def _get_initialization_generator(data_dir):
"""
Creates the pre-processing pipeline used during initialization of the other pre-processing pipelines.
Parameters
----------
data_dir: string
The directory which contains the respective data.
Returns
-------
data_pipeline: iterable
An iterable over a number of datapoints where each datapoint is a tuple of a list of inputs and a parameter dictionary and
the first input in the list is the network input, while the second input is the corresponding ground truth.
"""
global input_identifier
global ground_truth_identifier
with dpp.Pipeline() as init:
node = dpp.reader.file_paths(
data_dir,
input_identifier=input_identifier,
ground_truth_identifier=ground_truth_identifier,
random=False,
iterations=1)
node = vol.load_volume(node)
node = seg.mask_img_background(node, 0, pixel_value=0.)
node = seg.clip_img_values(node, 0., 200.)
return init
def generate_predictions(test_data_dir, save_name=""):
"""
Creates the pre-processing pipeline used when generating predictions.
Parameters
----------
test_data_dir: string
The directory which contains the test data.
save_name: string
The name used to save temporary values needed for pre-processing and specific to a given dataset.
These can be re-loaded, which greatly speeds up initialization.
Returns
-------
test_pipeline: iterable
An iterable over a number of datapoints where each datapoint is a tuple of a list of inputs and a parameter dictionary and
the first input in the list is the network input, while the second input is the corresponding ground truth.
"""
global input_identifier
global ground_truth_identifier
# Volumes for initialization #
init = _get_initialization_generator(test_data_dir)
with dpp.Pipeline(storage_name=save_name,
initialization_generator=init) as pipe:
node = dpp.reader.file_paths(
test_data_dir,
input_identifier=input_identifier,
ground_truth_identifier=ground_truth_identifier,
random=False,
iterations=1)
node = seg.medical.load_all_slices(node,
slice_type='axial',
depth=5,
single_label_slice=False)
# Adjust labels and colours
node = seg.crop_to_label(node,
1,
max_reduction=10,
square_crop=True,
default_label=0)
node = seg.mask_img_background(node, 0, pixel_value=0.)
node = seg.reduce_to_single_label_slice(node)
node = seg.clip_img_values(node, 0., 200.)
# Prepare as network input
node = seg.resize(node, [256, 256])
node = seg.transpose(node)
# node = seg.pad(node, 92, mode='constant', image_only=True, default_pixel=0.)
node = seg.robust_img_scaling(node,
ignore_values=[0., 200.],
initialize=True)
return pipe
def _lesions_initialization_generator(data_dir, input_identifier,
prediction_identifier):
with dpp.Pipeline() as init:
node = dpp.reader.file_paths(
data_dir,
input_identifier=input_identifier,
ground_truth_identifier=prediction_identifier,
random=False,
iterations=1)
node = vol.load_volume(node)
node = seg.mask_img_background(node, 0, pixel_value=0.)
node = seg.clip_img_values(node, 0., 200.)
return init
def _test_val_tail(node):
# Adjust labels and colours
# node = seg.crop_to_label(node, 1, max_reduction=10, square_crop=True, default_label=0)
# node = seg.resize(node, [312, 312])
# node = seg.decreasing_mask_img_background(node, 0, pixel_value=0.)
node = seg.mask_img_background(node, 0, pixel_value=0.)
node = seg.reduce_to_single_label_slice(node)
node = seg.clip_img_values(node, 0, 200.)
# node = seg.decreasing_clip_img_values(node, 0, 200.)
node = seg.fuse_labels_other_than(node, 2)
# Prepare as network input
# node = seg.resize(node, [256, 256])
node = seg.transpose(node)
# node = seg.robust_img_scaling(node, ignore_values=[0., 200.], initialize=True)
return node
def lesions(data_dir, input_identifier, prediction_identifier, save_name=""):
# Volumes for initialization #
init = _lesions_initialization_generator(data_dir, input_identifier,
prediction_identifier)
with dpp.Pipeline(storage_name=save_name,
initialization_generator=init) as pipe:
node = dpp.reader.file_paths(
data_dir,
input_identifier=input_identifier,
ground_truth_identifier=prediction_identifier,
random=False,
iterations=1)
node = seg.medical.load_all_slices(node,
slice_type='axial',
depth=5,
single_label_slice=False)
# Adjust labels and colours
node = seg.crop_to_label(node,
1,
max_reduction=10,
square_crop=True,
default_label=0)
node = seg.mask_img_background(node, 0, pixel_value=0.)
node = seg.reduce_to_single_label_slice(node)
node = seg.clip_img_values(node, 0., 200.)
# Prepare as network input
node = seg.resize(node, [256, 256])
node = seg.transpose(node)
node = seg.robust_img_scaling(node,
ignore_values=[0., 200.],
initialize=True)
return pipe
def training(train_data_dir, save_name=""):
"""
Creates the pre-processing pipeline used during training.
Parameters
----------
train_data_dir: string
The directory which contains the training and validation data.
The data is split into a training and validation set using the global seed.
save_name: string
The name used to save temporary values needed for pre-processing and specific to a given dataset.
These can be re-loaded, which greatly speeds up initialization.
Returns
-------
training_pipeline: iterable
An iterable over a number of datapoints where each datapoint is a tuple of a list of inputs and a parameter dictionary and
the first input in the list is the network input, while the second input is the corresponding ground truth.
"""
global input_identifier
global ground_truth_identifier
global split
global seed
# Volumes for initialization #
init = _get_initialization_generator(train_data_dir)
# Slices for training #
with dpp.Pipeline(storage_name=save_name,
initialization_generator=init) as pipe:
# Load slices
node = dpp.reader.file_paths(
train_data_dir,
input_identifier=input_identifier,
ground_truth_identifier=ground_truth_identifier,
random=True,
iterations=0,
split=split,
split_selector=0,
seed=seed)
node = seg.medical.load_slice_filtered(node,
label_of_interest=2,
label_required=1,
min_frequency=0.8,
max_tries=20,
slice_type='axial',
depth=5,
single_label_slice=False)
# Random rotation then crop to fit
node = seg.random_rotation(node, probability=1.0, upper_bound=180)
node = seg.crop_to_label(node,
1,
max_reduction=10,
square_crop=True,
default_label=0)
# Random transformations
node = seg.random_resize(node, [256, 256],
probability=1.0,
lower_bound=0.8,
upper_bound=1.1,
default_pixel=0.,
default_label=0)
node = seg.random_translation(node,
probability=0.5,
border_usage=0.5,
default_border=0.25,
label_of_interest=1,
default_pixel=0.,
default_label=0)
# Adjust colours and labels
node = seg.mask_img_background(node, 0, pixel_value=0.)
node = seg.reduce_to_single_label_slice(node)
node = seg.clip_img_values(node, 0., 200.)
node = seg.fuse_labels_other_than(node, 2)
# Prepare as input
node = seg.transpose(node)
node = seg.robust_img_scaling(node,
ignore_values=[0., 200.],
initialize=True)
return pipe