def post_process(img: structure.BrainImage, segmentation: sitk.Image,
probability: sitk.Image, **kwargs) -> sitk.Image:
"""Post-processes a segmentation.
Args:
img (structure.BrainImage): The image.
segmentation (sitk.Image): The segmentation (label image).
probability (sitk.Image): The probabilities images (a vector image).
Returns:
sitk.Image: The post-processed image.
"""
print('-' * 10, 'Post-processing', img.id_)
# construct pipeline
pipeline = fltr.FilterPipeline()
if kwargs.get('crf_post', False):
pipeline.add_filter(fltr_postp.DenseCRF())
pipeline.set_param(
fltr_postp.DenseCRFParams(img.images[structure.BrainImageTypes.T1],
img.images[structure.BrainImageTypes.T2],
probability), 0)
return pipeline.execute(segmentation)
def post_process(img: structure.BrainImage, segmentation: sitk.Image, probability: sitk.Image,
**kwargs) -> sitk.Image:
"""Post-processes a segmentation.
Args:
img (structure.BrainImage): The image.
segmentation (sitk.Image): The segmentation (label image).
probability (sitk.Image): The probabilities images (a vector image).
Returns:
sitk.Image: The post-processed image.
"""
print('-' * 10, 'Post-processing', img.id_)
# construct pipeline
pipeline = fltr.FilterPipeline()
if kwargs.get('simple_post', False):
pipeline.add_filter(fltr_postp.ImagePostProcessing())
pipeline.set_param(fltr_postp.ImagePostProcessingParam(probability,
variance=kwargs.get('variance', 1.0),
preserve_background=kwargs.get('preserve_background', False)), len(pipeline.filters) - 1)
if kwargs.get('crf_post', False):
pipeline.add_filter(fltr_postp.DenseCRF())
pipeline.set_param(fltr_postp.DenseCRFParams(img.images[structure.BrainImageTypes.T1w],
img.images[structure.BrainImageTypes.T2w],
probability), len(pipeline.filters) - 1)
return pipeline.execute(segmentation)
def main(data_dir: str):
# initialize filter pipeline with filters
filters = [
prep.GradientAnisotropicDiffusion(time_step=0.0625),
prep.HistogramMatcher()
]
histogram_matching_filter_idx = 1 # we need the index later to update the HistogramMatcher's parameters
pipeline = flt.FilterPipeline(filters)
# get subjects to evaluate
subject_dirs = [
subject for subject in glob.glob(os.path.join(data_dir, '*'))
if os.path.isdir(subject)
and os.path.basename(subject).startswith('Subject')
]
for subject_dir in subject_dirs:
subject_id = os.path.basename(subject_dir)
print(f'Filtering {subject_id}...')
# load the T1- and T2-weighted MR images
t1_image = sitk.ReadImage(
os.path.join(subject_dir, f'{subject_id}_T1.mha'))
t2_image = sitk.ReadImage(
os.path.join(subject_dir, f'{subject_id}_T2.mha'))
# set the T2-weighted MR image as reference for the histogram matching
pipeline.set_param(prep.HistogramMatcherParams(t2_image),
histogram_matching_filter_idx)
# execute filtering pipeline on the T1-weighted image
filtered_t1_image = pipeline.execute(t1_image)
# plot filtering result
slice_no_for_plot = t1_image.GetSize()[2] // 2
fig, axs = plt.subplots(1, 2)
axs[0].imshow(sitk.GetArrayFromImage(t1_image[:, :,
slice_no_for_plot]),
cmap='gray')
axs[0].set_title('Original image')
axs[1].imshow(sitk.GetArrayFromImage(
filtered_t1_image[:, :, slice_no_for_plot]),
cmap='gray')
axs[1].set_title('Filtered image')
fig.suptitle(f'{subject_id}', fontsize=16)
plt.show()
def pre_process(id_: str, paths: dict, **kwargs) -> structure.BrainImage:
"""Loads and processes an image.
The processing includes:
- Registration
- Pre-processing
- Feature extraction
Args:
id_ (str): An image identifier.
paths (dict): A dict, where the keys are an image identifier of type structure.BrainImageTypes
and the values are paths to the images.
Returns:
(structure.BrainImage):
"""
print('-' * 10, 'Processing', id_)
# load image
path = paths.pop(
id_, '') # the value with key id_ is the root directory of the image
img = {img_key: sitk.ReadImage(path) for img_key, path in paths.items()}
img = structure.BrainImage(id_, path, img)
# construct T1 pipeline
pipeline_t1 = fltr.FilterPipeline()
if kwargs.get('zscore_pre', False):
pipeline_t1.add_filter(fltr_prep.NormalizeZScore())
if kwargs.get('registration_pre', False):
pipeline_t1.add_filter(fltr_reg.MultiModalRegistration())
pipeline_t1.set_param(fltr_reg.MultiModalRegistrationParams(atlas_t1),
1)
# execute pipeline on T1 image
img.images[structure.BrainImageTypes.T1] = pipeline_t1.execute(
img.images[structure.BrainImageTypes.T1])
# construct T2 pipeline
pipeline_t2 = fltr.FilterPipeline()
if kwargs.get('zscore_pre', False):
pipeline_t2.add_filter(fltr_prep.NormalizeZScore())
# execute pipeline on T2 image
img.images[structure.BrainImageTypes.T2] = pipeline_t2.execute(
img.images[structure.BrainImageTypes.T2])
if kwargs.get('registration_pre', False):
# get transformation
transform = pipeline_t1.filters[1].transform
# apply transformation of T1 image registration to T2 image
image_t2 = img.images[structure.BrainImageTypes.T2]
image_t2 = sitk.Resample(image_t2, atlas_t1, transform,
sitk.sitkLinear, 0.0,
image_t2.GetPixelIDValue())
img.images[structure.BrainImageTypes.T2] = image_t2
# apply transformation of T1 image registration to ground truth
image_ground_truth = img.images[structure.BrainImageTypes.GroundTruth]
image_ground_truth = sitk.Resample(
image_ground_truth, atlas_t1, transform, sitk.sitkNearestNeighbor,
0, image_ground_truth.GetPixelIDValue())
img.images[structure.BrainImageTypes.GroundTruth] = image_ground_truth
# update image properties to atlas image properties after registration
img.image_properties = conversion.ImageProperties(atlas_t1)
# extract the features
feature_extractor = FeatureExtractor(img, **kwargs)
img = feature_extractor.execute()
img.feature_images = {}
return img
def pre_process(id_: str, paths: dict, **kwargs) -> structure.BrainImage:
"""Loads and processes an image.
The processing includes:
- Registration
- Pre-processing
- Feature extraction
Args:
id_ (str): An image identifier.
paths (dict): A dict, where the keys are an image identifier of type structure.BrainImageTypes
and the values are paths to the images.
Returns:
(structure.BrainImage):
"""
print('-' * 10, 'Processing', id_)
# load image
path = paths.pop(
id_, '') # the value with key id_ is the root directory of the image
path_to_transform = paths.pop(
structure.BrainImageTypes.RegistrationTransform, '')
img = {img_key: sitk.ReadImage(path) for img_key, path in paths.items()}
transform = sitk.ReadTransform(path_to_transform)
img = structure.BrainImage(id_, path, img, transform)
# construct pipeline for brain mask registration
# we need to perform this before the T1w and T2w pipeline because the registered mask is used for skull-stripping
pipeline_brain_mask = fltr.FilterPipeline()
if kwargs.get('registration_pre', False):
pipeline_brain_mask.add_filter(fltr_prep.ImageRegistration())
pipeline_brain_mask.set_param(
fltr_prep.ImageRegistrationParameters(atlas_t1, img.transformation,
True),
len(pipeline_brain_mask.filters) - 1)
# execute pipeline on the brain mask image
img.images[
structure.BrainImageTypes.BrainMask] = pipeline_brain_mask.execute(
img.images[structure.BrainImageTypes.BrainMask])
# construct pipeline for T1w image pre-processing
pipeline_t1 = fltr.FilterPipeline()
if kwargs.get('registration_pre', False):
pipeline_t1.add_filter(fltr_prep.ImageRegistration())
pipeline_t1.set_param(
fltr_prep.ImageRegistrationParameters(atlas_t1,
img.transformation),
len(pipeline_t1.filters) - 1)
if kwargs.get('skullstrip_pre', False):
pipeline_t1.add_filter(fltr_prep.SkullStripping())
pipeline_t1.set_param(
fltr_prep.SkullStrippingParameters(
img.images[structure.BrainImageTypes.BrainMask]),
len(pipeline_t1.filters) - 1)
if kwargs.get('normalization_pre', False):
pipeline_t1.add_filter(fltr_prep.ImageNormalization())
# execute pipeline on the T1w image
img.images[structure.BrainImageTypes.T1w] = pipeline_t1.execute(
img.images[structure.BrainImageTypes.T1w])
# construct pipeline for T2w image pre-processing
pipeline_t2 = fltr.FilterPipeline()
if kwargs.get('registration_pre', False):
pipeline_t2.add_filter(fltr_prep.ImageRegistration())
pipeline_t2.set_param(
fltr_prep.ImageRegistrationParameters(atlas_t2,
img.transformation),
len(pipeline_t2.filters) - 1)
if kwargs.get('skullstrip_pre', False):
pipeline_t2.add_filter(fltr_prep.SkullStripping())
pipeline_t2.set_param(
fltr_prep.SkullStrippingParameters(
img.images[structure.BrainImageTypes.BrainMask]),
len(pipeline_t2.filters) - 1)
if kwargs.get('normalization_pre', False):
pipeline_t2.add_filter(fltr_prep.ImageNormalization())
# execute pipeline on the T2w image
img.images[structure.BrainImageTypes.T2w] = pipeline_t2.execute(
img.images[structure.BrainImageTypes.T2w])
# construct pipeline for ground truth image pre-processing
pipeline_gt = fltr.FilterPipeline()
if kwargs.get('registration_pre', False):
pipeline_gt.add_filter(fltr_prep.ImageRegistration())
pipeline_gt.set_param(
fltr_prep.ImageRegistrationParameters(atlas_t1, img.transformation,
True),
len(pipeline_gt.filters) - 1)
# execute pipeline on the ground truth image
img.images[structure.BrainImageTypes.GroundTruth] = pipeline_gt.execute(
img.images[structure.BrainImageTypes.GroundTruth])
# update image properties to atlas image properties after registration
img.image_properties = conversion.ImageProperties(
img.images[structure.BrainImageTypes.T1w])
# extract the features
feature_extractor = FeatureExtractor(img, **kwargs)
img = feature_extractor.execute()
img.feature_images = {
} # we free up memory because we only need the img.feature_matrix
# for training of the classifier
return img
