def preprocess_image(im_path, n_levels, crop_shape=None, padding=None, aff_ref='FS', dist_map=False):
# read image and corresponding info
im, shape, aff, n_dims, n_channels, header, im_res = utils.get_volume_info(im_path, return_volume=True)
if padding:
im = edit_volumes.pad_volume(im, padding_shape=padding)
pad_shape = im.shape[:n_dims]
else:
pad_shape = shape
# check that patch_shape or im_shape are divisible by 2**n_levels
if crop_shape is not None:
crop_shape = utils.reformat_to_list(crop_shape, length=n_dims, dtype='int')
if not all([pad_shape[i] >= crop_shape[i] for i in range(len(pad_shape))]):
crop_shape = [min(pad_shape[i], crop_shape[i]) for i in range(n_dims)]
if not all([size % (2**n_levels) == 0 for size in crop_shape]):
crop_shape = [utils.find_closest_number_divisible_by_m(size, 2 ** n_levels) for size in crop_shape]
else:
if not all([size % (2**n_levels) == 0 for size in pad_shape]):
crop_shape = [utils.find_closest_number_divisible_by_m(size, 2 ** n_levels) for size in pad_shape]
# crop image if necessary
if crop_shape is not None:
im, crop_idx = edit_volumes.crop_volume(im, cropping_shape=crop_shape, return_crop_idx=True)
else:
crop_idx = None
# align image to training axes and directions
if n_dims > 2:
if aff_ref == 'FS':
aff_ref = np.array([[-1., 0., 0., 0.], [0., 0., 1., 0.], [0., -1., 0., 0.], [0., 0., 0., 1.]])
im = edit_volumes.align_volume_to_ref(im, aff, aff_ref=aff_ref, return_aff=False, n_dims=n_dims)
elif aff_ref == 'identity':
aff_ref = np.eye(4)
im = edit_volumes.align_volume_to_ref(im, aff, aff_ref=aff_ref, return_aff=False, n_dims=n_dims)
# normalise image
if n_channels == 1:
m = np.min(im)
M = np.max(im)
if M == m:
im = np.zeros(im.shape)
else:
im = (im - m) / (M - m)
else:
for i in range(im.shape[-1]):
if (not dist_map) | (dist_map & (i % 2 == 0)):
channel = im[..., i]
m = np.min(channel)
M = np.max(channel)
if M == m:
im[..., i] = np.zeros(channel.shape)
else:
im[..., i] = (channel - m) / (M - m)
# add batch and channel axes
im = utils.add_axis(im) if n_channels > 1 else utils.add_axis(im, axis=[0, -1])
return im, aff, header, im_res, n_channels, n_dims, shape, pad_shape, crop_idx
def build_model_inputs(path_images,
path_label_maps,
batchsize=1):
# get label info
_, _, n_dims, n_channels, _, _ = utils.get_volume_info(path_images[0])
# Generate!
while True:
# randomly pick as many images as batchsize
indices = npr.randint(len(path_label_maps), size=batchsize)
# initialise input lists
list_images = list()
list_label_maps = list()
for idx in indices:
# add image
image = utils.load_volume(path_images[idx], aff_ref=np.eye(4))
if n_channels > 1:
list_images.append(utils.add_axis(image, axis=0))
else:
list_images.append(utils.add_axis(image, axis=[0, -1]))
# add labels
labels = utils.load_volume(path_label_maps[idx], dtype='int', aff_ref=np.eye(4))
list_label_maps.append(utils.add_axis(labels, axis=[0, -1]))
# build list of inputs of augmentation model
list_inputs = [list_images, list_label_maps]
if batchsize > 1: # concatenate individual input types if batchsize > 1
list_inputs = [np.concatenate(item, 0) for item in list_inputs]
else:
list_inputs = [item[0] for item in list_inputs]
yield list_inputs
def build_model_inputs(path_label_maps,
n_labels,
batchsize=1,
n_channels=1,
generation_classes=None,
prior_distributions='uniform',
prior_means=None,
prior_stds=None,
use_specific_stats_for_channel=False,
mix_prior_and_random=False,
apply_linear_trans=True,
scaling_bounds=None,
rotation_bounds=None,
shearing_bounds=None,
background_paths=None):
"""
This function builds a generator to be fed to the lab2im model. It enables to generate all the required inputs,
according to the operations performed in the model.
:param path_label_maps: list of the paths of the input label maps.
:param n_labels: number of labels in the input label maps.
: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 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 or 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.
: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, or 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.
: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.
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 background_paths: (optional) list of paths of label maps to replace the soft brain tissues (label 258) with.
"""
# get label info
_, _, n_dims, _, _, _ = utils.get_volume_info(path_label_maps[0])
# allocate unique class to each label if generation classes is not given
if generation_classes is None:
generation_classes = np.arange(n_labels)
# Generate!
while True:
# randomly pick as many images as batchsize
indices = npr.randint(len(path_label_maps), size=batchsize)
# initialise input lists
list_label_maps = []
list_means = []
list_stds = []
list_affine_transforms = []
for idx in indices:
# add labels to inputs
y = utils.load_volume(path_label_maps[idx], dtype='int', aff_ref=np.eye(4))
if background_paths is not None:
idx_258 = np.where(y == 258)
if np.any(idx_258):
background = utils.load_volume(background_paths[npr.randint(len(background_paths))],
dtype='int', aff_ref=np.eye(4))
background_shape = background.shape
if np.all(np.array(background_shape) == background_shape[0]): # flip if same dimensions
background = np.flip(background, tuple([i for i in range(3) if np.random.normal() > 0]))
assert background.shape == y.shape, 'background patches should have same shape than training ' \
'labels. Had {0} and {1}'.format(background.shape, y.shape)
y[idx_258] = background[idx_258]
list_label_maps.append(utils.add_axis(y, axis=-2))
# add means and standard deviations to inputs
means = np.empty((n_labels, 0))
stds = np.empty((n_labels, 0))
for channel in range(n_channels):
# retrieve channel specific stats if necessary
if isinstance(prior_means, np.ndarray):
if (prior_means.shape[0] > 2) & use_specific_stats_for_channel:
if prior_means.shape[0] / 2 != n_channels:
raise ValueError("the number of blocks in prior_means does not match n_channels. This "
"message is printed because use_specific_stats_for_channel is True.")
tmp_prior_means = prior_means[2 * channel:2 * channel + 2, :]
else:
tmp_prior_means = prior_means
else:
tmp_prior_means = prior_means
if (prior_means is not None) & mix_prior_and_random & (npr.uniform() > 0.5):
tmp_prior_means = None
if isinstance(prior_stds, np.ndarray):
if (prior_stds.shape[0] > 2) & use_specific_stats_for_channel:
if prior_stds.shape[0] / 2 != n_channels:
raise ValueError("the number of blocks in prior_stds does not match n_channels. This "
"message is printed because use_specific_stats_for_channel is True.")
tmp_prior_stds = prior_stds[2 * channel:2 * channel + 2, :]
else:
tmp_prior_stds = prior_stds
else:
tmp_prior_stds = prior_stds
if (prior_stds is not None) & mix_prior_and_random & (npr.uniform() > 0.5):
tmp_prior_stds = None
# draw means and std devs from priors
tmp_classes_means = utils.draw_value_from_distribution(tmp_prior_means, n_labels, prior_distributions,
125., 100., positive_only=True)
tmp_classes_stds = utils.draw_value_from_distribution(tmp_prior_stds, n_labels, prior_distributions,
15., 10., positive_only=True)
tmp_means = utils.add_axis(tmp_classes_means[generation_classes], -1)
tmp_stds = utils.add_axis(tmp_classes_stds[generation_classes], -1)
means = np.concatenate([means, tmp_means], axis=1)
stds = np.concatenate([stds, tmp_stds], axis=1)
list_means.append(utils.add_axis(means))
list_stds.append(utils.add_axis(stds))
# add linear transform to inputs
if apply_linear_trans:
# get affine transformation: rotate, scale, shear (translation done during random cropping)
scaling = utils.draw_value_from_distribution(scaling_bounds, size=n_dims, centre=1, default_range=.15)
if n_dims == 2:
rotation = utils.draw_value_from_distribution(rotation_bounds, default_range=15.0)
else:
rotation = utils.draw_value_from_distribution(rotation_bounds, size=n_dims, default_range=15.0)
shearing = utils.draw_value_from_distribution(shearing_bounds, size=n_dims**2-n_dims, default_range=.01)
affine_transform = utils.create_affine_transformation_matrix(n_dims, scaling, rotation, shearing)
list_affine_transforms.append(utils.add_axis(affine_transform))
# build list of inputs of augmentation model
list_inputs = [list_label_maps, list_means, list_stds]
if apply_linear_trans:
list_inputs.append(list_affine_transforms)
# concatenate individual input types if batchsize > 1
if batchsize > 1:
list_inputs = [np.concatenate(item, 0) for item in list_inputs]
else:
list_inputs = [item[0] for item in list_inputs]
yield list_inputs
def predict(path_images,
path_model,
segmentation_label_list,
dist_map=False,
path_segmentations=None,
path_posteriors=None,
path_volumes=None,
segmentation_names_list=None,
padding=None,
cropping=None,
resample=None,
aff_ref='FS',
sigma_smoothing=0,
keep_biggest_component=False,
conv_size=3,
n_levels=5,
nb_conv_per_level=2,
unet_feat_count=24,
feat_multiplier=2,
activation='elu',
gt_folder=None,
evaluation_label_list=None,
compute_distances=False,
recompute=True,
verbose=True):
"""
This function uses trained models to segment images.
It is crucial that the inputs match the architecture parameters of the trained model.
:param path_images: path of the images to segment. Can be the path to a directory or the path to a single image.
:param path_model: path ot the trained model.
:param segmentation_label_list: List of labels for which to compute Dice scores. It should contain the same values
as the segmentation label list used for training the network.
Can be a sequence, a 1d numpy array, or the path to a numpy 1d array.
:param dist_map: (optional) whether the input will contain distance maps channels (between each intenisty channels)
Default is False.
:param path_segmentations: (optional) path where segmentations will be writen.
Should be a dir, if path_images is a dir, and afile if path_images is a file.
Should not be None, if path_posteriors is None.
:param path_posteriors: (optional) path where posteriors will be writen.
Should be a dir, if path_images is a dir, and afile if path_images is a file.
Should not be None, if path_segmentations is None.
:param path_volumes: (optional) path of a csv file where the soft volumes of all segmented regions will be writen.
The rows of the csv file correspond to subjects, and the columns correspond to segmentation labels.
The soft volume of a structure corresponds to the sum of its predicted probability map.
:param segmentation_names_list: (optional) List of names correponding to the names of the segmentation labels.
Only used when path_volumes is provided. Must be of the same size as segmentation_label_list. Can be given as a
list, a numpy array of strings, or the path to such a numpy array. Default is None.
:param padding: (optional) pad the images to the specified shape before predicting the segmentation maps.
Can be an int, a sequence or a 1d numpy array.
:param cropping: (optional) crop the images to the specified shape before predicting the segmentation maps.
If padding and cropping are specified, images are padded before being cropped.
Can be an int, a sequence or a 1d numpy array.
:param resample: (optional) resample the images to the specified resolution before predicting the segmentation maps.
Can be an int, a sequence or a 1d numpy array.
:param aff_ref: (optional) type of affine matrix of the images used for training. By default this is set to the
FreeSurfer orientation ('FS'), as it was the configuration in which SynthSeg was trained. However, the new models
are now trained on data aligned with identity vox2ras matrix, so you need to change aff_ref to 'identity'.
:param sigma_smoothing: (optional) If not None, the posteriors are smoothed with a gaussian kernel of the specified
standard deviation.
:param keep_biggest_component: (optional) whether to only keep the biggest component in the predicted segmentation.
:param conv_size: (optional) size of unet's convolution masks. Default is 3.
:param n_levels: (optional) number of levels for unet. Default is 5.
:param nb_conv_per_level: (optional) number of convolution layers per level. Default is 2.
:param unet_feat_count: (optional) number of features for the first layer of the unet. Default is 24.
:param feat_multiplier: (optional) multiplicative factor for the number of feature for each new level. Default is 2.
:param activation: (optional) activation function. Can be 'elu', 'relu'.
:param gt_folder: (optional) folder containing ground truth files for evaluation.
A numpy array containing all dice scores (labels in rows, subjects in columns) will be writen either at
segmentations_dir (if not None), or posteriors_dir.
:param evaluation_label_list: (optional) if gt_folder is True you can evaluate the Dice scores on a subset of the
segmentation labels, by providing another label list here. Can be a sequence, a 1d numpy array, or the path to a
numpy 1d array. Default is the same as segmentation_label_list.
:param recompute: (optional) whether to recompute segmentations that were already computed. This also applies to
Dice scores, if gt_folder is not None. Default is True.
:param verbose: (optional) whether to print out info about the remaining number of cases.
"""
# prepare output filepaths
images_to_segment, path_segmentations, path_posteriors, path_volumes, compute = \
prepare_output_files(path_images, path_segmentations, path_posteriors, path_volumes, recompute)
# get label and classes lists
label_list, n_neutral_labels = utils.get_list_labels(label_list=segmentation_label_list, FS_sort=True)
if evaluation_label_list is None:
evaluation_label_list = segmentation_label_list
# prepare volume file if needed
if path_volumes is not None:
if segmentation_names_list is not None:
csv_header = [[''] + utils.reformat_to_list(segmentation_names_list, load_as_numpy=True)]
csv_header += [[''] + [str(lab) for lab in label_list[1:]]]
else:
csv_header = [['subjects'] + [str(lab) for lab in label_list[1:]]]
with open(path_volumes, 'w') as csvFile:
writer = csv.writer(csvFile)
writer.writerows(csv_header)
csvFile.close()
# perform segmentation
net = None
previous_model_input_shape = None
loop_info = utils.LoopInfo(len(images_to_segment), 10, 'predicting', True)
for idx, (path_image, path_segmentation, path_posterior, tmp_compute) in enumerate(zip(images_to_segment,
path_segmentations,
path_posteriors,
compute)):
# compute segmentation only if needed
if tmp_compute:
# preprocess image and get information
image, aff, h, im_res, n_channels, n_dims, shape, pad_shape, crop_idx = \
preprocess_image(path_image, n_levels, cropping, padding, aff_ref=aff_ref, dist_map=dist_map)
model_input_shape = list(image.shape[1:])
# prepare net for first image or if input's size has changed
if (net is None) | (previous_model_input_shape != model_input_shape):
# check for image size compatibility
if (net is not None) & (previous_model_input_shape != model_input_shape) & verbose:
print('image of different shape as previous ones, redefining network')
previous_model_input_shape = model_input_shape
# build network
net = build_model(path_model, model_input_shape, resample, im_res, n_levels, len(label_list), conv_size,
nb_conv_per_level, unet_feat_count, feat_multiplier, activation, sigma_smoothing)
if verbose:
loop_info.update(idx)
# predict posteriors
prediction_patch = net.predict(image)
# get posteriors and segmentation
seg, posteriors = postprocess(prediction_patch, pad_shape, shape, crop_idx, n_dims, label_list,
keep_biggest_component, aff, aff_ref=aff_ref,
keep_biggest_of_each_group=keep_biggest_component,
n_neutral_labels=n_neutral_labels)
# write results to disk
if path_segmentation is not None:
utils.save_volume(seg.astype('int'), aff, h, path_segmentation)
if path_posterior is not None:
if n_channels > 1:
posteriors = utils.add_axis(posteriors, axis=[0, -1])
utils.save_volume(posteriors.astype('float'), aff, h, path_posterior)
else:
if path_volumes is not None:
posteriors, _, _, _, _, _, im_res = utils.get_volume_info(path_posterior, True, aff_ref=np.eye(4))
else:
posteriors = im_res = None
# compute volumes
if path_volumes is not None:
volumes = np.sum(posteriors[..., 1:], axis=tuple(range(0, len(posteriors.shape) - 1)))
volumes = np.around(volumes * np.prod(im_res), 3)
row = [os.path.basename(path_image).replace('.nii.gz', '')] + [str(vol) for vol in volumes]
with open(path_volumes, 'a') as csvFile:
writer = csv.writer(csvFile)
writer.writerow(row)
csvFile.close()
# evaluate
if gt_folder is not None:
# find path evaluation folder
path_first_result = path_segmentations[0] if (path_segmentations[0] is not None) else path_posteriors[0]
eval_folder = os.path.dirname(path_first_result)
# compute evaluation metrics
evaluate.dice_evaluation(gt_folder,
eval_folder,
evaluation_label_list,
compute_distances=compute_distances,
compute_score_whole_structure=False,
path_dice=os.path.join(eval_folder, 'dice.npy'),
path_hausdorff=os.path.join(eval_folder, 'hausdorff.npy'),
path_mean_distance=os.path.join(eval_folder, 'mean_distance.npy'),
recompute=recompute,
verbose=verbose)
def preprocess_image(im_path, n_levels, crop_shape=None, padding=None):
# read image and corresponding info
im, shape, aff, n_dims, n_channels, header, labels_res = utils.get_volume_info(
im_path, return_volume=True)
if padding:
if n_channels == 1:
im = np.pad(im, padding, mode='constant')
pad_shape = im.shape
else:
im = np.pad(im,
tuple([(padding, padding)] * n_dims + [(0, 0)]),
mode='constant')
pad_shape = im.shape[:-1]
else:
pad_shape = shape
# check that patch_shape or im_shape are divisible by 2**n_levels
if crop_shape is not None:
crop_shape = utils.reformat_to_list(crop_shape,
length=n_dims,
dtype='int')
if not all(
[pad_shape[i] >= crop_shape[i] for i in range(len(pad_shape))]):
crop_shape = [
min(pad_shape[i], crop_shape[i]) for i in range(n_dims)
]
print(
'cropping dimensions are higher than image size, changing cropping size to {}'
.format(crop_shape))
if not all([size % (2**n_levels) == 0 for size in crop_shape]):
crop_shape = [
utils.find_closest_number_divisible_by_m(size, 2**n_levels)
for size in crop_shape
]
else:
if not all([size % (2**n_levels) == 0 for size in pad_shape]):
crop_shape = [
utils.find_closest_number_divisible_by_m(size, 2**n_levels)
for size in pad_shape
]
# crop image if necessary
if crop_shape is not None:
crop_idx = np.round(
(pad_shape - np.array(crop_shape)) / 2).astype('int')
crop_idx = np.concatenate((crop_idx, crop_idx + crop_shape), axis=0)
im = edit_volumes.crop_volume_with_idx(im, crop_idx=crop_idx)
else:
crop_idx = None
# align image
# ref_axes = np.array([0, 2, 1])
# ref_signs = np.array([-1, 1, -1])
# im_axes, img_signs = utils.get_ras_axis_and_signs(aff, n_dims=n_dims)
# im = edit_volume.align_volume_to_ref(im, ref_axes, ref_signs, im_axes, img_signs)
# normalise image
m = np.min(im)
M = np.max(im)
if M == m:
im = np.zeros(im.shape)
else:
im = (im - m) / (M - m)
# add batch and channel axes
if n_channels > 1:
im = utils.add_axis(im)
else:
im = utils.add_axis(im, -2)
return im, aff, header, n_channels, n_dims, shape, pad_shape, crop_shape, crop_idx
def means_stds_fs_labels_with_relations(means_range, std_devs_range, min_diff=15, head=True):
# draw gm wm and csf means
gm_wm_csf_means = np.zeros(3)
while (abs(gm_wm_csf_means[1] - gm_wm_csf_means[0]) < min_diff) | \
(abs(gm_wm_csf_means[1] - gm_wm_csf_means[2]) < min_diff) | \
(abs(gm_wm_csf_means[0] - gm_wm_csf_means[2]) < min_diff):
gm_wm_csf_means = utils.draw_value_from_distribution(means_range, 3, 'uniform', 125., 100., positive_only=True)
gm_wm_csf_means = utils.add_axis(gm_wm_csf_means, -1)
# apply relations
wm = gm_wm_csf_means[0]
gm = gm_wm_csf_means[1]
csf = gm_wm_csf_means[2]
csf_like = csf * npr.uniform(low=0.95, high=1.05)
alpha_thalamus = npr.uniform(low=0.4, high=0.9)
thalamus = alpha_thalamus*gm + (1-alpha_thalamus)*wm
cerebellum_wm = wm * npr.uniform(low=0.7, high=1.3)
cerebellum_gm = gm * npr.uniform(low=0.7, high=1.3)
caudate = gm * npr.uniform(low=0.9, high=1.1)
putamen = gm * npr.uniform(low=0.9, high=1.1)
hippocampus = gm * npr.uniform(low=0.9, high=1.1)
amygdala = gm * npr.uniform(low=0.9, high=1.1)
accumbens = caudate * npr.uniform(low=0.9, high=1.1)
pallidum = wm * npr.uniform(low=0.8, high=1.2)
brainstem = wm * npr.uniform(low=0.8, high=1.2)
alpha_ventralDC = npr.uniform(low=0.1, high=0.6)
ventralDC = alpha_ventralDC*gm + (1-alpha_ventralDC)*wm
alpha_choroid = npr.uniform(low=0.0, high=1.0)
choroid = alpha_choroid*csf + (1-alpha_choroid)*wm
# regroup structures
neutral_means = [np.zeros(1), csf_like, csf_like, brainstem, csf]
sided_means = [wm, gm, csf_like, csf_like, cerebellum_wm, cerebellum_gm, thalamus, caudate, putamen, pallidum,
hippocampus, amygdala, accumbens, ventralDC, choroid]
# draw std deviations
std = utils.draw_value_from_distribution(std_devs_range, 17, 'uniform', 15., 10., positive_only=True)
std = utils.add_axis(std, -1)
neutral_stds = [np.zeros(1), std[1], std[1], std[2], std[3]]
sided_stds = [std[4], std[5], std[1], std[1], std[6], std[7], std[8], std[9], std[10], std[11], std[12], std[13],
std[14], std[15], std[16]]
# add means and variances for extra head labels if necessary
if head:
# means
extra_means = utils.draw_value_from_distribution(means_range, 2, 'uniform', 125., 100., positive_only=True)
extra_means = utils.add_axis(extra_means, -1)
skull = extra_means[0]
soft_non_brain = extra_means[1]
eye = csf * npr.uniform(low=0.95, high=1.05)
optic_chiasm = wm * npr.uniform(low=0.8, high=1.2)
vessel = csf * npr.uniform(low=0.7, high=1.3)
neutral_means += [csf_like, optic_chiasm, skull, soft_non_brain, eye]
sided_means.insert(-1, vessel)
# std dev
extra_std = utils.draw_value_from_distribution(std_devs_range, 4, 'uniform', 15., 10., positive_only=True)
extra_std = utils.add_axis(extra_std, -1)
neutral_stds += [std[1], extra_std[0], extra_std[1], extra_std[2], std[1]]
sided_stds.insert(-1, extra_std[3])
means = np.concatenate([np.array(neutral_means), np.array(sided_means), np.array(sided_means)])
stds = np.concatenate([np.array(neutral_stds), np.array(sided_stds), np.array(sided_stds)])
return means, stds
def build_model_input_generator(images_paths,
labels_paths,
n_channels,
im_shape,
scaling_range=None,
rotation_range=None,
shearing_range=None,
nonlin_shape_fact=0.0625,
nonlin_std_dev=3,
batch_size=1):
# Generate!
while True:
# randomly pick as many images as batch_size
indices = npr.randint(len(images_paths), size=batch_size)
# initialise input tensors
images_all = []
labels_all = []
aff_all = []
nonlinear_field_all = []
for idx in indices:
# add image
image = utils.load_volume(images_paths[idx])
if n_channels > 1:
images_all.append(utils.add_axis(image, axis=0))
else:
images_all.append(utils.add_axis(image, axis=-2))
# add labels
labels = utils.load_volume(labels_paths[idx], dtype='int')
labels_all.append(utils.add_axis(labels, axis=-2))
# get affine transformation: rotate, scale, shear (translation done during random cropping)
n_dims, _ = utils.get_dims(im_shape)
scaling = utils.draw_value_from_distribution(scaling_range, size=n_dims, centre=1, default_range=.15)
if n_dims == 2:
rotation_angle = utils.draw_value_from_distribution(rotation_range, default_range=15.0)
else:
rotation_angle = utils.draw_value_from_distribution(rotation_range, size=n_dims, default_range=15.0)
shearing = utils.draw_value_from_distribution(shearing_range, size=n_dims ** 2 - n_dims, default_range=.01)
aff = utils.create_affine_transformation_matrix(n_dims, scaling, rotation_angle, shearing)
aff_all.append(utils.add_axis(aff))
# add non linear field
deform_shape = utils.get_resample_shape(im_shape, nonlin_shape_fact, len(im_shape))
nonlinear_field = npr.normal(loc=0, scale=nonlin_std_dev * npr.rand(), size=deform_shape)
nonlinear_field_all.append(utils.add_axis(nonlinear_field))
# build list of inputs of the augmentation model
inputs_vals = [images_all, labels_all, aff_all, nonlinear_field_all]
# put images and labels (concatenated if batch_size>1) into a tuple of 2 elements: (cat_images, cat_labels)
if batch_size > 1:
inputs_vals = [np.concatenate(item, 0) for item in inputs_vals]
else:
inputs_vals = [item[0] for item in inputs_vals]
yield inputs_vals
def preprocess_image(im_path, n_levels, crop_shape=None, padding=None):
# read image and corresponding info
im, shape, aff, n_dims, n_channels, header, im_res = utils.get_volume_info(
im_path, True, np.eye(4))
if padding:
im = edit_volumes.pad_volume(im, padding_shape=padding)
pad_shape = im.shape[:n_dims]
else:
pad_shape = shape
# check that patch_shape or im_shape are divisible by 2**n_levels
if crop_shape is not None:
crop_shape = utils.reformat_to_list(crop_shape,
length=n_dims,
dtype='int')
if not all(
[pad_shape[i] >= crop_shape[i] for i in range(len(pad_shape))]):
crop_shape = [
min(pad_shape[i], crop_shape[i]) for i in range(n_dims)
]
if not all([size % (2**n_levels) == 0 for size in crop_shape]):
crop_shape = [
utils.find_closest_number_divisible_by_m(size, 2**n_levels)
for size in crop_shape
]
else:
if not all([size % (2**n_levels) == 0 for size in pad_shape]):
crop_shape = [
utils.find_closest_number_divisible_by_m(size, 2**n_levels)
for size in pad_shape
]
# crop image if necessary
if crop_shape is not None:
im, crop_idx = edit_volumes.crop_volume(im,
cropping_shape=crop_shape,
return_crop_idx=True)
else:
crop_idx = None
# normalise image
if n_channels == 1:
m = np.min(im)
M = np.max(im)
if M == m:
im = np.zeros(im.shape)
else:
im = (im - m) / (M - m)
else:
for i in range(im.shape[-1]):
channel = im[..., i]
m = np.min(channel)
M = np.max(channel)
if M == m:
im[..., i] = np.zeros(channel.shape)
else:
im[..., i] = (channel - m) / (M - m)
# add batch and channel axes
im = utils.add_axis(im) if n_channels > 1 else utils.add_axis(im,
axis=[0, -1])
return im, aff, header, im_res, n_channels, n_dims, shape, pad_shape, crop_idx
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 preprocess_image(im_path,
n_levels,
target_res,
crop=None,
padding=None,
flip=False,
path_resample=None):
# read image and corresponding info
im, _, aff, n_dims, n_channels, header, im_res = utils.get_volume_info(
im_path, True)
# resample image if necessary
if target_res is not None:
target_res = np.squeeze(
utils.reformat_to_n_channels_array(target_res, n_dims))
if np.any((im_res > target_res + 0.05) | (im_res < target_res - 0.05)):
im_res = target_res
im, aff = edit_volumes.resample_volume(im, aff, im_res)
if path_resample is not None:
utils.save_volume(im, aff, header, path_resample)
# align image
im = edit_volumes.align_volume_to_ref(im,
aff,
aff_ref=np.eye(4),
n_dims=n_dims)
shape = list(im.shape)
# pad image if specified
if padding:
im = edit_volumes.pad_volume(im, padding_shape=padding)
pad_shape = im.shape[:n_dims]
else:
pad_shape = shape
# check that patch_shape or im_shape are divisible by 2**n_levels
if crop is not None:
crop = utils.reformat_to_list(crop, length=n_dims, dtype='int')
if not all([pad_shape[i] >= crop[i] for i in range(len(pad_shape))]):
crop = [min(pad_shape[i], crop[i]) for i in range(n_dims)]
if not all([size % (2**n_levels) == 0 for size in crop]):
crop = [
utils.find_closest_number_divisible_by_m(size, 2**n_levels)
for size in crop
]
else:
if not all([size % (2**n_levels) == 0 for size in pad_shape]):
crop = [
utils.find_closest_number_divisible_by_m(size, 2**n_levels)
for size in pad_shape
]
# crop image if necessary
if crop is not None:
im, crop_idx = edit_volumes.crop_volume(im,
cropping_shape=crop,
return_crop_idx=True)
else:
crop_idx = None
# normalise image
if n_channels == 1:
im = edit_volumes.rescale_volume(im,
new_min=0.,
new_max=1.,
min_percentile=0.5,
max_percentile=99.5)
else:
for i in range(im.shape[-1]):
im[..., i] = edit_volumes.rescale_volume(im[..., i],
new_min=0.,
new_max=1.,
min_percentile=0.5,
max_percentile=99.5)
# flip image along right/left axis
if flip & (n_dims > 2):
im_flipped = edit_volumes.flip_volume(im,
direction='rl',
aff=np.eye(4))
im_flipped = utils.add_axis(
im_flipped) if n_channels > 1 else utils.add_axis(im_flipped,
axis=[0, -1])
else:
im_flipped = None
# add batch and channel axes
im = utils.add_axis(im) if n_channels > 1 else utils.add_axis(im,
axis=[0, -1])
return im, aff, header, im_res, n_channels, n_dims, shape, pad_shape, crop_idx, im_flipped
def predict(path_images,
path_segmentations,
path_model,
segmentation_labels,
n_neutral_labels=None,
path_posteriors=None,
path_resampled=None,
path_volumes=None,
segmentation_label_names=None,
padding=None,
cropping=None,
target_res=1.,
gradients=False,
flip=True,
topology_classes=None,
sigma_smoothing=0.5,
keep_biggest_component=True,
conv_size=3,
n_levels=5,
nb_conv_per_level=2,
unet_feat_count=24,
feat_multiplier=2,
activation='elu',
gt_folder=None,
evaluation_labels=None,
mask_folder=None,
list_incorrect_labels=None,
list_correct_labels=None,
compute_distances=False,
recompute=True,
verbose=True):
"""
This function uses trained models to segment images.
It is crucial that the inputs match the architecture parameters of the trained model.
:param path_images: path of the images to segment. Can be the path to a directory or the path to a single image.
:param path_segmentations: path where segmentations will be writen.
Should be a dir, if path_images is a dir, and a file if path_images is a file.
:param path_model: path ot the trained model.
:param segmentation_labels: List of labels for which to compute Dice scores. It should be the same list as the
segmentation_labels used in training.
:param n_neutral_labels: (optional) if the label maps contain some right/left specific labels and if test-time
flipping is applied (see parameter 'flip'), please provide the number of non-sided labels (including background).
It should be the same value as for training. Default is None.
:param path_posteriors: (optional) path where posteriors will be writen.
Should be a dir, if path_images is a dir, and a file if path_images is a file.
:param path_resampled: (optional) path where images resampled to 1mm isotropic will be writen.
We emphasise that images are resampled as soon as the resolution in one of the axes is not in the range [0.9; 1.1].
Should be a dir, if path_images is a dir, and a file if path_images is a file. Default is None, where resampled
images are not saved.
:param path_volumes: (optional) path of a csv file where the soft volumes of all segmented regions will be writen.
The rows of the csv file correspond to subjects, and the columns correspond to segmentation labels.
The soft volume of a structure corresponds to the sum of its predicted probability map.
:param segmentation_label_names: (optional) List of names correponding to the names of the segmentation labels.
Only used when path_volumes is provided. Must be of the same size as segmentation_labels. Can be given as a
list, a numpy array of strings, or the path to such a numpy array. Default is None.
:param padding: (optional) pad the images to the specified shape before predicting the segmentation maps.
Can be an int, a sequence or a 1d numpy array.
:param cropping: (optional) crop the images to the specified shape before predicting the segmentation maps.
If padding and cropping are specified, images are padded before being cropped.
Can be an int, a sequence or a 1d numpy array.
:param target_res: (optional) target resolution at which the network operates (and thus resolution of the output
segmentations). This must match the resolution of the training data ! target_res is used to automatically resampled
the images with resolutions outside [target_res-0.05, target_res+0.05].
Can be a sequence, a 1d numpy array. Set to None to disable the automatic resampling. Default is 1mm.
:param flip: (optional) whether to perform test-time augmentation, where the input image is segmented along with
a right/left flipped version on it. If set to True (default), be careful because this requires more memory.
:param topology_classes: List of classes corresponding to all segmentation labels, in order to group them into
classes, for each of which we will operate a smooth version of biggest connected component.
Can be a sequence, a 1d numpy array, or the path to a numpy 1d array in the same order as segmentation_labels.
Default is None, where no topological analysis is performed.
:param sigma_smoothing: (optional) If not None, the posteriors are smoothed with a gaussian kernel of the specified
standard deviation.
:param keep_biggest_component: (optional) whether to only keep the biggest component in the predicted segmentation.
This is applied independently of topology_classes, and it is applied to the whole segmentation
:param conv_size: (optional) size of unet's convolution masks. Default is 3.
:param n_levels: (optional) number of levels for unet. Default is 5.
:param nb_conv_per_level: (optional) number of convolution layers per level. Default is 2.
:param unet_feat_count: (optional) number of features for the first layer of the unet. Default is 24.
:param feat_multiplier: (optional) multiplicative factor for the number of feature for each new level. Default is 2.
:param activation: (optional) activation function. Can be 'elu', 'relu'.
:param gt_folder: (optional) path of the ground truth label maps corresponding to the input images. Should be a dir,
if path_images is a dir, or a file if path_images is a file.
Providing a gt_folder will trigger a Dice evaluation, where scores will be writen along with the path_segmentations.
Specifically, the scores are contained in a numpy array, where labels are in rows, and subjects in columns.
:param evaluation_labels: (optional) if gt_folder is True you can evaluate the Dice scores on a subset of the
segmentation labels, by providing another label list here. Can be a sequence, a 1d numpy array, or the path to a
numpy 1d array. Default is np.unique(segmentation_labels).
:param mask_folder: (optional) path of masks that will be used to mask out some parts of the obtained segmentations
during the evaluation. Default is None, where nothing is masked.
:param list_incorrect_labels: (optional) this option enables to replace some label values in the obtained
segmentations by other label values. Can be a list, a 1d numpy array, or the path to such an array.
: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 compute_distances: (optional) whether to add Hausdorff and mean surface distance evaluations to the default
Dice evaluation. Default is True.
:param recompute: (optional) whether to recompute segmentations that were already computed. This also applies to
Dice scores, if gt_folder is not None. Default is True.
:param verbose: (optional) whether to print out info about the remaining number of cases.
"""
# prepare input/output filepaths
path_images, path_segmentations, path_posteriors, path_resampled, path_volumes, compute = \
prepare_output_files(path_images, path_segmentations, path_posteriors, path_resampled, path_volumes, recompute)
# get label list
segmentation_labels, _ = utils.get_list_labels(
label_list=segmentation_labels)
n_labels = len(segmentation_labels)
# get unique label values, and build correspondance table between contralateral structures if necessary
if (n_neutral_labels is not None) & flip:
n_sided_labels = int((n_labels - n_neutral_labels) / 2)
lr_corresp = np.stack([
segmentation_labels[n_neutral_labels:n_neutral_labels +
n_sided_labels],
segmentation_labels[n_neutral_labels + n_sided_labels:]
])
segmentation_labels, indices = np.unique(segmentation_labels,
return_index=True)
lr_corresp_unique, lr_corresp_indices = np.unique(lr_corresp[0, :],
return_index=True)
lr_corresp_unique = np.stack(
[lr_corresp_unique, lr_corresp[1, lr_corresp_indices]])
lr_corresp_unique = lr_corresp_unique[:, 1:] if not np.all(
lr_corresp_unique[:, 0]) else lr_corresp_unique
lr_indices = np.zeros_like(lr_corresp_unique)
for i in range(lr_corresp_unique.shape[0]):
for j, lab in enumerate(lr_corresp_unique[i]):
lr_indices[i, j] = np.where(segmentation_labels == lab)[0]
else:
segmentation_labels, indices = np.unique(segmentation_labels,
return_index=True)
lr_indices = None
# prepare topology classes
if topology_classes is not None:
topology_classes = utils.load_array_if_path(
topology_classes, load_as_numpy=True)[indices]
# prepare volume file if needed
if path_volumes is not None:
if segmentation_label_names is not None:
segmentation_label_names = utils.load_array_if_path(
segmentation_label_names)[indices]
csv_header = [[''] + segmentation_label_names[1:].tolist()]
csv_header += [[''] +
[str(lab) for lab in segmentation_labels[1:]]]
else:
csv_header = [['subjects'] +
[str(lab) for lab in segmentation_labels[1:]]]
with open(path_volumes, 'w') as csvFile:
writer = csv.writer(csvFile)
writer.writerows(csv_header)
csvFile.close()
# build network
_, _, n_dims, n_channels, _, _ = utils.get_volume_info(path_images[0])
model_input_shape = [None] * n_dims + [n_channels]
net = build_model(path_model, model_input_shape, n_levels,
len(segmentation_labels), conv_size, nb_conv_per_level,
unet_feat_count, feat_multiplier, activation,
sigma_smoothing, gradients)
# perform segmentation
loop_info = utils.LoopInfo(len(path_images), 10, 'predicting', True)
for idx, (path_image, path_segmentation, path_posterior, path_resample, tmp_compute) in \
enumerate(zip(path_images, path_segmentations, path_posteriors, path_resampled, compute)):
# compute segmentation only if needed
if tmp_compute:
if verbose:
loop_info.update(idx)
# preprocessing
image, aff, h, im_res, _, _, shape, pad_shape, crop_idx, im_flipped = \
preprocess_image(path_image, n_levels, target_res, cropping, padding, flip, path_resample)
# prediction
prediction_patch = net.predict(image)
prediction_patch_flip = net.predict(im_flipped) if flip else None
# postprocessing
seg, posteriors = postprocess(
prediction_patch,
pad_shape,
shape,
crop_idx,
n_dims,
segmentation_labels,
lr_indices,
keep_biggest_component,
aff,
topology_classes=topology_classes,
post_patch_flip=prediction_patch_flip)
# write results to disk
if path_segmentation is not None:
utils.save_volume(seg,
aff,
h,
path_segmentation,
dtype='int32')
if path_posterior is not None:
if n_channels > 1:
posteriors = utils.add_axis(posteriors, axis=[0, -1])
utils.save_volume(posteriors,
aff,
h,
path_posterior,
dtype='float32')
else:
if path_volumes is not None:
posteriors, _, _, _, _, _, im_res = utils.get_volume_info(
path_posterior, True, aff_ref=np.eye(4))
else:
posteriors = im_res = None
# compute volumes
if path_volumes is not None:
volumes = np.sum(posteriors[..., 1:],
axis=tuple(range(0,
len(posteriors.shape) - 1)))
volumes = np.around(volumes * np.prod(im_res), 3)
row = [os.path.basename(path_image).replace('.nii.gz', '')
] + [str(vol) for vol in volumes]
with open(path_volumes, 'a') as csvFile:
writer = csv.writer(csvFile)
writer.writerow(row)
csvFile.close()
# evaluate
if gt_folder is not None:
# find path where segmentations are saved evaluation folder, and get labels on which to evaluate
eval_folder = os.path.dirname(path_segmentations[0])
if evaluation_labels is None:
evaluation_labels = segmentation_labels
# set path of result arrays for surface distance if necessary
if compute_distances:
path_hausdorff = os.path.join(eval_folder, 'hausdorff.npy')
path_hausdorff_99 = os.path.join(eval_folder, 'hausdorff_99.npy')
path_hausdorff_95 = os.path.join(eval_folder, 'hausdorff_95.npy')
path_mean_distance = os.path.join(eval_folder, 'mean_distance.npy')
else:
path_hausdorff = path_hausdorff_99 = path_hausdorff_95 = path_mean_distance = None
# compute evaluation metrics
evaluate.evaluation(gt_folder,
eval_folder,
evaluation_labels,
mask_dir=mask_folder,
path_dice=os.path.join(eval_folder, 'dice.npy'),
path_hausdorff=path_hausdorff,
path_hausdorff_99=path_hausdorff_99,
path_hausdorff_95=path_hausdorff_95,
path_mean_distance=path_mean_distance,
list_incorrect_labels=list_incorrect_labels,
list_correct_labels=list_correct_labels,
recompute=recompute,
verbose=verbose)
def build_model_inputs(path_label_maps,
n_labels,
batchsize=1,
n_channels=1,
generation_classes=None,
prior_distributions='uniform',
prior_means=None,
prior_stds=None,
use_specific_stats_for_channel=False,
mix_prior_and_random=False):
"""
This function builds a generator to be fed to the lab2im model. It enables to generate all the required inputs,
according to the operations performed in the model.
:param path_label_maps: list of the paths of the input label maps.
:param n_labels: number of labels in the input label maps.
: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 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 or 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.
: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, or 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.
"""
# get label info
_, _, n_dims, _, _, _ = utils.get_volume_info(path_label_maps[0])
# allocate unique class to each label if generation classes is not given
if generation_classes is None:
generation_classes = np.arange(n_labels)
# Generate!
while True:
# randomly pick as many images as batchsize
indices = npr.randint(len(path_label_maps), size=batchsize)
# initialise input lists
list_label_maps = []
list_means = []
list_stds = []
for idx in indices:
# add labels to inputs
lab = utils.load_volume(path_label_maps[idx], dtype='int', aff_ref=np.eye(4))
list_label_maps.append(utils.add_axis(lab, axis=[0, -1]))
# add means and standard deviations to inputs
means = np.empty((1, n_labels, 0))
stds = np.empty((1, n_labels, 0))
for channel in range(n_channels):
# retrieve channel specific stats if necessary
if isinstance(prior_means, np.ndarray):
if (prior_means.shape[0] > 2) & use_specific_stats_for_channel:
if prior_means.shape[0] / 2 != n_channels:
raise ValueError("the number of blocks in prior_means does not match n_channels. This "
"message is printed because use_specific_stats_for_channel is True.")
tmp_prior_means = prior_means[2 * channel:2 * channel + 2, :]
else:
tmp_prior_means = prior_means
else:
tmp_prior_means = prior_means
if (prior_means is not None) & mix_prior_and_random & (npr.uniform() > 0.5):
tmp_prior_means = None
if isinstance(prior_stds, np.ndarray):
if (prior_stds.shape[0] > 2) & use_specific_stats_for_channel:
if prior_stds.shape[0] / 2 != n_channels:
raise ValueError("the number of blocks in prior_stds does not match n_channels. This "
"message is printed because use_specific_stats_for_channel is True.")
tmp_prior_stds = prior_stds[2 * channel:2 * channel + 2, :]
else:
tmp_prior_stds = prior_stds
else:
tmp_prior_stds = prior_stds
if (prior_stds is not None) & mix_prior_and_random & (npr.uniform() > 0.5):
tmp_prior_stds = None
# draw means and std devs from priors
tmp_classes_means = utils.draw_value_from_distribution(tmp_prior_means, n_labels, prior_distributions,
125., 100., positive_only=True)
tmp_classes_stds = utils.draw_value_from_distribution(tmp_prior_stds, n_labels, prior_distributions,
15., 10., positive_only=True)
if npr.uniform() > 0.95: # reset the background to 0 in 10% of cases
tmp_classes_means[0] = 0
tmp_classes_stds[0] = 0
tmp_means = utils.add_axis(tmp_classes_means[generation_classes], axis=[0, -1])
tmp_stds = utils.add_axis(tmp_classes_stds[generation_classes], axis=[0, -1])
means = np.concatenate([means, tmp_means], axis=-1)
stds = np.concatenate([stds, tmp_stds], axis=-1)
list_means.append(means)
list_stds.append(stds)
# build list of inputs for generation model
list_inputs = [list_label_maps, list_means, list_stds]
if batchsize > 1: # concatenate each input type if batchsize > 1
list_inputs = [np.concatenate(item, 0) for item in list_inputs]
else:
list_inputs = [item[0] for item in list_inputs]
yield list_inputs
def preprocess_image(im_path, n_levels, crop_shape=None, padding=None, aff_ref='FS'):
# read image and corresponding info
im, shape, aff, n_dims, n_channels, header, im_res = utils.get_volume_info(im_path, return_volume=True)
if padding:
if n_channels == 1:
im = np.pad(im, padding, mode='constant')
pad_shape = im.shape
else:
im = np.pad(im, tuple([(padding, padding)] * n_dims + [(0, 0)]), mode='constant')
pad_shape = im.shape[:-1]
else:
pad_shape = shape
# check that patch_shape or im_shape are divisible by 2**n_levels
if crop_shape is not None:
crop_shape = utils.reformat_to_list(crop_shape, length=n_dims, dtype='int')
if not all([pad_shape[i] >= crop_shape[i] for i in range(len(pad_shape))]):
crop_shape = [min(pad_shape[i], crop_shape[i]) for i in range(n_dims)]
print('cropping dimensions are higher than image size, changing cropping size to {}'.format(crop_shape))
if not all([size % (2**n_levels) == 0 for size in crop_shape]):
crop_shape = [utils.find_closest_number_divisible_by_m(size, 2 ** n_levels) for size in crop_shape]
else:
if not all([size % (2**n_levels) == 0 for size in pad_shape]):
crop_shape = [utils.find_closest_number_divisible_by_m(size, 2 ** n_levels) for size in pad_shape]
# crop image if necessary
if crop_shape is not None:
crop_idx = np.round((pad_shape - np.array(crop_shape)) / 2).astype('int')
crop_idx = np.concatenate((crop_idx, crop_idx + crop_shape), axis=0)
im = edit_volumes.crop_volume_with_idx(im, crop_idx=crop_idx)
else:
crop_idx = None
# align image to training axes and directions
if n_dims > 2:
if aff_ref == 'FS':
aff_ref = np.array([[-1., 0., 0., 0.], [0., 0., 1., 0.], [0., -1., 0., 0.], [0., 0., 0., 1.]])
im = edit_volumes.align_volume_to_ref(im, aff, aff_ref=aff_ref, return_aff=False)
elif aff_ref == 'identity':
aff_ref = np.eye(4)
im = edit_volumes.align_volume_to_ref(im, aff, aff_ref=aff_ref, return_aff=False)
elif aff_ref == 'MS':
aff_ref = np.array([[-1., 0., 0., 0.], [0., -1., 0., 0.], [0., 0., 1., 0.], [0., 0., 0., 1.]])
im = edit_volumes.align_volume_to_ref(im, aff, aff_ref=aff_ref, return_aff=False)
# normalise image
if n_channels == 1:
m = np.min(im)
M = np.max(im)
if M == m:
im = np.zeros(im.shape)
else:
im = (im - m) / (M - m)
if n_channels > 1:
for i in range(im.shape[-1]):
channel = im[..., i]
m = np.min(channel)
M = np.max(channel)
if M == m:
im[..., i] = np.zeros(channel.shape)
else:
im[..., i] = (channel - m) / (M - m)
# add batch and channel axes
if n_channels > 1:
im = utils.add_axis(im)
else:
im = utils.add_axis(im, -2)
return im, aff, header, im_res, n_channels, n_dims, shape, pad_shape, crop_shape, crop_idx
