs, steps_per_epoch, model_dir, 'dice', checkpoint)
# Prepare list of images to process
path_t1_images = os.path.abspath(args['path_t1_images'])
path_t2_images = os.path.abspath(args['path_t2_images'])
basename_t1 = os.path.basename(path_t1_images)
basename_t2 = os.path.basename(path_t2_images)
path_predictions = os.path.abspath(args['path_predictions'])
# prepare input/output volumes
# First case: you're providing directories
if ('.nii.gz' not in basename_t1) & ('.nii' not in basename_t1) & (
'.mgz' not in basename_t1) & ('.npz' not in basename_t1):
if os.path.isfile(path_t1_images):
raise Exception(
'extension not supported for %s, only use: nii.gz, .nii, .mgz, or .npz'
% path_t1_images)
images_to_segment_t1 = utils.list_images_in_folder(path_t1_images)
images_to_segment_t2 = utils.list_images_in_folder(path_t2_images)
utils.mkdir(path_predictions)
path_predictions = [
os.path.join(path_predictions,
os.path.basename(image)).replace('.nii', '_SynthSR.nii')
for image in images_to_segment_t1
]
path_predictions = [
seg_path.replace('.mgz', '_SynthSR.mgz')
for seg_path in path_predictions
]
path_predictions = [
seg_path.replace('.npz', '_SynthSR.npz')
for seg_path in path_predictions
]
def prepare_output_files(path_images, out_seg, out_posteriors, out_resampled,
out_volumes, recompute):
'''
Prepare output files.
'''
# check inputs
assert path_images is not None, 'please specify an input file/folder (--i)'
assert out_seg is not None, 'please specify an output file/folder (--o)'
# convert path to absolute paths
path_images = os.path.abspath(path_images)
basename = os.path.basename(path_images)
out_seg = os.path.abspath(out_seg) if (out_seg is not None) else out_seg
out_posteriors = os.path.abspath(out_posteriors) if (
out_posteriors is not None) else out_posteriors
out_resampled = os.path.abspath(out_resampled) if (
out_resampled is not None) else out_resampled
out_volumes = os.path.abspath(out_volumes) if (
out_volumes is not None) else out_volumes
# path_images is a folder
if ('.nii.gz' not in basename) & ('.nii' not in basename) & (
'.mgz' not in basename) & ('.npz' not in basename):
if os.path.isfile(path_images):
raise Exception(
'Extension not supported for %s, only use: nii.gz, .nii, .mgz, or .npz'
% path_images)
path_images = utils.list_images_in_folder(path_images)
if (out_seg[-7:] == '.nii.gz') | (out_seg[-4:] == '.nii') | (
out_seg[-4:] == '.mgz') | (out_seg[-4:] == '.npz'):
raise Exception(
'Output folders cannot have extensions: .nii.gz, .nii, .mgz, or .npz, had %s'
% out_seg)
utils.mkdir(out_seg)
out_seg = [
os.path.join(out_seg, os.path.basename(image)).replace(
'.nii', '_synthseg.nii') for image in path_images
]
out_seg = [
seg_path.replace('.mgz', '_synthseg.mgz') for seg_path in out_seg
]
out_seg = [
seg_path.replace('.npz', '_synthseg.npz') for seg_path in out_seg
]
recompute_seg = [not os.path.isfile(path_seg) for path_seg in out_seg]
if out_posteriors is not None:
if (out_posteriors[-7:] == '.nii.gz') | (out_posteriors[-4:] == '.nii') | \
(out_posteriors[-4:] == '.mgz') | (out_posteriors[-4:] == '.npz'):
raise Exception('Output folders cannot have extensions: '
'.nii.gz, .nii, .mgz, or .npz, had %s' %
out_posteriors)
utils.mkdir(out_posteriors)
out_posteriors = [
os.path.join(out_posteriors, os.path.basename(image)).replace(
'.nii', '_posteriors.nii') for image in path_images
]
out_posteriors = [
posteriors_path.replace('.mgz', '_posteriors.mgz')
for posteriors_path in out_posteriors
]
out_posteriors = [
posteriors_path.replace('.npz', '_posteriors.npz')
for posteriors_path in out_posteriors
]
recompute_post = [
not os.path.isfile(path_post) for path_post in out_posteriors
]
else:
out_posteriors = [out_posteriors] * len(path_images)
recompute_post = [out_volumes is not None] * len(path_images)
if out_resampled is not None:
if (out_resampled[-7:] == '.nii.gz') | (out_resampled[-4:] == '.nii') | \
(out_resampled[-4:] == '.mgz') | (out_resampled[-4:] == '.npz'):
raise Exception('Output folders cannot have extensions: '
'.nii.gz, .nii, .mgz, or .npz, had %s' %
out_resampled)
utils.mkdir(out_resampled)
out_resampled = [
os.path.join(out_resampled, os.path.basename(image)).replace(
'.nii', '_resampled.nii') for image in path_images
]
out_resampled = [
resampled_path.replace('.mgz', '_resampled.mgz')
for resampled_path in out_resampled
]
out_resampled = [
resampled_path.replace('.npz', '_resampled.npz')
for resampled_path in out_resampled
]
recompute_resampled = [
not os.path.isfile(path_post) for path_post in out_resampled
]
else:
out_resampled = [out_resampled] * len(path_images)
recompute_resampled = [out_volumes is not None] * len(path_images)
# path_images is an image
else:
assert os.path.isfile(path_images), "file does not exist: %s \n" \
"please make sure the path and the extension are correct" % path_images
path_images = [path_images]
if ('.nii.gz' not in out_seg) & ('.nii' not in out_seg) & (
'.mgz' not in out_seg) & ('.npz' not in out_seg):
utils.mkdir(out_seg)
filename = os.path.basename(path_images[0]).replace(
'.nii', '_synthseg.nii')
filename = filename.replace('.mgz', '_synthseg.mgz')
filename = filename.replace('.npz', '_synthseg.npz')
out_seg = os.path.join(out_seg, filename)
else:
utils.mkdir(os.path.dirname(out_seg))
out_seg = [out_seg]
recompute_seg = [not os.path.isfile(out_seg[0])]
if out_posteriors is not None:
if ('.nii.gz' not in out_posteriors) & ('.nii' not in out_posteriors) &\
('.mgz' not in out_posteriors) & ('.npz' not in out_posteriors):
utils.mkdir(out_posteriors)
filename = os.path.basename(path_images[0]).replace(
'.nii', '_posteriors.nii')
filename = filename.replace('.mgz', '_posteriors.mgz')
filename = filename.replace('.npz', '_posteriors.npz')
out_posteriors = os.path.join(out_posteriors, filename)
else:
utils.mkdir(os.path.dirname(out_posteriors))
recompute_post = [not os.path.isfile(out_posteriors)]
else:
recompute_post = [out_volumes is not None]
out_posteriors = [out_posteriors]
if out_resampled is not None:
if ('.nii.gz' not in out_resampled) & ('.nii' not in out_resampled) &\
('.mgz' not in out_resampled) & ('.npz' not in out_resampled):
utils.mkdir(out_resampled)
filename = os.path.basename(path_images[0]).replace(
'.nii', '_resampled.nii')
filename = filename.replace('.mgz', '_resampled.mgz')
filename = filename.replace('.npz', '_resampled.npz')
out_resampled = os.path.join(out_resampled, filename)
else:
utils.mkdir(os.path.dirname(out_resampled))
recompute_resampled = [not os.path.isfile(out_resampled)]
else:
recompute_resampled = [out_volumes is not None]
out_resampled = [out_resampled]
recompute_list = [
recompute | re_seg | re_post | re_res
for (re_seg, re_post,
re_res) in zip(recompute_seg, recompute_post, recompute_resampled)
]
if out_volumes is not None:
if out_volumes[-4:] != '.csv':
print(
'Path for volume outputs provided without csv extension. Adding csv extension.'
)
out_volumes += '.csv'
utils.mkdir(os.path.dirname(out_volumes))
return path_images, out_seg, out_posteriors, out_resampled, out_volumes, recompute_list
def __init__(self,
labels_dir,
generation_labels=None,
output_labels=None,
n_neutral_labels=None,
padding_margin=None,
batch_size=1,
n_channels=1,
target_res=None,
output_shape=None,
output_div_by_n=None,
prior_distributions='uniform',
generation_classes=None,
prior_means=None,
prior_stds=None,
use_specific_stats_for_channel=False,
flipping=True,
apply_linear_trans=True,
scaling_bounds=None,
rotation_bounds=None,
shearing_bounds=None,
apply_nonlin_trans=True,
nonlin_std=3.,
nonlin_shape_factor=0.0625,
blur_background=True,
data_res=None,
thickness=None,
downsample=False,
blur_range=1.15,
crop_channel_2=None,
apply_bias_field=True,
bias_field_std=0.3,
bias_shape_factor=0.025):
"""
This class is wrapper around the labels_to_image_model model. It contains the GPU model that generates images
from labels maps, and a python generator that suplies the input data for this model.
To generate pairs of image/labels you can just call the method generate_image() on an object of this class.
:param labels_dir: path of folder with all input label maps, or to a single label map.
# IMPORTANT !!!
# Each time we provide a parameter with separate values for each axis (e.g. with a numpy array or a sequence),
# these values refer to the axes of the raw label map (i.e. once it has been loaded in python).
# Depending on the label map orientation, the axes of its raw array may or may not correspond to the RAS axes.
# label maps-related parameters
:param generation_labels: (optional) list of all possible label values in the input label maps.
Default is None, where the label values are directly gotten from the provided label maps.
If not None, can be a sequence or a 1d numpy array, or the path to a 1d numpy array.
If flipping is true (i.e. right/left flipping is enabled), generation_labels should be organised as follows:
background label first, then non-sided labels (e.g. CSF, brainstem, etc.), then all the structures of the same
hemisphere (can be left or right), and finally all the corresponding contralateral structures in the same order.
:param output_labels: (optional) list of all the label values to keep in the output label maps (in no particular
order). Should be a subset of the values contained in generation_labels.
Label values that are in generation_labels but not in output_labels are reset to zero.
Can be a sequence, a 1d numpy array, or the path to a 1d numpy array.
:param n_neutral_labels: (optional) number of non-sided generation labels.
Default is total number of label values.
:param padding_margin: (optional) margin by which to pad the input labels with zeros.
Padding is applied prior to any other operation.
Can be an integer (same padding in all dimensions), a sequence, a 1d numpy array, or the path to a 1d numpy
array. Default is no padding.
# output-related parameters
:param batch_size: (optional) numbers of images to generate per mini-batch. Default is 1.
:param n_channels: (optional) number of channels to be synthetised. Default is 1.
:param target_res: (optional) target resolution of the generated images and corresponding label maps.
If None, the outputs will have the same resolution as the input label maps.
Can be a number (isotropic resolution), a sequence, a 1d numpy array, or the path to a 1d numpy array.
:param output_shape: (optional) shape of the output image, obtained by randomly cropping the generated image.
Can be an integer (same size in all dimensions), a sequence, a 1d numpy array, or the path to a 1d numpy array.
:param output_div_by_n: (optional) forces the output shape to be divisible by this value. It overwrites
output_shape if necessary. Can be an integer (same size in all dimensions), a sequence, a 1d numpy array, or
the path to a 1d numpy array.
# GMM-sampling parameters
:param generation_classes: (optional) Indices regrouping generation labels into classes of same intensity
distribution. Regouped labels will thus share the same Gaussian when samling a new image. Can be a sequence, a
1d numpy array, or the path to a 1d numpy array. It should have the same length as generation_labels, and
contain values between 0 and K-1, where K is the total number of classes.
Default is all labels have different classes (K=len(generation_labels)).
:param prior_distributions: (optional) type of distribution from which we sample the GMM parameters.
Can either be 'uniform', or 'normal'. Default is 'uniform'.
:param prior_means: (optional) hyperparameters controlling the prior distributions of the GMM means. Because
these prior distributions are uniform or normal, they require by 2 hyperparameters. Thus prior_means can be:
1) a sequence of length 2, directly defining the two hyperparameters: [min, max] if prior_distributions is
uniform, [mean, std] if the distribution is normal. The GMM means of are independently sampled at each
mini_batch from the same distribution.
2) an array of shape (2, K), where K is the number of classes (K=len(generation_labels) if generation_classes is
not given). The mean of the Gaussian distribution associated to class k in [0, ...K-1] is sampled at each
mini-batch from U(prior_means[0,k], prior_means[1,k]) if prior_distributions is uniform, and from
N(prior_means[0,k], prior_means[1,k]) if prior_distributions is normal.
3) an array of shape (2*n_mod, K), where each block of two rows is associated to hyperparameters derived
from different modalities. In this case, if use_specific_stats_for_channel is False, we first randomly select a
modality from the n_mod possibilities, and we sample the GMM means like in 2).
If use_specific_stats_for_channel is True, each block of two rows correspond to a different channel
(n_mod=n_channels), thus we select the corresponding block to each channel rather than randomly drawing it.
4) the path to such a numpy array.
Default is None, which corresponds to prior_means = [25, 225].
:param prior_stds: (optional) same as prior_means but for the standard deviations of the GMM.
Default is None, which corresponds to prior_stds = [5, 25].
:param use_specific_stats_for_channel: (optional) whether the i-th block of two rows in the prior arrays must be
only used to generate the i-th channel. If True, n_mod should be equal to n_channels. Default is False.
# spatial deformation parameters
:param flipping: (optional) whether to introduce right/left random flipping. Default is True.
:param apply_linear_trans: (optional) whether to apply affine deformation. Default is True.
:param scaling_bounds: (optional) if apply_linear_trans is True, the scaling factor for each dimension is
sampled from a uniform distribution of predefined bounds. Can either be:
1) a number, in which case the scaling factor is independently sampled from the uniform distribution of bounds
(1-scaling_bounds, 1+scaling_bounds) for each dimension.
2) a sequence, in which case the scaling factor is sampled from the uniform distribution of bounds
(1-scaling_bounds[i], 1+scaling_bounds[i]) for the i-th dimension.
3) a numpy array of shape (2, n_dims), in which case the scaling factor is sampled from the uniform distribution
of bounds (scaling_bounds[0, i], scaling_bounds[1, i]) for the i-th dimension.
4) the path to such a numpy array.
If None (default), scaling_range = 0.15
:param rotation_bounds: (optional) same as scaling bounds but for the rotation angle, except that for cases 1
and 2, the bounds are centred on 0 rather than 1, i.e. (0+rotation_bounds[i], 0-rotation_bounds[i]).
If None (default), rotation_bounds = 15.
:param shearing_bounds: (optional) same as scaling bounds. If None (default), shearing_bounds = 0.01.
:param apply_nonlin_trans: (optional) whether to apply non linear elastic deformation.
If true, a diffeomorphic deformation field is obtained by first sampling a small tensor from the normal
distribution, resizing it to image size, and integrationg it. Default is True.
:param nonlin_std: (optional) If apply_nonlin_trans is True, maximum value for the standard deviation of the
normal distribution from which we sample the first tensor for synthesising the deformation field.
:param nonlin_shape_factor: (optional) If apply_nonlin_trans is True, ratio between the size of the input label
maps and the size of the sampled tensor for synthesising the deformation field.
# blurring/resampling parameters
:param blur_background: (optional) whether to produce an unrealistic background or not.
If True, the background is generated/blurred with the other labels, according to the values of prior_means and
prior_stds. Also, it is reset to zero-background with a probability of 0.2.
If False, the background is reset to zero, or can be replaced by a low-intensity background with a probability
of 0.5. Additionally we correct for edge blurring effects.
Default is True.
:param data_res: (optional) acquisition resolution to mimick. If provided, the images sampled from the GMM are
blurred to mimick data that would be: 1) acquired at the given acquisition resolution, and 2) resample at
target_resolution.
Default is None, where images are isotropically blurred to introduce some spatial correlation between voxels.
If the generated images are uni-modal, data_res can be a number (isotropic acquisition resolution), a sequence,
a 1d numpy array, or the path to a 1d numy array. In the multi-modal case, it should be given as a numpy array
(or a path) of size (n_mod, n_dims), where each row is the acquisition resolution of the correspionding chanel.
:param thickness: (optional) if data_res is provided, we can further specify the slice thickness of the low
resolution images to mimick.
If the generated images are uni-modal, data_res can be a number (isotropic acquisition resolution), a sequence,
a 1d numpy array, or the path to a 1d numy array. In the multi-modal case, it should be given as a numpy array
(or a path) of size (n_mod, n_dims), where each row is the acquisition resolution of the correspionding chanel.
:param downsample: (optional) whether to actually downsample the volume image to data_res.
Default is False, except when thickness is provided, and thickness < data_res.
:param blur_range: (optional) Randomise the standard deviation of the blurring kernels, (whether data_res is
given or not). At each mini_batch, the standard deviation of the blurring kernels are multiplied by a
coefficient sampled from a uniform distribution with bounds [1/blur_range, blur_range].
If None, no randomisation. Default is 1.15.
:param crop_channel_2: (optional) stats for cropping second channel along the anterior-posterior axis.
Should be a vector of length 4, with bounds of uniform distribution for cropping the front and back of the image
(in percentage). None is no croppping.
# bias field parameters
:param apply_bias_field: (optional) whether to apply a bias field to the final image. Default is True.
If True, the bias field is obtained by sampling a first tensor from normal distribution, resizing it to image
size, and rescaling the values to positive number by taking the voxel-wise exponential. Default is True.
:param bias_field_std: (optional) If apply_nonlin_trans is True, maximum value for the standard deviation of the
normal distribution from which we sample the first tensor for synthesising the bias field.
:param bias_shape_factor: (optional) If apply_bias_field is True, ratio between the size of the input
label maps and the size of the sampled tensor for synthesising the bias field.
"""
# prepare data files
if ('.nii.gz' in labels_dir) | ('.nii' in labels_dir) | (
'.mgz' in labels_dir) | ('.npz' in labels_dir):
self.labels_paths = [labels_dir]
else:
self.labels_paths = utils.list_images_in_folder(labels_dir)
assert len(self.labels_paths) > 0, "Could not find any training data"
# generation parameters
self.labels_shape, self.aff, self.n_dims, _, self.header, self.atlas_res = \
utils.get_volume_info(self.labels_paths[0])
self.n_channels = n_channels
if generation_labels is not None:
self.generation_labels = utils.load_array_if_path(
generation_labels)
else:
self.generation_labels = utils.get_list_labels(
labels_dir=labels_dir)
if output_labels is not None:
self.output_labels = utils.load_array_if_path(output_labels)
else:
self.output_labels = self.generation_labels
if n_neutral_labels is not None:
self.n_neutral_labels = n_neutral_labels
else:
self.n_neutral_labels = self.generation_labels.shape[0]
self.target_res = utils.load_array_if_path(target_res)
# preliminary operations
self.padding_margin = utils.load_array_if_path(padding_margin)
self.flipping = flipping
self.output_shape = utils.load_array_if_path(output_shape)
self.output_div_by_n = output_div_by_n
# GMM parameters
self.prior_distributions = prior_distributions
if generation_classes is not None:
self.generation_classes = utils.load_array_if_path(
generation_classes)
assert self.generation_classes.shape == self.generation_labels.shape, \
'if provided, generation labels should have the same shape as generation_labels'
unique_classes = np.unique(self.generation_classes)
assert np.array_equal(unique_classes, np.arange(np.max(unique_classes)+1)), \
'generation_classes should a linear range between 0 and its maximum value.'
else:
self.generation_classes = np.arange(
self.generation_labels.shape[0])
self.prior_means = utils.load_array_if_path(prior_means)
self.prior_stds = utils.load_array_if_path(prior_stds)
self.use_specific_stats_for_channel = use_specific_stats_for_channel
# linear transformation parameters
self.apply_linear_trans = apply_linear_trans
self.scaling_bounds = utils.load_array_if_path(scaling_bounds)
self.rotation_bounds = utils.load_array_if_path(rotation_bounds)
self.shearing_bounds = utils.load_array_if_path(shearing_bounds)
# elastic transformation parameters
self.apply_nonlin_trans = apply_nonlin_trans
self.nonlin_std = nonlin_std
self.nonlin_shape_factor = nonlin_shape_factor
# blurring parameters
self.blur_background = blur_background
self.data_res = utils.load_array_if_path(data_res)
self.thickness = utils.load_array_if_path(thickness)
self.downsample = downsample
self.blur_range = blur_range
self.crop_second_channel = utils.load_array_if_path(crop_channel_2)
# bias field parameters
self.apply_bias_field = apply_bias_field
self.bias_field_std = bias_field_std
self.bias_shape_factor = bias_shape_factor
# build transformation model
self.labels_to_image_model, self.model_output_shape = self._build_labels_to_image_model(
)
# build generator for model inputs
self.model_inputs_generator = self._build_model_inputs_generator(
batch_size)
# build brain generator
self.brain_generator = self._build_brain_generator()
def __init__(self,
labels_dir,
generation_labels=None,
output_labels=None,
n_neutral_labels=None,
batchsize=1,
n_channels=1,
target_res=None,
output_shape=None,
output_div_by_n=None,
prior_distributions='uniform',
generation_classes=None,
prior_means=None,
prior_stds=None,
use_specific_stats_for_channel=False,
mix_prior_and_random=False,
flipping=True,
scaling_bounds=0.15,
rotation_bounds=15,
shearing_bounds=.012,
translation_bounds=False,
nonlin_std=4.,
nonlin_shape_factor=0.0625,
randomise_res=False,
buil_distance_maps=False,
data_res=None,
thickness=None,
downsample=False,
blur_range=1.15,
bias_field_std=0.5,
bias_shape_factor=0.025):
"""
This class is wrapper around the labels_to_image_model model. It contains the GPU model that generates images
from labels maps, and a python generator that suplies the input data for this model.
To generate pairs of image/labels you can just call the method generate_image() on an object of this class.
:param labels_dir: path of folder with all input label maps, or to a single label map.
# IMPORTANT !!!
# Each time we provide a parameter with separate values for each axis (e.g. with a numpy array or a sequence),
# these values refer to the RAS axes.
# label maps-related parameters
:param generation_labels: (optional) list of all possible label values in the input label maps.
Default is None, where the label values are directly gotten from the provided label maps.
If not None, can be a sequence or a 1d numpy array, or the path to a 1d numpy array.
If flipping is true (i.e. right/left flipping is enabled), generation_labels should be organised as follows:
background label first, then non-sided labels (e.g. CSF, brainstem, etc.), then all the structures of the same
hemisphere (can be left or right), and finally all the corresponding contralateral structures in the same order.
:param output_labels: (optional) list of all the label values to keep in the output label maps (in no particular
order). Should be a subset of the values contained in generation_labels.
Label values that are in generation_labels but not in output_labels are reset to zero.
Can be a sequence, a 1d numpy array, or the path to a 1d numpy array.
By default output labels are equal to generation labels.
:param n_neutral_labels: (optional) number of non-sided generation labels.
Default is total number of label values.
# output-related parameters
:param batchsize: (optional) numbers of images to generate per mini-batch. Default is 1.
:param n_channels: (optional) number of channels to be synthetised. Default is 1.
:param target_res: (optional) target resolution of the generated images and corresponding label maps.
If None, the outputs will have the same resolution as the input label maps.
Can be a number (isotropic resolution), a sequence, a 1d numpy array, or the path to a 1d numpy array.
:param output_shape: (optional) shape of the output image, obtained by randomly cropping the generated image.
Can be an integer (same size in all dimensions), a sequence, a 1d numpy array, or the path to a 1d numpy array.
Default is None, where no cropping is performed.
:param output_div_by_n: (optional) forces the output shape to be divisible by this value. It overwrites
output_shape if necessary. Can be an integer (same size in all dimensions), a sequence, a 1d numpy array, or
the path to a 1d numpy array.
# GMM-sampling parameters
:param generation_classes: (optional) Indices regrouping generation labels into classes of same intensity
distribution. Regouped labels will thus share the same Gaussian when samling a new image. Can be a sequence, a
1d numpy array, or the path to a 1d numpy array. It should have the same length as generation_labels, and
contain values between 0 and K-1, where K is the total number of classes.
Default is all labels have different classes (K=len(generation_labels)).
:param prior_distributions: (optional) type of distribution from which we sample the GMM parameters.
Can either be 'uniform', or 'normal'. Default is 'uniform'.
:param prior_means: (optional) hyperparameters controlling the prior distributions of the GMM means. Because
these prior distributions are uniform or normal, they require by 2 hyperparameters. Thus prior_means can be:
1) a sequence of length 2, directly defining the two hyperparameters: [min, max] if prior_distributions is
uniform, [mean, std] if the distribution is normal. The GMM means of are independently sampled at each
mini_batch from the same distribution.
2) an array of shape (2, K), where K is the number of classes (K=len(generation_labels) if generation_classes is
not given). The mean of the Gaussian distribution associated to class k in [0, ...K-1] is sampled at each
mini-batch from U(prior_means[0,k], prior_means[1,k]) if prior_distributions is uniform, and from
N(prior_means[0,k], prior_means[1,k]) if prior_distributions is normal.
3) an array of shape (2*n_mod, K), where each block of two rows is associated to hyperparameters derived
from different modalities. In this case, if use_specific_stats_for_channel is False, we first randomly select a
modality from the n_mod possibilities, and we sample the GMM means like in 2).
If use_specific_stats_for_channel is True, each block of two rows correspond to a different channel
(n_mod=n_channels), thus we select the corresponding block to each channel rather than randomly drawing it.
4) the path to such a numpy array.
Default is None, which corresponds to prior_means = [25, 225].
:param prior_stds: (optional) same as prior_means but for the standard deviations of the GMM.
Default is None, which corresponds to prior_stds = [5, 25].
:param use_specific_stats_for_channel: (optional) whether the i-th block of two rows in the prior arrays must be
only used to generate the i-th channel. If True, n_mod should be equal to n_channels. Default is False.
:param mix_prior_and_random: (optional) if prior_means is not None, enables to reset the priors to their default
values for half of thes cases, and thus generate images of random contrast.
# spatial deformation parameters
:param flipping: (optional) whether to introduce right/left random flipping. Default is True.
:param scaling_bounds: (optional) range of the random saling to apply at each mini-batch. The scaling factor for
each dimension is sampled from a uniform distribution of predefined bounds. Can either be:
1) a number, in which case the scaling factor is independently sampled from the uniform distribution of bounds
[1-scaling_bounds, 1+scaling_bounds] for each dimension.
2) a sequence, in which case the scaling factor is sampled from the uniform distribution of bounds
(1-scaling_bounds[i], 1+scaling_bounds[i]) for the i-th dimension.
3) a numpy array of shape (2, n_dims), in which case the scaling factor is sampled from the uniform distribution
of bounds (scaling_bounds[0, i], scaling_bounds[1, i]) for the i-th dimension.
4) False, in which case scaling is completely turned off.
Default is scaling_bounds = 0.15 (case 1)
:param rotation_bounds: (optional) same as scaling bounds but for the rotation angle, except that for cases 1
and 2, the bounds are centred on 0 rather than 1, i.e. [0+rotation_bounds[i], 0-rotation_bounds[i]].
Default is rotation_bounds = 15.
:param shearing_bounds: (optional) same as scaling bounds. Default is shearing_bounds = 0.012.
:param translation_bounds: (optional) same as scaling bounds. Default is translation_bounds = False, but we
encourage using it when cropping is deactivated (i.e. when output_shape=None in BrainGenerator).
:param nonlin_std: (optional) Maximum value for the standard deviation of the normal distribution from which we
sample the first tensor for synthesising the deformation field. Set to 0 if you wish to completely turn the
elastic deformation off.
:param nonlin_shape_factor: (optional) if nonlin_std is strictly positive, factor between the shapes of the
input label maps and the shape of the input non-linear tensor.
# blurring/resampling parameters
:param randomise_res: (optional) whether to mimic images that would have been 1) acquired at low resolution, and
2) resampled to high esolution. The low resolution is uniformly resampled at each minibatch from [1mm, 9mm].
In that process, the images generated by sampling the GMM are:
1) blurred at the sampled LR, 2) downsampled at LR, and 3) resampled at target_resolution.
:param data_res: (optional) specific acquisition resolution to mimic, as opposed to random resolution sampled
when randomis_res is True. This triggers a blurring which mimics the acquisition resolution, but downsampling
is optional (see param downsample). Default for data_res is None, where images are slighlty blurred.
If the generated images are uni-modal, data_res can be a number (isotropic acquisition resolution), a sequence,
a 1d numpy array, or the path to a 1d numy array. In the multi-modal case, it should be given as a umpy array (
or a path) of size (n_mod, n_dims), where each row is the acquisition resolution of the corresponding channel.
:param thickness: (optional) if data_res is provided, we can further specify the slice thickness of the low
resolution images to mimic. Must be provided in the same format as data_res. Default thickness = data_res.
:param downsample: (optional) whether to actually downsample the volume images to data_res after blurring.
Default is False, except when thickness is provided, and thickness < data_res.
:param blur_range: (optional) Randomise the standard deviation of the blurring kernels, (whether data_res is
given or not). At each mini_batch, the standard deviation of the blurring kernels are multiplied by a
coefficient sampled from a uniform distribution with bounds [1/blur_range, blur_range].
If None, no randomisation. Default is 1.15.
# bias field parameters
:param bias_field_std: (optional) If strictly positive, this triggers the corruption of synthesised images with
a bias field. It is obtained by sampling a first small tensor from a normal distribution, resizing it to full
size, and rescaling it to positive values by taking the voxel-wise exponential. bias_field_std designates the
std dev of the normal distribution from which we sample the first tensor. Set to 0 to deactivate bias field.
:param bias_shape_factor: (optional) If bias_field_std is strictly positive, this designates the ratio between
the size of the input label maps and the size of the first sampled tensor for synthesising the bias field.
"""
# prepare data files
self.labels_paths = utils.list_images_in_folder(labels_dir)
# generation parameters
self.labels_shape, self.aff, self.n_dims, _, self.header, self.atlas_res = \
utils.get_volume_info(self.labels_paths[0], aff_ref=np.eye(4))
self.n_channels = n_channels
if generation_labels is not None:
self.generation_labels = utils.load_array_if_path(generation_labels)
else:
self.generation_labels, _ = utils.get_list_labels(labels_dir=labels_dir)
if output_labels is not None:
self.output_labels = utils.load_array_if_path(output_labels)
else:
self.output_labels = self.generation_labels
if n_neutral_labels is not None:
self.n_neutral_labels = n_neutral_labels
else:
self.n_neutral_labels = self.generation_labels.shape[0]
self.target_res = utils.load_array_if_path(target_res)
self.batchsize = batchsize
# preliminary operations
self.flipping = flipping
self.output_shape = utils.load_array_if_path(output_shape)
self.output_div_by_n = output_div_by_n
# GMM parameters
self.prior_distributions = prior_distributions
if generation_classes is not None:
self.generation_classes = utils.load_array_if_path(generation_classes)
assert self.generation_classes.shape == self.generation_labels.shape, \
'if provided, generation labels should have the same shape as generation_labels'
unique_classes = np.unique(self.generation_classes)
assert np.array_equal(unique_classes, np.arange(np.max(unique_classes)+1)), \
'generation_classes should a linear range between 0 and its maximum value.'
else:
self.generation_classes = np.arange(self.generation_labels.shape[0])
self.prior_means = utils.load_array_if_path(prior_means)
self.prior_stds = utils.load_array_if_path(prior_stds)
self.use_specific_stats_for_channel = use_specific_stats_for_channel
# linear transformation parameters
self.scaling_bounds = utils.load_array_if_path(scaling_bounds)
self.rotation_bounds = utils.load_array_if_path(rotation_bounds)
self.shearing_bounds = utils.load_array_if_path(shearing_bounds)
self.translation_bounds = utils.load_array_if_path(translation_bounds)
# elastic transformation parameters
self.nonlin_std = nonlin_std
self.nonlin_shape_factor = nonlin_shape_factor
# blurring parameters
self.randomise_res = randomise_res
self.buil_distance_maps = buil_distance_maps
self.data_res = utils.load_array_if_path(data_res)
assert not (self.randomise_res & (self.data_res is not None)), \
'randomise_res and data_res cannot be provided at the same time'
self.thickness = utils.load_array_if_path(thickness)
self.downsample = downsample
self.blur_range = blur_range
# bias field parameters
self.bias_field_std = bias_field_std
self.bias_shape_factor = bias_shape_factor
# build transformation model
self.labels_to_image_model, self.model_output_shape = self._build_labels_to_image_model()
# build generator for model inputs
self.model_inputs_generator = self._build_model_inputs_generator(mix_prior_and_random)
# build brain generator
self.brain_generator = self._build_brain_generator()
def dice_evaluation(gt_folder, seg_folder, path_segmentation_label_list,
path_result_dice_array):
"""Computes Dice scores for all labels contained in path_segmentation_label_list. Files in gt_folder and seg_folder
are matched by sorting order.
:param gt_folder: folder containing ground truth files.
:param seg_folder: folder containing evaluation files.
:param path_segmentation_label_list: path of numpy vector containing all labels to compute the Dice for.
:param path_result_dice_array: path where the resulting Dice will be writen as numpy array.
:return: numpy array containing all dice scores (labels in rows, subjects in columns).
"""
# get list of automated and manual segmentations
list_path_gt_labels = utils.list_images_in_folder(gt_folder)
list_path_segs = utils.list_images_in_folder(seg_folder)
if len(list_path_gt_labels) != len(list_path_segs):
warnings.warn(
'both data folders should have the same length, had {} and {}'.
format(len(list_path_gt_labels), len(list_path_segs)))
# load labels list
label_list, neutral_labels = utils.get_list_labels(
label_list=path_segmentation_label_list,
FS_sort=True,
labels_dir=gt_folder)
# create result folder
if not os.path.exists(os.path.dirname(path_result_dice_array)):
os.mkdir(os.path.dirname(path_result_dice_array))
# initialise result matrix
dice_coefs = np.zeros((label_list.shape[0], len(list_path_segs)))
# start analysis
for im_idx, (path_gt, path_seg) in enumerate(
zip(list_path_gt_labels, list_path_segs)):
utils.print_loop_info(im_idx, len(list_path_segs), 10)
# load gt labels and segmentation
gt_labels = utils.load_volume(path_gt, dtype='int')
seg = utils.load_volume(path_seg, dtype='int')
n_dims = len(gt_labels.shape)
# crop images
gt_labels, cropping = edit_volumes.crop_volume_around_region(gt_labels,
margin=10)
if n_dims == 2:
seg = seg[cropping[0]:cropping[2], cropping[1]:cropping[3]]
elif n_dims == 3:
seg = seg[cropping[0]:cropping[3], cropping[1]:cropping[4],
cropping[2]:cropping[5]]
else:
raise Exception(
'cannot evaluate images with more than 3 dimensions')
# extract list of unique labels
label_list_sorted = np.sort(label_list)
tmp_dice = fast_dice(gt_labels, seg, label_list_sorted)
dice_coefs[:,
im_idx] = tmp_dice[np.searchsorted(label_list_sorted,
label_list)]
# write dice results
np.save(path_result_dice_array, dice_coefs)
return dice_coefs
def prepare_hippo_training_atlases(labels_dir,
result_dir,
image_dir=None,
image_result_dir=None,
smooth=True,
crop_margin=50,
recompute=True):
"""This function prepares training label maps from CobraLab. It first crops each atlas around the right and left
hippocampi, with a margin. It then equalises the shape of these atlases by croppping them to the size of the
smallest hippocampus. Finally it realigns the obtained atlases to FS orientation axes.
:param labels_dir: path of directory with label maps to prepare
:param result_dir: path of directory where prepared atlases will be writen
:param image_dir: (optional) path of directory with images corresponding to the label maps to prepare.
This can be sued to prepare a dataset of real images for supervised training.
:param image_result_dir: (optional) path of directory where images corresponding to prepared atlases will be writen
:param smooth: (optional) whether to smooth the final cropped label maps
:param crop_margin: (optional) margin to add around hippocampi when cropping
:param recompute: (optional) whether to recompute result files even if they already exists"""
# create results dir
if not os.path.exists(result_dir):
os.mkdir(result_dir)
tmp_result_dir = os.path.join(result_dir, 'first_cropping')
if not os.path.exists(tmp_result_dir):
os.mkdir(tmp_result_dir)
if image_dir is not None:
assert image_result_dir is not None, 'image_result_dir should not be None if image_dir is specified'
if not os.path.exists(image_result_dir):
os.mkdir(image_result_dir)
tmp_image_result_dir = os.path.join(image_result_dir, 'first_cropping')
if not os.path.exists(tmp_image_result_dir):
os.mkdir(tmp_image_result_dir)
else:
tmp_image_result_dir = None
# list labels and images
labels_paths = utils.list_images_in_folder(labels_dir)
if image_dir is not None:
path_images = utils.list_images_in_folder(image_dir)
else:
path_images = [None] * len(labels_paths)
# crop all atlases around hippo
print('\ncropping around hippo')
shape_array = np.zeros((len(labels_paths)*2, 3))
for idx, (path_label, path_image) in enumerate(zip(labels_paths, path_images)):
utils.print_loop_info(idx, len(labels_paths), 1)
# crop left hippo first
path_label_first_crop_l = os.path.join(tmp_result_dir,
os.path.basename(path_label).replace('.nii', '_left.nii'))
lab, aff, h = utils.load_volume(path_label, im_only=False)
lab_l, croppping_idx, aff_l = edit_volumes.crop_volume_around_region(lab, crop_margin,
list(range(20101, 20109)), aff=aff)
if (not os.path.exists(path_label_first_crop_l)) | recompute:
utils.save_volume(lab_l, aff_l, h, path_label_first_crop_l)
else:
lab_l = utils.load_volume(path_label_first_crop_l)
if path_image is not None:
path_image_first_crop_l = os.path.join(tmp_image_result_dir,
os.path.basename(path_image).replace('.nii', '_left.nii'))
if (not os.path.exists(path_image_first_crop_l)) | recompute:
im, aff, h = utils.load_volume(path_image, im_only=False)
im, aff = edit_volumes.crop_volume_with_idx(im, croppping_idx, aff)
utils.save_volume(im, aff, h, path_image_first_crop_l)
shape_array[2*idx, :] = np.array(lab_l.shape)
# crop right hippo and flip them
path_label_first_crop_r = os.path.join(tmp_result_dir,
os.path.basename(path_label).replace('.nii', '_right_flipped.nii'))
lab, aff, h = utils.load_volume(path_label, im_only=False)
lab_r, croppping_idx, aff_r = edit_volumes.crop_volume_around_region(lab, crop_margin,
list(range(20001, 20009)), aff=aff)
if (not os.path.exists(path_label_first_crop_r)) | recompute:
lab_r = edit_volumes.flip_volume(lab_r, direction='rl', aff=aff_r)
utils.save_volume(lab_r, aff_r, h, path_label_first_crop_r)
else:
lab_r = utils.load_volume(path_label_first_crop_r)
if path_image is not None:
path_image_first_crop_r = os.path.join(tmp_image_result_dir,
os.path.basename(path_image).replace('.nii', '_right.nii'))
if (not os.path.exists(path_image_first_crop_r)) | recompute:
im, aff, h = utils.load_volume(path_image, im_only=False)
im, aff = edit_volumes.crop_volume_with_idx(im, croppping_idx, aff)
im = edit_volumes.flip_volume(im, direction='rl', aff=aff)
utils.save_volume(im, aff, h, path_image_first_crop_r)
shape_array[2*idx+1, :] = np.array(lab_r.shape)
# list croppped files
path_labels_first_cropped = utils.list_images_in_folder(tmp_result_dir)
if tmp_image_result_dir is not None:
path_images_first_cropped = utils.list_images_in_folder(tmp_image_result_dir)
else:
path_images_first_cropped = [None] * len(path_labels_first_cropped)
# crop all label maps to same size
print('\nequalising shapes')
new_shape = np.min(shape_array, axis=0).astype('int32')
for i, (path_label, path_image) in enumerate(zip(path_labels_first_cropped, path_images_first_cropped)):
utils.print_loop_info(i, len(path_labels_first_cropped), 1)
# get cropping indices
path_lab_cropped = os.path.join(result_dir, os.path.basename(path_label))
lab, aff, h = utils.load_volume(path_label, im_only=False)
lab_shape = lab.shape
min_cropping = np.array([np.maximum(int((lab_shape[i]-new_shape[i])/2), 0) for i in range(3)])
max_cropping = np.array([min_cropping[i] + new_shape[i] for i in range(3)])
# crop labels and realign on adni format
if (not os.path.exists(path_lab_cropped)) | recompute:
lab, aff = edit_volumes.crop_volume_with_idx(lab, np.concatenate([min_cropping, max_cropping]), aff)
# realign on adni format
lab = np.flip(lab, axis=2)
aff[0:3, 0:3] = np.array([[-0.6, 0, 0], [0, 0, -0.6], [0, -0.6, 0]])
utils.save_volume(lab, aff, h, path_lab_cropped)
# crop image and realign on adni format
if path_image is not None:
path_im_cropped = os.path.join(image_result_dir, os.path.basename(path_image))
if (not os.path.exists(path_im_cropped)) | recompute:
im, aff, h = utils.load_volume(path_image, im_only=False)
im, aff = edit_volumes.crop_volume_with_idx(im, np.concatenate([min_cropping, max_cropping]), aff)
im = np.flip(im, axis=2)
aff[0:3, 0:3] = np.array([[-0.6, 0, 0], [0, 0, -0.6], [0, -0.6, 0]])
im = edit_volumes.mask_volume(im, lab)
utils.save_volume(im, aff, h, path_im_cropped)
# correct all labels to left values
print('\ncorrecting labels')
list_incorrect_labels = [77, 80, 251, 252, 253, 254, 255, 29, 41, 42, 43, 44, 46, 47, 49, 50, 51, 52, 54, 58, 60,
61, 62, 63, 7012, 20001, 20002, 20004, 20005, 20006, 20007, 20008]
list_correct_labels = [2, 3, 2, 2, 2, 2, 2, 2, 2, 3, 4, 5, 7, 8, 10, 11, 12, 13, 18, 26, 28, 2, 30, 31, 20108,
20101, 20102, 20104, 20105, 20106, 20107, 20108]
edit_volumes.correct_labels_in_dir(result_dir, list_incorrect_labels, list_correct_labels, result_dir)
# smooth labels
if smooth:
print('\nsmoothing labels')
edit_volumes.smooth_labels_in_dir(result_dir, result_dir)
def prepare_output_files(path_images, out_seg, out_posteriors, out_volumes):
# prepare input/output volumes
if ('nii.gz' not in path_images) & ('.mgz' not in path_images) & (
'.npz' not in path_images):
images_to_segment = utils.list_images_in_folder(path_images)
assert len(images_to_segment) > 0, "Could not find any training data"
if out_seg:
if not os.path.exists(out_seg):
os.mkdir(out_seg)
out_seg = [
os.path.join(out_seg, os.path.basename(image)).replace(
'.nii.gz', '_seg.nii.gz') for image in images_to_segment
]
out_seg = [
seg_path.replace('.mgz', '_seg.mgz') for seg_path in out_seg
]
out_seg = [
seg_path.replace('.npz', '_seg.npz') for seg_path in out_seg
]
else:
out_seg = [out_seg] * len(images_to_segment)
if out_posteriors:
if not os.path.exists(out_posteriors):
os.mkdir(out_posteriors)
out_posteriors = [
os.path.join(out_posteriors, os.path.basename(image)).replace(
'.nii.gz', '_posteriors.nii.gz')
for image in images_to_segment
]
out_posteriors = [
posteriors_path.replace('.mgz', '_posteriors.mgz')
for posteriors_path in out_posteriors
]
out_posteriors = [
posteriors_path.replace('.npz', '_posteriors.npz')
for posteriors_path in out_posteriors
]
else:
out_posteriors = [out_posteriors] * len(images_to_segment)
if out_volumes:
if out_volumes[-4:] != '.csv':
out_volumes += '.csv'
if not os.path.exists(os.path.dirname(out_volumes)):
os.mkdir(os.path.dirname(out_volumes))
else:
assert os.path.exists(path_images), "Could not find image to segment"
images_to_segment = [path_images]
if out_seg is not None:
if ('nii.gz' not in out_seg) & ('.mgz' not in out_seg) & (
'.npz' not in out_seg):
if not os.path.exists(out_seg):
os.mkdir(out_seg)
filename = os.path.basename(path_images).replace(
'.nii.gz', '_seg.nii.gz')
filename = filename.replace('mgz', '_seg.mgz')
filename = filename.replace('.npz', '_seg.npz')
out_seg = os.path.join(out_seg, filename)
out_seg = [out_seg]
if out_posteriors is not None:
if ('nii.gz' not in out_posteriors) & (
'.mgz' not in out_posteriors) & ('.npz'
not in out_posteriors):
if not os.path.exists(out_posteriors):
os.mkdir(out_posteriors)
filename = os.path.basename(path_images).replace(
'.nii.gz', '_posteriors.nii.gz')
filename = filename.replace('mgz', '_posteriors.mgz')
filename = filename.replace('.npz', '_posteriors.npz')
out_posteriors = os.path.join(out_posteriors, filename)
out_posteriors = [out_posteriors]
return images_to_segment, out_seg, out_posteriors, out_volumes
def preprocess_asegs(aseg_dir,
lesion_gt_dir,
list_incorrect,
list_correct,
lesion_label_in_gt=77,
dilate=2,
recompute=False):
# align asegs to gt dir (cropping to same dimension)
cropped_dir = aseg_dir + '_cropped'
edit_volumes.mri_convert_images_in_dir(aseg_dir,
cropped_dir,
interpolation='nearest',
reference_dir=lesion_gt_dir,
recompute=recompute)
# correct for aseg labels
corrected_dir = cropped_dir + '_corrected'
edit_volumes.correct_labels_in_dir(cropped_dir,
list_incorrect,
list_correct,
corrected_dir,
smooth=False,
recompute=recompute)
# list gt and aseg, and create result dir
list_lesion_labels = utils.list_images_in_folder(lesion_gt_dir)
list_aseg_labels = utils.list_images_in_folder(corrected_dir)
inpainted_dir = corrected_dir + '_lesion_inpainted'
utils.mkdir(inpainted_dir)
# loop over subjects
for path_lesion_label, path_aseg_label in zip(list_lesion_labels,
list_aseg_labels):
path_result = os.path.join(inpainted_dir,
os.path.basename(path_aseg_label))
if (not os.path.isfile(path_result)) | recompute:
# paste lesion label
lesions = utils.load_volume(path_lesion_label)
aseg_label, aff, h = utils.load_volume(path_aseg_label,
im_only=False)
lesion_mask = lesions == lesion_label_in_gt
aseg_label[lesion_mask] = 77
utils.save_volume(aseg_label, aff, h, path_result)
# dilate lesion and ventricle
dilated_dir = inpainted_dir + '_dilated'
utils.mkdir(dilated_dir)
list_inpainted_aseg = utils.list_images_in_folder(inpainted_dir)
for path_aseg in list_inpainted_aseg:
path_result = os.path.join(dilated_dir, os.path.basename(path_aseg))
if (not os.path.isfile(path_result)) | recompute:
# define lesion, WM, and LV masks
aseg, aff, h = utils.load_volume(path_aseg, im_only=False)
WM = aseg == 2
LV = aseg == 4
lesion = aseg == 77
# morphological operations to bridge the gaps between lesions and LV
morph_struct = utils.build_binary_structure(
dilate, len(aseg.shape))
dilated_LV_or_lesion = binary_dilation(LV | lesion, morph_struct)
filled_LV_or_lesion = binary_erosion(dilated_LV_or_lesion,
morph_struct)
LV = LV | (filled_LV_or_lesion & WM)
aseg[LV] = 4
# save map
utils.save_volume(aseg, aff, h, path_result)
def dice_evaluation(gt_dir,
seg_dir,
label_list,
compute_distances=False,
compute_score_whole_structure=False,
path_dice=None,
path_hausdorff=None,
path_mean_distance=None,
crop_margin_around_gt=10,
recompute=True,
verbose=True):
"""This function computes Dice scores between two sets of labels maps in gt_dir (ground truth) and seg_dir
(typically predictions). Labels maps in both folders are matched by sorting order.
:param gt_dir: path of directory with gt label maps
:param seg_dir: path of directory with label maps to compare to gt_dir. Matched to gt label maps by sorting order.
:param label_list: list of label values for which to compute evaluation metrics. Can be a sequence, a 1d numpy
array, or the path to such array.
:param compute_distances: (optional) whether to compute distances (Hausdorff and mean distance) between the surfaces
of GT and predicted labels. Default is False.
:param compute_score_whole_structure: (optional) whether to also compute the selected scores for the whole segmented
structure (i.e. scores are computed for a single structure obtained by regrouping all non-zero values). If True, the
resulting scores are added as an extra row to the result matrices. Default is False.
:param path_dice: path where the resulting Dice will be writen as numpy array.
Default is None, where the array is not saved.
:param path_hausdorff: path where the resulting Hausdorff distances will be writen as numpy array (only if
compute_distances is True). Default is None, where the array is not saved.
:param path_mean_distance: path where the resulting mean distances will be writen as numpy array (only if
compute_distances is True). Default is None, where the array is not saved.
:param crop_margin_around_gt: (optional) margin by which to crop around the gt volumes, in order to copute the
scores more efficiently. If None, no cropping is performed.
:param recompute: (optional) whether to recompute the already existing results. Default is True.
:param verbose: (optional) whether to print out info about the remaining number of cases.
:return: numpy array containing all Dice scores (labels in rows, subjects in columns). Also returns numpy arrays
with the same structures for Hausdorff and mean distances if compute_distances is True.
"""
# check whether to recompute
compute_dice = not os.path.isfile(path_dice) if (path_dice
is not None) else True
if compute_distances:
compute_hausdorff = not os.path.isfile(path_hausdorff) if (
path_hausdorff is not None) else True
compute_mean_dist = not os.path.isfile(path_mean_distance) if (
path_mean_distance is not None) else True
else:
compute_hausdorff = compute_mean_dist = False
if compute_dice | compute_hausdorff | compute_mean_dist | recompute:
# get list label maps to compare
path_gt_labels = utils.list_images_in_folder(gt_dir)
path_segs = utils.list_images_in_folder(seg_dir)
if len(path_gt_labels) != len(path_segs):
print(
'gt and segmentation folders must have the same amount of label maps.'
)
# load labels list
label_list, _ = utils.get_list_labels(label_list=label_list,
FS_sort=True,
labels_dir=gt_dir)
n_labels = len(label_list)
# initialise result matrices
if compute_score_whole_structure:
max_dists = np.zeros((n_labels + 1, len(path_segs)))
mean_dists = np.zeros((n_labels + 1, len(path_segs)))
dice_coefs = np.zeros((n_labels + 1, len(path_segs)))
else:
max_dists = np.zeros((n_labels, len(path_segs)))
mean_dists = np.zeros((n_labels, len(path_segs)))
dice_coefs = np.zeros((n_labels, len(path_segs)))
# loop over segmentations
loop_info = utils.LoopInfo(len(path_segs), 10, 'evaluating')
for idx, (path_gt,
path_seg) in enumerate(zip(path_gt_labels, path_segs)):
if verbose:
loop_info.update(idx)
# load gt labels and segmentation
gt_labels = utils.load_volume(path_gt, dtype='int')
seg = utils.load_volume(path_seg, dtype='int')
# crop images
if crop_margin_around_gt is not None:
gt_labels, cropping = edit_volumes.crop_volume_around_region(
gt_labels, margin=crop_margin_around_gt)
seg = edit_volumes.crop_volume_with_idx(seg, cropping)
# compute Dice scores
dice_coefs[:n_labels, idx] = fast_dice(gt_labels, seg, label_list)
# compute Dice scores for whole structures
if compute_score_whole_structure:
temp_gt = (gt_labels > 0) * 1
temp_seg = (seg > 0) * 1
dice_coefs[-1, idx] = dice(temp_gt, temp_seg)
else:
temp_gt = temp_seg = None
# compute average and Hausdorff distances
if compute_distances:
# compute unique label values
unique_gt_labels = np.unique(gt_labels)
unique_seg_labels = np.unique(seg)
# compute max/mean surface distances for all labels
for index, label in enumerate(label_list):
if (label in unique_gt_labels) & (label
in unique_seg_labels):
mask_gt = np.where(gt_labels == label, True, False)
mask_seg = np.where(seg == label, True, False)
max_dists[index,
idx], mean_dists[index,
idx] = surface_distances(
mask_gt, mask_seg)
else:
max_dists[index, idx] = max(gt_labels.shape)
mean_dists[index, idx] = max(gt_labels.shape)
# compute max/mean distances for whole structure
if compute_score_whole_structure:
max_dists[-1, idx], mean_dists[-1,
idx] = surface_distances(
temp_gt, temp_seg)
# write results
if path_dice is not None:
utils.mkdir(os.path.dirname(path_dice))
np.save(path_dice, dice_coefs)
if compute_distances and path_hausdorff is not None:
utils.mkdir(os.path.dirname(path_hausdorff))
np.save(path_hausdorff, max_dists)
if compute_distances and path_mean_distance is not None:
utils.mkdir(os.path.dirname(path_mean_distance))
np.save(path_mean_distance, mean_dists)
else:
dice_coefs = np.load(path_dice)
if compute_distances:
max_dists = np.load(path_hausdorff)
mean_dists = np.load(path_mean_distance)
else:
max_dists = mean_dists = None
if compute_distances:
return dice_coefs, max_dists, mean_dists
else:
return dice_coefs, None, None
def postprocess_samseg(list_samseg_dir,
list_gt_dir,
path_segmentation_labels,
incorrect_labels,
correct_labels,
list_posteriors_dir=None,
list_thresholds=None,
recompute=False):
""" This function processes the samseg segmentations: it corrects the labels (right/left and 99 to 77), resamples
them to the space of gt_dir, and computes the Dice scores for 1) all_subjects vs. testing subjects only, and 2) all
ROIs vs. lesions only.
It requires that all segmentations are sorted in three subfolders inside samseg_main_dir: t1, flair, and t1_flair.
IMPORTANT: Images are expected to have to following naming convention: <subject_id>.samseg.<contrast>.lesion.mgz,
where <contrast> must either be t1, flair, ***t1_flair***
:param list_samseg_dir: main samseg dir containing the three subfolders t1, flair, t1_flair
:param list_gt_dir: folder with the gt label maps for all subjects
:param path_segmentation_labels: list of segmentation labels
:param incorrect_labels: list of samseg incorrect labels
:param correct_labels: list of labels to correct the wrong one with
:param recompute: whether to recompute files
"""
if list_posteriors_dir is None:
list_posteriors_dir = [None] * len(list_samseg_dir)
for samseg_dir, gt_dir, posteriors_dir, threshold in zip(
list_samseg_dir, list_gt_dir, list_posteriors_dir,
list_thresholds):
# define result directories
samseg_corrected_dir = samseg_dir + '_corrected'
samseg_preprocessed_dir = samseg_dir + '_preprocessed'
if (not os.path.isdir(samseg_preprocessed_dir)) | recompute:
# regroup right/left labels and change 99 to 77
edit_volumes.correct_labels_in_dir(samseg_dir,
incorrect_labels,
correct_labels,
samseg_corrected_dir,
recompute=recompute)
# resample to gt format
edit_volumes.mri_convert_images_in_dir(samseg_corrected_dir,
samseg_preprocessed_dir,
interpolation='nearest',
reference_dir=gt_dir,
recompute=recompute)
# replace lesions by thresholded lesion posteriors
if posteriors_dir is not None:
# resample posteriors to gt format
posteriors_preprocessed_dir = posteriors_dir + '_preprocessed'
edit_volumes.mri_convert_images_in_dir(posteriors_dir,
posteriors_preprocessed_dir,
reference_dir=gt_dir,
recompute=recompute)
# list hard segmentations and posteriors
samseg_postprocessed_dir = samseg_dir + '_postprocessed'
utils.mkdir(samseg_postprocessed_dir)
path_segs = [
path for path in utils.list_images_in_folder(
samseg_preprocessed_dir)
]
path_posteriors = [
path for path in utils.list_images_in_folder(
posteriors_preprocessed_dir)
]
for subject_idx, (path_seg, path_post) in enumerate(
zip(path_segs, path_posteriors)):
path_result = os.path.join(samseg_postprocessed_dir,
os.path.basename(path_seg))
if (not os.path.isfile(path_result)) | recompute:
# replace segmented lesions by thresholded posteriors
seg, aff, h = utils.load_volume(path_seg, im_only=False)
posteriors = utils.load_volume(path_post)
seg[seg == 77] = 2
seg[posteriors > threshold] = 77
utils.save_volume(seg, aff, h, path_result)
else:
samseg_postprocessed_dir = samseg_preprocessed_dir
# compute dice scores with
path_dice_testing = os.path.join(samseg_postprocessed_dir, 'dice.npy')
path_dice_lesions_testing = os.path.join(samseg_postprocessed_dir,
'dice_lesions.npy')
if (not os.path.isfile(path_dice_testing)) | recompute:
dice_evaluation(gt_dir, samseg_postprocessed_dir,
path_segmentation_labels, path_dice_testing)
if (not os.path.isfile(path_dice_lesions_testing)) | recompute:
dice = np.load(path_dice_testing)
np.save(path_dice_lesions_testing, dice[4, :])
unet_model.load_weights(str(model_path), by_name=True)
# Prepare list of images to process
path_images = os.path.abspath(args['path_images'])
basename = os.path.basename(path_images)
path_predictions = os.path.abspath(args['path_predictions'])
# prepare input/output volumes
# First case: you're providing directories
if ('.nii.gz' not in basename) & ('.nii' not in basename) & (
'.mgz' not in basename) & ('.npz' not in basename):
if os.path.isfile(path_images):
raise Exception(
'extension not supported for %s, only use: nii.gz, .nii, .mgz, or .npz'
% path_images)
images_to_segment = utils.list_images_in_folder(path_images)
utils.mkdir(path_predictions)
path_predictions = [
os.path.join(path_predictions,
os.path.basename(image)).replace('.nii', '_SynthSR.nii')
for image in images_to_segment
]
path_predictions = [
seg_path.replace('.mgz', '_SynthSR.mgz')
for seg_path in path_predictions
]
path_predictions = [
seg_path.replace('.npz', '_SynthSR.npz')
for seg_path in path_predictions
]
def evaluation(gt_dir,
seg_dir,
label_list,
mask_dir=None,
compute_score_whole_structure=False,
path_dice=None,
path_hausdorff=None,
path_hausdorff_99=None,
path_hausdorff_95=None,
path_mean_distance=None,
crop_margin_around_gt=10,
list_incorrect_labels=None,
list_correct_labels=None,
use_nearest_label=False,
recompute=True,
verbose=True):
"""This function computes Dice scores, as well as surface distances, between two sets of labels maps in gt_dir
(ground truth) and seg_dir (typically predictions). Labels maps in both folders are matched by sorting order.
The resulting scores are saved at the specified locations.
:param gt_dir: path of directory with gt label maps
:param seg_dir: path of directory with label maps to compare to gt_dir. Matched to gt label maps by sorting order.
:param label_list: list of label values for which to compute evaluation metrics. Can be a sequence, a 1d numpy
array, or the path to such array.
:param mask_dir: (optional) path of directory with masks of areas to ignore for each evaluated segmentation.
Matched to gt label maps by sorting order. Default is None, where nothing is masked.
:param compute_score_whole_structure: (optional) whether to also compute the selected scores for the whole segmented
structure (i.e. scores are computed for a single structure obtained by regrouping all non-zero values). If True, the
resulting scores are added as an extra row to the result matrices. Default is False.
:param path_dice: path where the resulting Dice will be writen as numpy array.
Default is None, where the array is not saved.
:param path_hausdorff: path where the resulting Hausdorff distances will be writen as numpy array (only if
compute_distances is True). Default is None, where the array is not saved.
:param path_hausdorff_99: same as for path_hausdorff but for the 99th percentile of the boundary distance.
:param path_hausdorff_95: same as for path_hausdorff but for the 95th percentile of the boundary distance.
:param path_mean_distance: path where the resulting mean distances will be writen as numpy array (only if
compute_distances is True). Default is None, where the array is not saved.
:param crop_margin_around_gt: (optional) margin by which to crop around the gt volumes, in order to copute the
scores more efficiently. If None, no cropping is performed.
:param list_incorrect_labels: (optional) this option enables to replace some label values in the maps in seg_dir by
other label values. Can be a list, a 1d numpy array, or the path to such an array.
The incorrect labels can then be replaced either by specified values, or by the nearest value (see below).
:param list_correct_labels: (optional) list of values to correct the labels specified in list_incorrect_labels.
Correct values must have the same order as their corresponding value in list_incorrect_labels.
:param use_nearest_label: (optional) whether to correct the incorrect lavel values with the nearest labels.
:param recompute: (optional) whether to recompute the already existing results. Default is True.
:param verbose: (optional) whether to print out info about the remaining number of cases.
"""
# check whether to recompute
compute_dice = not os.path.isfile(path_dice) if (path_dice
is not None) else True
compute_hausdorff = not os.path.isfile(path_hausdorff) if (
path_hausdorff is not None) else False
compute_hausdorff_99 = not os.path.isfile(path_hausdorff_99) if (
path_hausdorff_99 is not None) else False
compute_hausdorff_95 = not os.path.isfile(path_hausdorff_95) if (
path_hausdorff_95 is not None) else False
compute_mean_dist = not os.path.isfile(path_mean_distance) if (
path_mean_distance is not None) else False
compute_hd = [
compute_hausdorff, compute_hausdorff_99, compute_hausdorff_95
]
if compute_dice | any(compute_hd) | compute_mean_dist | recompute:
# get list label maps to compare
path_gt_labels = utils.list_images_in_folder(gt_dir)
path_segs = utils.list_images_in_folder(seg_dir)
path_gt_labels = utils.reformat_to_list(path_gt_labels,
length=len(path_segs))
if len(path_gt_labels) != len(path_segs):
print(
'gt and segmentation folders must have the same amount of label maps.'
)
if mask_dir is not None:
path_masks = utils.list_images_in_folder(mask_dir)
if len(path_masks) != len(path_segs):
print('not the same amount of masks and segmentations.')
else:
path_masks = [None] * len(path_segs)
# load labels list
label_list, _ = utils.get_list_labels(label_list=label_list,
FS_sort=True,
labels_dir=gt_dir)
n_labels = len(label_list)
max_label = np.max(label_list) + 1
# initialise result matrices
if compute_score_whole_structure:
max_dists = np.zeros((n_labels + 1, len(path_segs), 3))
mean_dists = np.zeros((n_labels + 1, len(path_segs)))
dice_coefs = np.zeros((n_labels + 1, len(path_segs)))
else:
max_dists = np.zeros((n_labels, len(path_segs), 3))
mean_dists = np.zeros((n_labels, len(path_segs)))
dice_coefs = np.zeros((n_labels, len(path_segs)))
# loop over segmentations
loop_info = utils.LoopInfo(len(path_segs),
10,
'evaluating',
print_time=True)
for idx, (path_gt, path_seg, path_mask) in enumerate(
zip(path_gt_labels, path_segs, path_masks)):
if verbose:
loop_info.update(idx)
# load gt labels and segmentation
gt_labels = utils.load_volume(path_gt, dtype='int')
seg = utils.load_volume(path_seg, dtype='int')
if path_mask is not None:
mask = utils.load_volume(path_mask, dtype='bool')
gt_labels[mask] = max_label
seg[mask] = max_label
# crop images
if crop_margin_around_gt is not None:
gt_labels, cropping = edit_volumes.crop_volume_around_region(
gt_labels, margin=crop_margin_around_gt)
seg = edit_volumes.crop_volume_with_idx(seg, cropping)
if list_incorrect_labels is not None:
seg = edit_volumes.correct_label_map(seg,
list_incorrect_labels,
list_correct_labels,
use_nearest_label)
# compute Dice scores
dice_coefs[:n_labels, idx] = fast_dice(gt_labels, seg, label_list)
# compute Dice scores for whole structures
if compute_score_whole_structure:
temp_gt = (gt_labels > 0) * 1
temp_seg = (seg > 0) * 1
dice_coefs[-1, idx] = dice(temp_gt, temp_seg)
else:
temp_gt = temp_seg = None
# compute average and Hausdorff distances
if any(compute_hd) | compute_mean_dist:
# compute unique label values
unique_gt_labels = np.unique(gt_labels)
unique_seg_labels = np.unique(seg)
# compute max/mean surface distances for all labels
for index, label in enumerate(label_list):
if (label in unique_gt_labels) & (label
in unique_seg_labels):
mask_gt = np.where(gt_labels == label, True, False)
mask_seg = np.where(seg == label, True, False)
tmp_max_dists, mean_dists[index,
idx] = surface_distances(
mask_gt, mask_seg,
[100, 99, 95])
max_dists[index, idx, :] = np.array(tmp_max_dists)
else:
mean_dists[index, idx] = max(gt_labels.shape)
max_dists[index, idx, :] = np.array(
[max(gt_labels.shape)] * 3)
# compute max/mean distances for whole structure
if compute_score_whole_structure:
tmp_max_dists, mean_dists[-1, idx] = surface_distances(
temp_gt, temp_seg, [100, 99, 95])
max_dists[-1, idx, :] = np.array(tmp_max_dists)
# write results
if path_dice is not None:
utils.mkdir(os.path.dirname(path_dice))
np.save(path_dice, dice_coefs)
if path_hausdorff is not None:
utils.mkdir(os.path.dirname(path_hausdorff))
np.save(path_hausdorff, max_dists[..., 0])
if path_hausdorff_99 is not None:
utils.mkdir(os.path.dirname(path_hausdorff_99))
np.save(path_hausdorff_99, max_dists[..., 1])
if path_hausdorff_95 is not None:
utils.mkdir(os.path.dirname(path_hausdorff_95))
np.save(path_hausdorff_95, max_dists[..., 2])
if path_mean_distance is not None:
utils.mkdir(os.path.dirname(path_mean_distance))
np.save(path_mean_distance, max_dists[..., 2])
def prepare_output_files(path_images, out_seg, out_posteriors, out_volumes):
# convert path to absolute paths
path_images = os.path.abspath(path_images)
if out_seg is not None:
out_seg = os.path.abspath(out_seg)
if out_posteriors is not None:
out_posteriors = os.path.abspath(out_posteriors)
if out_volumes is not None:
out_volumes = os.path.abspath(out_volumes)
# prepare input/output volumes
if ('.nii.gz' not in path_images) & ('.nii' not in path_images) & ('.mgz' not in path_images) & \
('.npz' not in path_images):
images_to_segment = utils.list_images_in_folder(path_images)
assert len(images_to_segment) > 0, "Could not find any training data"
if out_seg:
utils.mkdir(out_seg)
out_seg = [os.path.join(out_seg, os.path.basename(image)).replace('.nii', '_seg.nii') for image in
images_to_segment]
out_seg = [seg_path.replace('.mgz', '_seg.mgz') for seg_path in out_seg]
out_seg = [seg_path.replace('.npz', '_seg.npz') for seg_path in out_seg]
else:
out_seg = [out_seg] * len(images_to_segment)
if out_posteriors:
utils.mkdir(out_posteriors)
out_posteriors = [os.path.join(out_posteriors, os.path.basename(image)).replace('.nii',
'_posteriors.nii') for image in images_to_segment]
out_posteriors = [posteriors_path.replace('.mgz', '_posteriors.mgz')
for posteriors_path in out_posteriors]
out_posteriors = [posteriors_path.replace('.npz', '_posteriors.npz')
for posteriors_path in out_posteriors]
else:
out_posteriors = [out_posteriors] * len(images_to_segment)
else:
assert os.path.exists(path_images), "Could not find image to segment"
images_to_segment = [path_images]
if out_seg is not None:
if ('.nii.gz' not in out_seg) & ('.nii' not in out_seg) & ('.mgz' not in out_seg) & ('.npz' not in out_seg):
utils.mkdir(out_seg)
filename = os.path.basename(path_images).replace('.nii', '_seg.nii')
filename = filename.replace('mgz', '_seg.mgz')
filename = filename.replace('.npz', '_seg.npz')
out_seg = os.path.join(out_seg, filename)
else:
utils.mkdir(os.path.dirname(out_seg))
out_seg = [out_seg]
if out_posteriors is not None:
if ('.nii.gz' not in out_posteriors) & ('.nii' not in out_posteriors) & ('.mgz' not in out_posteriors) & \
('.npz' not in out_posteriors):
utils.mkdir(out_posteriors)
filename = os.path.basename(path_images).replace('.nii', '_posteriors.nii')
filename = filename.replace('mgz', '_posteriors.mgz')
filename = filename.replace('.npz', '_posteriors.npz')
out_posteriors = os.path.join(out_posteriors, filename)
else:
utils.mkdir(os.path.dirname(out_posteriors))
out_posteriors = [out_posteriors]
if out_volumes:
if out_volumes[-4:] != '.csv':
print('out_volumes provided without csv extension. Adding csv extension to output_volumes.')
out_volumes += '.csv'
utils.mkdir(os.path.dirname(out_volumes))
return images_to_segment, out_seg, out_posteriors, out_volumes
