def regression_match_image(source_image,
reference_image,
poly_order=1,
truncate=True):
"""
Image intensity normalization using linear regression.
Arguments
---------
source_image : ANTsImage
Image whose intensities are matched to the reference
image.
reference_image : ANTsImage
Defines the reference intensity function.
poly_order : integer
Polynomial order of fit. Default is 1 (linear fit).
truncate : boolean
Turns on/off the clipping of intensities.
Returns
-------
ANTs image (i.e., source_image) matched to the (reference_image).
Example
-------
>>> import ants
>>> source_image = ants.image_read(ants.get_ants_data('r16'))
>>> reference_image = ants.image_read(ants.get_ants_data('r64'))
>>> matched_image = regression_match_image(source_image, reference_image)
"""
if source_image.shape != reference_image.shape:
raise ValueError("Images do not have the same dimension.")
source_intensities = np.expand_dims((source_image.numpy()).flatten(), axis=1)
reference_intensities = np.expand_dims((reference_image.numpy()).flatten(), axis=1)
poly_features = PolynomialFeatures(degree=poly_order)
source_intensities_poly = poly_features.fit_transform(source_intensities)
model = LinearRegression()
model.fit(source_intensities_poly, reference_intensities)
matched_source_intensities = model.predict(source_intensities_poly)
if truncate == True:
min_reference_value = reference_intensities.min()
max_reference_value = reference_intensities.max()
matched_source_intensities[matched_source_intensities < min_reference_value] = min_reference_value
matched_source_intensities[matched_source_intensities > max_reference_value] = max_reference_value
matched_source_image = ants.make_image(source_image.shape, matched_source_intensities)
matched_source_image = ants.copy_image_info(source_image, matched_source_image)
return(matched_source_image)
def test_ndimage_to_list(self):
image = ants.image_read(ants.get_ants_data('r16'))
image2 = ants.image_read(ants.get_ants_data('r64'))
ants.set_spacing(image, (2, 2))
ants.set_spacing(image2, (2, 2))
imageTar = ants.make_image((*image2.shape, 2))
ants.set_spacing(imageTar, (2, 2, 2))
image3 = ants.list_to_ndimage(imageTar, [image, image2])
self.assertEqual(image3.dimension, 3)
ants.set_direction(image3, np.eye(3) * 2)
images_unmerged = ants.ndimage_to_list(image3)
self.assertEqual(len(images_unmerged), 2)
self.assertEqual(images_unmerged[0].dimension, 2)
def test_make_image(self):
self.setUp()
for arr in self.arrs:
voxval = 6.
img = ants.make_image(arr.shape, voxval=voxval)
self.assertTrue(img.dimension, arr.ndim)
self.assertTrue(img.shape, arr.shape)
nptest.assert_allclose(img.mean(), voxval)
new_origin = tuple([6.9] * arr.ndim)
new_spacing = tuple([3.6] * arr.ndim)
new_direction = np.eye(arr.ndim) * 9.6
img2 = ants.make_image(arr.shape,
voxval=voxval,
origin=new_origin,
spacing=new_spacing,
direction=new_direction)
self.assertTrue(img2.dimension, arr.ndim)
self.assertTrue(img2.shape, arr.shape)
nptest.assert_allclose(img2.mean(), voxval)
self.assertEqual(img2.origin, new_origin)
self.assertEqual(img2.spacing, new_spacing)
nptest.assert_allclose(img2.direction, new_direction)
for ptype in self.pixeltypes:
img = ants.make_image(arr.shape, voxval=1., pixeltype=ptype)
self.assertEqual(img.pixeltype, ptype)
# test with components
img = ants.make_image((69, 70, 4), has_components=True)
self.assertEqual(img.components, 4)
self.assertEqual(img.dimension, 2)
nptest.assert_allclose(img.mean(), 0.)
img = ants.make_image((69, 70, 71, 4), has_components=True)
self.assertEqual(img.components, 4)
self.assertEqual(img.dimension, 3)
nptest.assert_allclose(img.mean(), 0.)
# set from image
for img in self.imgs:
mask = img > img.mean()
arr = img[mask]
img2 = ants.make_image(mask, voxval=arr)
nptest.assert_allclose(img2.numpy(), (img * mask).numpy())
self.assertTrue(ants.image_physical_space_consistency(img2, mask))
# set with arr.ndim > 1
img2 = ants.make_image(mask, voxval=np.expand_dims(arr, -1))
nptest.assert_allclose(img2.numpy(), (img * mask).numpy())
self.assertTrue(ants.image_physical_space_consistency(img2, mask))
def test_transform_index_to_physical_point(self):
img = ants.make_image((10, 10), np.random.randn(100))
pt = ants.transform_index_to_physical_point(img, (2, 2))
# image not ANTsImage
with self.assertRaises(Exception):
ants.transform_index_to_physical_point(2, (2, 2))
# index not tuple/list
with self.assertRaises(Exception):
ants.transform_index_to_physical_point(img, img)
# index not same size as dimension
with self.assertRaises(Exception):
ants.transform_index_to_physical_point(img,
[2] * (img.dimension + 1))
def lung_extraction(image,
modality="proton",
antsxnet_cache_directory=None,
verbose=False):
"""
Perform proton or ct lung extraction using U-net.
Arguments
---------
image : ANTsImage
input image
modality : string
Modality image type. Options include "ct", "proton", "protonLobes",
"maskLobes", and "ventilation".
antsxnet_cache_directory : string
Destination directory for storing the downloaded template and model weights.
Since these can be resused, if is None, these data will be downloaded to a
~/.keras/ANTsXNet/.
verbose : boolean
Print progress to the screen.
Returns
-------
Dictionary of ANTs segmentation and probability images.
Example
-------
>>> output = lung_extraction(lung_image, modality="proton")
"""
from ..architectures import create_unet_model_2d
from ..architectures import create_unet_model_3d
from ..utilities import get_pretrained_network
from ..utilities import get_antsxnet_data
from ..utilities import pad_or_crop_image_to_size
if image.dimension != 3:
raise ValueError( "Image dimension must be 3." )
if antsxnet_cache_directory == None:
antsxnet_cache_directory = "ANTsXNet"
image_mods = [modality]
channel_size = len(image_mods)
weights_file_name = None
unet_model = None
if modality == "proton":
weights_file_name = get_pretrained_network("protonLungMri",
antsxnet_cache_directory=antsxnet_cache_directory)
classes = ("background", "left_lung", "right_lung")
number_of_classification_labels = len(classes)
reorient_template_file_name_path = get_antsxnet_data("protonLungTemplate",
antsxnet_cache_directory=antsxnet_cache_directory)
reorient_template = ants.image_read(reorient_template_file_name_path)
resampled_image_size = reorient_template.shape
unet_model = create_unet_model_3d((*resampled_image_size, channel_size),
number_of_outputs=number_of_classification_labels,
number_of_layers=4, number_of_filters_at_base_layer=16, dropout_rate=0.0,
convolution_kernel_size=(7, 7, 5), deconvolution_kernel_size=(7, 7, 5))
unet_model.load_weights(weights_file_name)
if verbose == True:
print("Lung extraction: normalizing image to the template.")
center_of_mass_template = ants.get_center_of_mass(reorient_template * 0 + 1)
center_of_mass_image = ants.get_center_of_mass(image * 0 + 1)
translation = np.asarray(center_of_mass_image) - np.asarray(center_of_mass_template)
xfrm = ants.create_ants_transform(transform_type="Euler3DTransform",
center=np.asarray(center_of_mass_template), translation=translation)
warped_image = ants.apply_ants_transform_to_image(xfrm, image, reorient_template)
batchX = np.expand_dims(warped_image.numpy(), axis=0)
batchX = np.expand_dims(batchX, axis=-1)
batchX = (batchX - batchX.mean()) / batchX.std()
predicted_data = unet_model.predict(batchX, verbose=int(verbose))
origin = warped_image.origin
spacing = warped_image.spacing
direction = warped_image.direction
probability_images_array = list()
for i in range(number_of_classification_labels):
probability_images_array.append(
ants.from_numpy(np.squeeze(predicted_data[0, :, :, :, i]),
origin=origin, spacing=spacing, direction=direction))
if verbose == True:
print("Lung extraction: renormalize probability mask to native space.")
for i in range(number_of_classification_labels):
probability_images_array[i] = ants.apply_ants_transform_to_image(
ants.invert_ants_transform(xfrm), probability_images_array[i], image)
image_matrix = ants.image_list_to_matrix(probability_images_array, image * 0 + 1)
segmentation_matrix = np.argmax(image_matrix, axis=0)
segmentation_image = ants.matrix_to_images(
np.expand_dims(segmentation_matrix, axis=0), image * 0 + 1)[0]
return_dict = {'segmentation_image' : segmentation_image,
'probability_images' : probability_images_array}
return(return_dict)
if modality == "protonLobes" or modality == "maskLobes":
reorient_template_file_name_path = get_antsxnet_data("protonLungTemplate",
antsxnet_cache_directory=antsxnet_cache_directory)
reorient_template = ants.image_read(reorient_template_file_name_path)
resampled_image_size = reorient_template.shape
spatial_priors_file_name_path = get_antsxnet_data("protonLobePriors",
antsxnet_cache_directory=antsxnet_cache_directory)
spatial_priors = ants.image_read(spatial_priors_file_name_path)
priors_image_list = ants.ndimage_to_list(spatial_priors)
channel_size = 1 + len(priors_image_list)
number_of_classification_labels = 1 + len(priors_image_list)
unet_model = create_unet_model_3d((*resampled_image_size, channel_size),
number_of_outputs=number_of_classification_labels, mode="classification",
number_of_filters_at_base_layer=16, number_of_layers=4,
convolution_kernel_size=(3, 3, 3), deconvolution_kernel_size=(2, 2, 2),
dropout_rate=0.0, weight_decay=0, additional_options=("attentionGating",))
if modality == "protonLobes":
penultimate_layer = unet_model.layers[-2].output
outputs2 = Conv3D(filters=1,
kernel_size=(1, 1, 1),
activation='sigmoid',
kernel_regularizer=regularizers.l2(0.0))(penultimate_layer)
unet_model = Model(inputs=unet_model.input, outputs=[unet_model.output, outputs2])
weights_file_name = get_pretrained_network("protonLobes",
antsxnet_cache_directory=antsxnet_cache_directory)
else:
weights_file_name = get_pretrained_network("maskLobes",
antsxnet_cache_directory=antsxnet_cache_directory)
unet_model.load_weights(weights_file_name)
if verbose == True:
print("Lung extraction: normalizing image to the template.")
center_of_mass_template = ants.get_center_of_mass(reorient_template * 0 + 1)
center_of_mass_image = ants.get_center_of_mass(image * 0 + 1)
translation = np.asarray(center_of_mass_image) - np.asarray(center_of_mass_template)
xfrm = ants.create_ants_transform(transform_type="Euler3DTransform",
center=np.asarray(center_of_mass_template), translation=translation)
warped_image = ants.apply_ants_transform_to_image(xfrm, image, reorient_template)
warped_array = warped_image.numpy()
if modality == "protonLobes":
warped_array = (warped_array - warped_array.mean()) / warped_array.std()
else:
warped_array[warped_array != 0] = 1
batchX = np.zeros((1, *warped_array.shape, channel_size))
batchX[0,:,:,:,0] = warped_array
for i in range(len(priors_image_list)):
batchX[0,:,:,:,i+1] = priors_image_list[i].numpy()
predicted_data = unet_model.predict(batchX, verbose=int(verbose))
origin = warped_image.origin
spacing = warped_image.spacing
direction = warped_image.direction
probability_images_array = list()
for i in range(number_of_classification_labels):
if modality == "protonLobes":
probability_images_array.append(
ants.from_numpy(np.squeeze(predicted_data[0][0, :, :, :, i]),
origin=origin, spacing=spacing, direction=direction))
else:
probability_images_array.append(
ants.from_numpy(np.squeeze(predicted_data[0, :, :, :, i]),
origin=origin, spacing=spacing, direction=direction))
if verbose == True:
print("Lung extraction: renormalize probability images to native space.")
for i in range(number_of_classification_labels):
probability_images_array[i] = ants.apply_ants_transform_to_image(
ants.invert_ants_transform(xfrm), probability_images_array[i], image)
image_matrix = ants.image_list_to_matrix(probability_images_array, image * 0 + 1)
segmentation_matrix = np.argmax(image_matrix, axis=0)
segmentation_image = ants.matrix_to_images(
np.expand_dims(segmentation_matrix, axis=0), image * 0 + 1)[0]
if modality == "protonLobes":
whole_lung_mask = ants.from_numpy(np.squeeze(predicted_data[1][0, :, :, :, 0]),
origin=origin, spacing=spacing, direction=direction)
whole_lung_mask = ants.apply_ants_transform_to_image(
ants.invert_ants_transform(xfrm), whole_lung_mask, image)
return_dict = {'segmentation_image' : segmentation_image,
'probability_images' : probability_images_array,
'whole_lung_mask_image' : whole_lung_mask}
return(return_dict)
else:
return_dict = {'segmentation_image' : segmentation_image,
'probability_images' : probability_images_array}
return(return_dict)
elif modality == "ct":
################################
#
# Preprocess image
#
################################
if verbose == True:
print("Preprocess CT image.")
def closest_simplified_direction_matrix(direction):
closest = np.floor(np.abs(direction) + 0.5)
closest[direction < 0] *= -1.0
return closest
simplified_direction = closest_simplified_direction_matrix(image.direction)
reference_image_size = (128, 128, 128)
ct_preprocessed = ants.resample_image(image, reference_image_size, use_voxels=True, interp_type=0)
ct_preprocessed[ct_preprocessed < -1000] = -1000
ct_preprocessed[ct_preprocessed > 400] = 400
ct_preprocessed.set_direction(simplified_direction)
ct_preprocessed.set_origin((0, 0, 0))
ct_preprocessed.set_spacing((1, 1, 1))
################################
#
# Reorient image
#
################################
reference_image = ants.make_image(reference_image_size,
voxval=0,
spacing=(1, 1, 1),
origin=(0, 0, 0),
direction=np.identity(3))
center_of_mass_reference = np.floor(ants.get_center_of_mass(reference_image * 0 + 1))
center_of_mass_image = np.floor(ants.get_center_of_mass(ct_preprocessed * 0 + 1))
translation = np.asarray(center_of_mass_image) - np.asarray(center_of_mass_reference)
xfrm = ants.create_ants_transform(transform_type="Euler3DTransform",
center=np.asarray(center_of_mass_reference), translation=translation)
ct_preprocessed = ((ct_preprocessed - ct_preprocessed.min()) /
(ct_preprocessed.max() - ct_preprocessed.min()))
ct_preprocessed_warped = ants.apply_ants_transform_to_image(
xfrm, ct_preprocessed, reference_image, interpolation="nearestneighbor")
ct_preprocessed_warped = ((ct_preprocessed_warped - ct_preprocessed_warped.min()) /
(ct_preprocessed_warped.max() - ct_preprocessed_warped.min())) - 0.5
################################
#
# Build models and load weights
#
################################
if verbose == True:
print("Build model and load weights.")
weights_file_name = get_pretrained_network("lungCtWithPriorsSegmentationWeights",
antsxnet_cache_directory=antsxnet_cache_directory)
classes = ("background", "left lung", "right lung", "airways")
number_of_classification_labels = len(classes)
luna16_priors = ants.ndimage_to_list(ants.image_read(get_antsxnet_data("luna16LungPriors")))
for i in range(len(luna16_priors)):
luna16_priors[i] = ants.resample_image(luna16_priors[i], reference_image_size, use_voxels=True)
channel_size = len(luna16_priors) + 1
unet_model = create_unet_model_3d((*reference_image_size, channel_size),
number_of_outputs=number_of_classification_labels, mode="classification",
number_of_layers=4, number_of_filters_at_base_layer=16, dropout_rate=0.0,
convolution_kernel_size=(3, 3, 3), deconvolution_kernel_size=(2, 2, 2),
weight_decay=1e-5, additional_options=("attentionGating",))
unet_model.load_weights(weights_file_name)
################################
#
# Do prediction and normalize to native space
#
################################
if verbose == True:
print("Prediction.")
batchX = np.zeros((1, *reference_image_size, channel_size))
batchX[:,:,:,:,0] = ct_preprocessed_warped.numpy()
for i in range(len(luna16_priors)):
batchX[:,:,:,:,i+1] = luna16_priors[i].numpy() - 0.5
predicted_data = unet_model.predict(batchX, verbose=verbose)
probability_images = list()
for i in range(number_of_classification_labels):
if verbose == True:
print("Reconstructing image", classes[i])
probability_image = ants.from_numpy(np.squeeze(predicted_data[:,:,:,:,i]),
origin=ct_preprocessed_warped.origin, spacing=ct_preprocessed_warped.spacing,
direction=ct_preprocessed_warped.direction)
probability_image = ants.apply_ants_transform_to_image(
ants.invert_ants_transform(xfrm), probability_image, ct_preprocessed)
probability_image = ants.resample_image(probability_image,
resample_params=image.shape, use_voxels=True, interp_type=0)
probability_image = ants.copy_image_info(image, probability_image)
probability_images.append(probability_image)
image_matrix = ants.image_list_to_matrix(probability_images, image * 0 + 1)
segmentation_matrix = np.argmax(image_matrix, axis=0)
segmentation_image = ants.matrix_to_images(
np.expand_dims(segmentation_matrix, axis=0), image * 0 + 1)[0]
return_dict = {'segmentation_image' : segmentation_image,
'probability_images' : probability_images}
return(return_dict)
elif modality == "ventilation":
################################
#
# Preprocess image
#
################################
if verbose == True:
print("Preprocess ventilation image.")
template_size = (256, 256)
image_modalities = ("Ventilation",)
channel_size = len(image_modalities)
preprocessed_image = (image - image.mean()) / image.std()
ants.set_direction(preprocessed_image, np.identity(3))
################################
#
# Build models and load weights
#
################################
unet_model = create_unet_model_2d((*template_size, channel_size),
number_of_outputs=1, mode='sigmoid',
number_of_layers=4, number_of_filters_at_base_layer=32, dropout_rate=0.0,
convolution_kernel_size=(3, 3), deconvolution_kernel_size=(2, 2),
weight_decay=0)
if verbose == True:
print("Whole lung mask: retrieving model weights.")
weights_file_name = get_pretrained_network("wholeLungMaskFromVentilation",
antsxnet_cache_directory=antsxnet_cache_directory)
unet_model.load_weights(weights_file_name)
################################
#
# Extract slices
#
################################
spacing = ants.get_spacing(preprocessed_image)
dimensions_to_predict = (spacing.index(max(spacing)),)
total_number_of_slices = 0
for d in range(len(dimensions_to_predict)):
total_number_of_slices += preprocessed_image.shape[dimensions_to_predict[d]]
batchX = np.zeros((total_number_of_slices, *template_size, channel_size))
slice_count = 0
for d in range(len(dimensions_to_predict)):
number_of_slices = preprocessed_image.shape[dimensions_to_predict[d]]
if verbose == True:
print("Extracting slices for dimension ", dimensions_to_predict[d], ".")
for i in range(number_of_slices):
ventilation_slice = pad_or_crop_image_to_size(ants.slice_image(preprocessed_image, dimensions_to_predict[d], i), template_size)
batchX[slice_count,:,:,0] = ventilation_slice.numpy()
slice_count += 1
################################
#
# Do prediction and then restack into the image
#
################################
if verbose == True:
print("Prediction.")
prediction = unet_model.predict(batchX, verbose=verbose)
permutations = list()
permutations.append((0, 1, 2))
permutations.append((1, 0, 2))
permutations.append((1, 2, 0))
probability_image = ants.image_clone(image) * 0
current_start_slice = 0
for d in range(len(dimensions_to_predict)):
current_end_slice = current_start_slice + preprocessed_image.shape[dimensions_to_predict[d]] - 1
which_batch_slices = range(current_start_slice, current_end_slice)
prediction_per_dimension = prediction[which_batch_slices,:,:,0]
prediction_array = np.transpose(np.squeeze(prediction_per_dimension), permutations[dimensions_to_predict[d]])
prediction_image = ants.copy_image_info(image,
pad_or_crop_image_to_size(ants.from_numpy(prediction_array),
image.shape))
probability_image = probability_image + (prediction_image - probability_image) / (d + 1)
current_start_slice = current_end_slice + 1
return(probability_image)
else:
return ValueError("Unrecognized modality.")
gmmat = ants.timeseries_to_matrix( simg, gmseg )
nuisance = mycompcor['components']
nuisance = np.c_[ nuisance, motCorr['FD'] ]
nuisance = np.c_[ nuisance, mycompcor['basis'] ]
gmmat = regress_components( gmmat[goodtimes[0],:], nuisance[goodtimes[0],:] )
dfnmat = ants.timeseries_to_matrix( simg, ants.threshold_image( dfnImg, 1, dfnImg.max() ) )
dfnmat = regress_components( dfnmat[goodtimes[0],:], nuisance[goodtimes[0],:] )
# dfnmatf = frequencyFilterfMRI( dfnmat, tr = tr, freqLo = 0.01, freqHi = 0.1, opt = "trig" )
dfnsignal = dfnmat.mean( axis = 1 )
gmmatDFNCorr = np.zeros( gmmat.shape[1] )
for k in range( gmmat.shape[1] ):
gmmatDFNCorr[ k ] = pearsonr( dfnsignal, gmmat[:,k] )[0]
corrImg = ants.make_image( gmseg, gmmatDFNCorr )
corrImgPos = corrImg * ants.threshold_image( corrImg, 0.25, 1 )
ants.plot( avgBold, corrImgPos, axis=2, overlay_alpha = 0.6, cbar=False, nslices = 24, ncol=8, cbar_length=0.3, cbar_vertical=True )
"""
needs to be translated to python still
TODO:
ilr - image based linear regression
frequencyFilterfMRI - scipy probably butter filtfilt
getNetworkBeta <-function( motcorrIn, networkName = 'Default Mode' ) {
def sysu_media_wmh_segmentation(flair,
t1=None,
do_preprocessing=True,
use_ensemble=True,
use_axial_slices_only=True,
antsxnet_cache_directory=None,
verbose=False):
"""
Perform WMH segmentation using the winning submission in the MICCAI
2017 challenge by the sysu_media team using FLAIR or T1/FLAIR. The
MICCAI challenge is discussed in
https://pubmed.ncbi.nlm.nih.gov/30908194/
with the sysu_media's team entry is discussed in
https://pubmed.ncbi.nlm.nih.gov/30125711/
with the original implementation available here:
https://github.com/hongweilibran/wmh_ibbmTum
Arguments
---------
flair : ANTsImage
input 3-D FLAIR brain image (not skull-stripped).
t1 : ANTsImage
input 3-D T1 brain image (not skull-stripped).
do_preprocessing : boolean
perform n4 bias correction?
use_ensemble : boolean
check whether to use all 3 sets of weights.
use_axial_slices_only : boolean
If True, use original implementation which was trained on axial slices.
If False, use ANTsXNet variant implementation which applies the slice-by-slice
models to all 3 dimensions and averages the results.
antsxnet_cache_directory : string
Destination directory for storing the downloaded template and model weights.
Since these can be resused, if is None, these data will be downloaded to a
~/.keras/ANTsXNet/.
verbose : boolean
Print progress to the screen.
Returns
-------
WMH segmentation probability image
Example
-------
>>> image = ants.image_read("flair.nii.gz")
>>> probability_mask = sysu_media_wmh_segmentation(image)
"""
from ..architectures import create_sysu_media_unet_model_2d
from ..utilities import brain_extraction
from ..utilities import crop_image_center
from ..utilities import get_pretrained_network
from ..utilities import preprocess_brain_image
from ..utilities import pad_or_crop_image_to_size
if flair.dimension != 3:
raise ValueError("Image dimension must be 3.")
if antsxnet_cache_directory == None:
antsxnet_cache_directory = "ANTsXNet"
################################
#
# Preprocess images
#
################################
flair_preprocessed = flair
if do_preprocessing == True:
flair_preprocessing = preprocess_brain_image(
flair,
truncate_intensity=(0.01, 0.99),
do_brain_extraction=False,
do_bias_correction=True,
do_denoising=False,
antsxnet_cache_directory=antsxnet_cache_directory,
verbose=verbose)
flair_preprocessed = flair_preprocessing["preprocessed_image"]
number_of_channels = 1
if t1 is not None:
t1_preprocessed = t1
if do_preprocessing == True:
t1_preprocessing = preprocess_brain_image(
t1,
truncate_intensity=(0.01, 0.99),
do_brain_extraction=False,
do_bias_correction=True,
do_denoising=False,
antsxnet_cache_directory=antsxnet_cache_directory,
verbose=verbose)
t1_preprocessed = t1_preprocessing["preprocessed_image"]
number_of_channels = 2
################################
#
# Estimate mask
#
################################
brain_mask = None
if verbose == True:
print("Estimating brain mask.")
if t1 is not None:
brain_mask = brain_extraction(t1, modality="t1")
else:
brain_mask = brain_extraction(flair, modality="flair")
reference_image = ants.make_image((200, 200, 200),
voxval=1,
spacing=(1, 1, 1),
origin=(0, 0, 0),
direction=np.identity(3))
center_of_mass_reference = ants.get_center_of_mass(reference_image)
center_of_mass_image = ants.get_center_of_mass(brain_mask)
translation = np.asarray(center_of_mass_image) - np.asarray(
center_of_mass_reference)
xfrm = ants.create_ants_transform(
transform_type="Euler3DTransform",
center=np.asarray(center_of_mass_reference),
translation=translation)
flair_preprocessed_warped = ants.apply_ants_transform_to_image(
xfrm, flair_preprocessed, reference_image)
brain_mask_warped = ants.threshold_image(
ants.apply_ants_transform_to_image(xfrm, brain_mask, reference_image),
0.5, 1.1, 1, 0)
if t1 is not None:
t1_preprocessed_warped = ants.apply_ants_transform_to_image(
xfrm, t1_preprocessed, reference_image)
################################
#
# Gaussian normalize intensity based on brain mask
#
################################
mean_flair = flair_preprocessed_warped[brain_mask_warped > 0].mean()
std_flair = flair_preprocessed_warped[brain_mask_warped > 0].std()
flair_preprocessed_warped = (flair_preprocessed_warped -
mean_flair) / std_flair
if number_of_channels == 2:
mean_t1 = t1_preprocessed_warped[brain_mask_warped > 0].mean()
std_t1 = t1_preprocessed_warped[brain_mask_warped > 0].std()
t1_preprocessed_warped = (t1_preprocessed_warped - mean_t1) / std_t1
################################
#
# Build models and load weights
#
################################
number_of_models = 1
if use_ensemble == True:
number_of_models = 3
unet_models = list()
for i in range(number_of_models):
if number_of_channels == 1:
weights_file_name = get_pretrained_network(
"sysuMediaWmhFlairOnlyModel" + str(i),
antsxnet_cache_directory=antsxnet_cache_directory)
else:
weights_file_name = get_pretrained_network(
"sysuMediaWmhFlairT1Model" + str(i),
antsxnet_cache_directory=antsxnet_cache_directory)
unet_models.append(
create_sysu_media_unet_model_2d((200, 200, number_of_channels)))
unet_models[i].load_weights(weights_file_name)
################################
#
# Extract slices
#
################################
dimensions_to_predict = [2]
if use_axial_slices_only == False:
dimensions_to_predict = list(range(3))
total_number_of_slices = 0
for d in range(len(dimensions_to_predict)):
total_number_of_slices += flair_preprocessed_warped.shape[
dimensions_to_predict[d]]
batchX = np.zeros((total_number_of_slices, 200, 200, number_of_channels))
slice_count = 0
for d in range(len(dimensions_to_predict)):
number_of_slices = flair_preprocessed_warped.shape[
dimensions_to_predict[d]]
if verbose == True:
print("Extracting slices for dimension ", dimensions_to_predict[d],
".")
for i in range(number_of_slices):
flair_slice = pad_or_crop_image_to_size(
ants.slice_image(flair_preprocessed_warped,
dimensions_to_predict[d], i), (200, 200))
batchX[slice_count, :, :, 0] = flair_slice.numpy()
if number_of_channels == 2:
t1_slice = pad_or_crop_image_to_size(
ants.slice_image(t1_preprocessed_warped,
dimensions_to_predict[d], i), (200, 200))
batchX[slice_count, :, :, 1] = t1_slice.numpy()
slice_count += 1
################################
#
# Do prediction and then restack into the image
#
################################
if verbose == True:
print("Prediction.")
prediction = unet_models[0].predict(batchX, verbose=verbose)
if number_of_models > 1:
for i in range(1, number_of_models, 1):
prediction += unet_models[i].predict(batchX, verbose=verbose)
prediction /= number_of_models
permutations = list()
permutations.append((0, 1, 2))
permutations.append((1, 0, 2))
permutations.append((1, 2, 0))
prediction_image_average = ants.image_clone(flair_preprocessed_warped) * 0
current_start_slice = 0
for d in range(len(dimensions_to_predict)):
current_end_slice = current_start_slice + flair_preprocessed_warped.shape[
dimensions_to_predict[d]] - 1
which_batch_slices = range(current_start_slice, current_end_slice)
prediction_per_dimension = prediction[which_batch_slices, :, :, :]
prediction_array = np.transpose(np.squeeze(prediction_per_dimension),
permutations[dimensions_to_predict[d]])
prediction_image = ants.copy_image_info(
flair_preprocessed_warped,
pad_or_crop_image_to_size(ants.from_numpy(prediction_array),
flair_preprocessed_warped.shape))
prediction_image_average = prediction_image_average + (
prediction_image - prediction_image_average) / (d + 1)
current_start_slice = current_end_slice + 1
probability_image = ants.apply_ants_transform_to_image(
ants.invert_ants_transform(xfrm), prediction_image_average, flair)
return (probability_image)
outputPrefix = outputDirectory + 'antspy'
numberOfOuterIterations = 5
image = ants.image_read( dataDirectory + 'KKI2009-01-MPRAGE_slice150.nii.gz', dimension = 2 )
mask = ants.image_read( dataDirectory + 'KKI2009-01-MPRAGE_slice150_mask.nii.gz', dimension = 2 )
weightMask = None
for i in range( numberOfOuterIterations ):
print( "*************** N4 <---> Atropos iteration ", i, " ******************\n" )
n4Results = ants.n4_bias_field_correction( image, mask = mask,
weight_mask = weightMask, verbose = True )
image = n4Results
atroposResults = ants.atropos( a = image, x = mask )
onesImage = ants.make_image( image.shape, voxval = 0,
spacing = ants.get_spacing( image ),
origin = ants.get_origin( image ),
direction = ants.get_direction( image ) )
weightMask = atroposResults['probabilityimages'][1] *\
( onesImage - atroposResults['probabilityimages'][0] ) *\
( onesImage - atroposResults['probabilityimages'][2] ) +\
atroposResults['probabilityimages'][2] *\
( onesImage - atroposResults['probabilityimages'][1] ) *\
( onesImage - atroposResults['probabilityimages'][0] )
ants.image_write( image, outputPrefix + "N4Corrected.nii.gz" )
ants.image_write( atroposResults['segmentation'], outputPrefix + "AtroposSegmentation.nii.gz" )
for i in range( len( atroposResults['probabilityimages'] ) ):
ants.image_write( atroposResults['probabilityimages'][i],
outputPrefix + "AtroposSegmentationProbability" + str( i ) + ".nii.gz" )
def test_convolve_image_example(self):
fi = ants.image_read(ants.get_ants_data('r16'))
convimg = ants.make_image((3, 3), (1, 0, 1, 0, -4, 0, 1, 0, 1))
convout = ants.convolve_image(fi, convimg)
convimg2 = ants.make_image((3, 3), (0, 1, 0, 1, 0, -1, 0, -1, 0))
convout2 = ants.convolve_image(fi, convimg2)
def claustrum_segmentation(t1,
do_preprocessing=True,
use_ensemble=True,
antsxnet_cache_directory=None,
verbose=False):
"""
Claustrum segmentation
Described here:
https://arxiv.org/abs/2008.03465
with the implementation available at:
https://github.com/hongweilibran/claustrum_multi_view
Arguments
---------
t1 : ANTsImage
input 3-D T1 brain image.
do_preprocessing : boolean
perform n4 bias correction.
use_ensemble : boolean
check whether to use all 3 sets of weights.
antsxnet_cache_directory : string
Destination directory for storing the downloaded template and model weights.
Since these can be resused, if is None, these data will be downloaded to a
~/.keras/ANTsXNet/.
verbose : boolean
Print progress to the screen.
Returns
-------
Claustrum segmentation probability image
Example
-------
>>> image = ants.image_read("t1.nii.gz")
>>> probability_mask = claustrum_segmentation(image)
"""
from ..architectures import create_sysu_media_unet_model_2d
from ..utilities import brain_extraction
from ..utilities import get_pretrained_network
from ..utilities import preprocess_brain_image
from ..utilities import pad_or_crop_image_to_size
if t1.dimension != 3:
raise ValueError("Image dimension must be 3.")
if antsxnet_cache_directory == None:
antsxnet_cache_directory = "ANTsXNet"
image_size = (180, 180)
################################
#
# Preprocess images
#
################################
number_of_channels = 1
t1_preprocessed = ants.image_clone(t1)
brain_mask = ants.threshold_image(t1, 0, 0, 0, 1)
if do_preprocessing == True:
t1_preprocessing = preprocess_brain_image(
t1,
truncate_intensity=(0.01, 0.99),
brain_extraction_modality="t1",
do_bias_correction=True,
do_denoising=True,
antsxnet_cache_directory=antsxnet_cache_directory,
verbose=verbose)
t1_preprocessed = t1_preprocessing["preprocessed_image"]
brain_mask = t1_preprocessing["brain_mask"]
reference_image = ants.make_image((170, 256, 256),
voxval=1,
spacing=(1, 1, 1),
origin=(0, 0, 0),
direction=np.identity(3))
center_of_mass_reference = ants.get_center_of_mass(reference_image)
center_of_mass_image = ants.get_center_of_mass(brain_mask)
translation = np.asarray(center_of_mass_image) - np.asarray(
center_of_mass_reference)
xfrm = ants.create_ants_transform(
transform_type="Euler3DTransform",
center=np.asarray(center_of_mass_reference),
translation=translation)
t1_preprocessed_warped = ants.apply_ants_transform_to_image(
xfrm, t1_preprocessed, reference_image)
brain_mask_warped = ants.threshold_image(
ants.apply_ants_transform_to_image(xfrm, brain_mask, reference_image),
0.5, 1.1, 1, 0)
################################
#
# Gaussian normalize intensity based on brain mask
#
################################
mean_t1 = t1_preprocessed_warped[brain_mask_warped > 0].mean()
std_t1 = t1_preprocessed_warped[brain_mask_warped > 0].std()
t1_preprocessed_warped = (t1_preprocessed_warped - mean_t1) / std_t1
t1_preprocessed_warped = t1_preprocessed_warped * brain_mask_warped
################################
#
# Build models and load weights
#
################################
number_of_models = 1
if use_ensemble == True:
number_of_models = 3
if verbose == True:
print("Claustrum: retrieving axial model weights.")
unet_axial_models = list()
for i in range(number_of_models):
weights_file_name = get_pretrained_network(
"claustrum_axial_" + str(i),
antsxnet_cache_directory=antsxnet_cache_directory)
unet_axial_models.append(
create_sysu_media_unet_model_2d((*image_size, number_of_channels),
anatomy="claustrum"))
unet_axial_models[i].load_weights(weights_file_name)
if verbose == True:
print("Claustrum: retrieving coronal model weights.")
unet_coronal_models = list()
for i in range(number_of_models):
weights_file_name = get_pretrained_network(
"claustrum_coronal_" + str(i),
antsxnet_cache_directory=antsxnet_cache_directory)
unet_coronal_models.append(
create_sysu_media_unet_model_2d((*image_size, number_of_channels),
anatomy="claustrum"))
unet_coronal_models[i].load_weights(weights_file_name)
################################
#
# Extract slices
#
################################
dimensions_to_predict = [1, 2]
batch_coronal_X = np.zeros(
(t1_preprocessed_warped.shape[1], *image_size, number_of_channels))
batch_axial_X = np.zeros(
(t1_preprocessed_warped.shape[2], *image_size, number_of_channels))
for d in range(len(dimensions_to_predict)):
number_of_slices = t1_preprocessed_warped.shape[
dimensions_to_predict[d]]
if verbose == True:
print("Extracting slices for dimension ", dimensions_to_predict[d],
".")
for i in range(number_of_slices):
t1_slice = pad_or_crop_image_to_size(
ants.slice_image(t1_preprocessed_warped,
dimensions_to_predict[d], i), image_size)
if dimensions_to_predict[d] == 1:
batch_coronal_X[i, :, :, 0] = np.rot90(t1_slice.numpy(), k=-1)
else:
batch_axial_X[i, :, :, 0] = np.rot90(t1_slice.numpy())
################################
#
# Do prediction and then restack into the image
#
################################
if verbose == True:
print("Coronal prediction.")
prediction_coronal = unet_coronal_models[0].predict(batch_coronal_X,
verbose=verbose)
if number_of_models > 1:
for i in range(1, number_of_models, 1):
prediction_coronal += unet_coronal_models[i].predict(
batch_coronal_X, verbose=verbose)
prediction_coronal /= number_of_models
for i in range(t1_preprocessed_warped.shape[1]):
prediction_coronal[i, :, :, 0] = np.rot90(
np.squeeze(prediction_coronal[i, :, :, 0]))
if verbose == True:
print("Axial prediction.")
prediction_axial = unet_axial_models[0].predict(batch_axial_X,
verbose=verbose)
if number_of_models > 1:
for i in range(1, number_of_models, 1):
prediction_axial += unet_axial_models[i].predict(batch_axial_X,
verbose=verbose)
prediction_axial /= number_of_models
for i in range(t1_preprocessed_warped.shape[2]):
prediction_axial[i, :, :,
0] = np.rot90(np.squeeze(prediction_axial[i, :, :,
0]),
k=-1)
if verbose == True:
print("Restack image and transform back to native space.")
permutations = list()
permutations.append((0, 1, 2))
permutations.append((1, 0, 2))
permutations.append((1, 2, 0))
prediction_image_average = ants.image_clone(t1_preprocessed_warped) * 0
for d in range(len(dimensions_to_predict)):
which_batch_slices = range(
t1_preprocessed_warped.shape[dimensions_to_predict[d]])
prediction_per_dimension = None
if dimensions_to_predict[d] == 1:
prediction_per_dimension = prediction_coronal[
which_batch_slices, :, :, :]
else:
prediction_per_dimension = prediction_axial[
which_batch_slices, :, :, :]
prediction_array = np.transpose(np.squeeze(prediction_per_dimension),
permutations[dimensions_to_predict[d]])
prediction_image = ants.copy_image_info(
t1_preprocessed_warped,
pad_or_crop_image_to_size(ants.from_numpy(prediction_array),
t1_preprocessed_warped.shape))
prediction_image_average = prediction_image_average + (
prediction_image - prediction_image_average) / (d + 1)
probability_image = ants.apply_ants_transform_to_image(
ants.invert_ants_transform(xfrm), prediction_image_average,
t1) * ants.threshold_image(brain_mask, 0.5, 1, 1, 0)
return (probability_image)
def apply_super_resolution_model_to_image(
image,
model,
target_range=(-127.5, 127.5),
batch_size=32,
regression_order=None,
verbose=False,
):
"""
Apply a pretrained deep back projection model for super resolution.
Helper function for applying a pretrained deep back projection model.
Apply a patch-wise trained network to perform super-resolution. Can be applied
to variable sized inputs. Warning: This function may be better used on CPU
unless the GPU can accommodate the full image size. Warning 2: The global
intensity range (min to max) of the output will match the input where the
range is taken over all channels.
Arguments
---------
image : ANTs image
input image.
model : keras object or string
pretrained keras model or filename.
target_range : 2-element tuple
a tuple or array defining the (min, max) of the input image
(e.g., -127.5, 127.5). Output images will be scaled back to original
intensity. This range should match the mapping used in the training
of the network.
batch_size : integer
Batch size used for the prediction call.
regression_order : integer
If specified, Apply the function regression_match_image with
poly_order=regression_order.
verbose : boolean
If True, show status messages.
Returns
-------
Super-resolution image upscaled to resolution specified by the network.
Example
-------
>>> import ants
>>> image = ants.image_read(ants.get_ants_data('r16'))
>>> image_sr = apply_super_resolution_model_to_image(image, get_pretrained_network("dbpn4x"))
"""
tflite_flag = False
channel_axis = 0
if K.image_data_format() == "channels_last":
channel_axis = -1
if target_range[0] > target_range[1]:
target_range = target_range[::-1]
start_time = time.time()
if isinstance(model, str):
if path.isfile(model):
if verbose:
print("Load model.")
if path.splitext(model)[1] == '.tflite':
interpreter = tf.lite.Interpreter(model)
interpreter.allocate_tensors()
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
shape_length = len(interpreter.get_input_details()[0]['shape'])
else:
model = load_model(model)
shape_length = len(model.input_shape)
if verbose:
elapsed_time = time.time() - start_time
print(" (elapsed time: ", elapsed_time, ")")
else:
raise ValueError("Model not found.")
else:
shape_length = len(model.input_shape)
if shape_length < 4 | shape_length > 5:
raise ValueError("Unexpected input shape.")
else:
if shape_length == 5 & image.dimension != 3:
raise ValueError("Expecting 3D input for this model.")
elif shape_length == 4 & image.dimension != 2:
raise ValueError("Expecting 2D input for this model.")
if channel_axis == -1:
channel_axis < shape_length
if tflite_flag:
channel_size = interpreter.get_input_details()[0]['shape'][channel_axis]
else:
channel_size = model.input_shape[channel_axis]
if channel_size != image.components:
raise ValueError(
"Channel size of model",
str(channel_size),
"does not match ncomponents=",
str(image.components),
"of the input image.",
)
image_patches = extract_image_patches(
image,
patch_size=image.shape,
max_number_of_patches=1,
stride_length=image.shape,
return_as_array=True,
)
if image.components == 1:
image_patches = np.expand_dims(image_patches, axis=-1)
image_patches = image_patches - image_patches.min()
image_patches = (
image_patches / image_patches.max() * (target_range[1] - target_range[0])
+ target_range[0]
)
if verbose:
print("Prediction")
start_time = time.time()
if tflite_flag:
image_patches = image_patches.astype('float32')
interpreter.set_tensor(input_details[0]['index'], image_patches)
interpreter.invoke()
out = interpreter.tensor(output_details[0]['index'])
prediction = out()
else:
prediction = model.predict(image_patches, batch_size=batch_size)
if verbose:
elapsed_time = time.time() - start_time
print(" (elapsed time: ", elapsed_time, ")")
if verbose:
print("Reconstruct intensities")
intensity_range = image.range()
prediction = prediction - prediction.min()
prediction = (
prediction / prediction.max() * (intensity_range[1] - intensity_range[0])
+ intensity_range[0]
)
def slice_array_channel(input_array, slice, channel_axis=-1):
if channel_axis == 0:
if shape_length == 4:
return input_array[slice, :, :, :]
else:
return input_array[slice, :, :, :, :]
else:
if shape_length == 4:
return input_array[:, :, :, slice]
else:
return input_array[:, :, :, :, slice]
expansion_factor = np.asarray(prediction.shape) / np.asarray(image_patches.shape)
if channel_axis == 0:
FIXME
expansion_factor = expansion_factor[1 : (len(expansion_factor) - 1)]
if verbose:
print("ExpansionFactor:", str(expansion_factor))
if image.components == 1:
image_array = slice_array_channel(prediction, 0, channel_axis)
prediction_image = ants.make_image(
(np.asarray(image.shape) * np.asarray(expansion_factor)).astype(int),
image_array,
)
if regression_order is not None:
reference_image = ants.resample_image_to_target(image, prediction_image)
prediction_image = regression_match_image(
prediction_image, reference_image, poly_order=regression_order
)
else:
image_component_list = list()
for k in range(image.components):
image_array = slice_array_channel(prediction, k, channel_axis)
image_component_list.append(
ants.make_image(
(np.asarray(image.shape) * np.asarray(expansion_factor)).astype(
int
),
image_array,
)
)
prediction_image = ants.merge_channels(image_component_list)
prediction_image = ants.copy_image_info(image, prediction_image)
ants.set_spacing(
prediction_image,
tuple(np.asarray(image.spacing) / np.asarray(expansion_factor)),
)
return prediction_image
def sysu_media_wmh_segmentation(flair,
t1=None,
use_ensemble=True,
antsxnet_cache_directory=None,
verbose=False):
"""
Perform WMH segmentation using the winning submission in the MICCAI
2017 challenge by the sysu_media team using FLAIR or T1/FLAIR. The
MICCAI challenge is discussed in
https://pubmed.ncbi.nlm.nih.gov/30908194/
with the sysu_media's team entry is discussed in
https://pubmed.ncbi.nlm.nih.gov/30125711/
with the original implementation available here:
https://github.com/hongweilibran/wmh_ibbmTum
The original implementation used global thresholding as a quick
brain extraction approach. Due to possible generalization difficulties,
we leave such post-processing steps to the user. For brain or white
matter masking see functions brain_extraction or deep_atropos,
respectively.
Arguments
---------
flair : ANTsImage
input 3-D FLAIR brain image (not skull-stripped).
t1 : ANTsImage
input 3-D T1 brain image (not skull-stripped).
use_ensemble : boolean
check whether to use all 3 sets of weights.
antsxnet_cache_directory : string
Destination directory for storing the downloaded template and model weights.
Since these can be resused, if is None, these data will be downloaded to a
~/.keras/ANTsXNet/.
verbose : boolean
Print progress to the screen.
Returns
-------
WMH segmentation probability image
Example
-------
>>> image = ants.image_read("flair.nii.gz")
>>> probability_mask = sysu_media_wmh_segmentation(image)
"""
from ..architectures import create_sysu_media_unet_model_2d
from ..utilities import get_pretrained_network
from ..utilities import pad_or_crop_image_to_size
from ..utilities import preprocess_brain_image
from ..utilities import binary_dice_coefficient
if flair.dimension != 3:
raise ValueError("Image dimension must be 3.")
if antsxnet_cache_directory == None:
antsxnet_cache_directory = "ANTsXNet"
image_size = (200, 200)
################################
#
# Preprocess images
#
################################
def closest_simplified_direction_matrix(direction):
closest = (np.abs(direction) + 0.5).astype(int).astype(float)
closest[direction < 0] *= -1.0
return closest
simplified_direction = closest_simplified_direction_matrix(flair.direction)
flair_preprocessing = preprocess_brain_image(
flair,
truncate_intensity=None,
brain_extraction_modality=None,
do_bias_correction=False,
do_denoising=False,
antsxnet_cache_directory=antsxnet_cache_directory,
verbose=verbose)
flair_preprocessed = flair_preprocessing["preprocessed_image"]
flair_preprocessed.set_direction(simplified_direction)
flair_preprocessed.set_origin((0, 0, 0))
flair_preprocessed.set_spacing((1, 1, 1))
number_of_channels = 1
t1_preprocessed = None
if t1 is not None:
t1_preprocessing = preprocess_brain_image(
t1,
truncate_intensity=None,
brain_extraction_modality=None,
do_bias_correction=False,
do_denoising=False,
antsxnet_cache_directory=antsxnet_cache_directory,
verbose=verbose)
t1_preprocessed = t1_preprocessing["preprocessed_image"]
t1_preprocessed.set_direction(simplified_direction)
t1_preprocessed.set_origin((0, 0, 0))
t1_preprocessed.set_spacing((1, 1, 1))
number_of_channels = 2
################################
#
# Reorient images
#
################################
reference_image = ants.make_image((256, 256, 256),
voxval=0,
spacing=(1, 1, 1),
origin=(0, 0, 0),
direction=np.identity(3))
center_of_mass_reference = np.floor(
ants.get_center_of_mass(reference_image * 0 + 1))
center_of_mass_image = np.floor(
ants.get_center_of_mass(flair_preprocessed))
translation = np.asarray(center_of_mass_image) - np.asarray(
center_of_mass_reference)
xfrm = ants.create_ants_transform(
transform_type="Euler3DTransform",
center=np.asarray(center_of_mass_reference),
translation=translation)
flair_preprocessed_warped = ants.apply_ants_transform_to_image(
xfrm,
flair_preprocessed,
reference_image,
interpolation="nearestneighbor")
crop_image = ants.image_clone(flair_preprocessed) * 0 + 1
crop_image_warped = ants.apply_ants_transform_to_image(
xfrm, crop_image, reference_image, interpolation="nearestneighbor")
flair_preprocessed_warped = ants.crop_image(flair_preprocessed_warped,
crop_image_warped, 1)
if t1 is not None:
t1_preprocessed_warped = ants.apply_ants_transform_to_image(
xfrm,
t1_preprocessed,
reference_image,
interpolation="nearestneighbor")
t1_preprocessed_warped = ants.crop_image(t1_preprocessed_warped,
crop_image_warped, 1)
################################
#
# Gaussian normalize intensity
#
################################
mean_flair = flair_preprocessed.mean()
std_flair = flair_preprocessed.std()
if number_of_channels == 2:
mean_t1 = t1_preprocessed.mean()
std_t1 = t1_preprocessed.std()
flair_preprocessed_warped = (flair_preprocessed_warped -
mean_flair) / std_flair
if number_of_channels == 2:
t1_preprocessed_warped = (t1_preprocessed_warped - mean_t1) / std_t1
################################
#
# Build models and load weights
#
################################
number_of_models = 1
if use_ensemble == True:
number_of_models = 3
if verbose == True:
print("White matter hyperintensity: retrieving model weights.")
unet_models = list()
for i in range(number_of_models):
if number_of_channels == 1:
weights_file_name = get_pretrained_network(
"sysuMediaWmhFlairOnlyModel" + str(i),
antsxnet_cache_directory=antsxnet_cache_directory)
else:
weights_file_name = get_pretrained_network(
"sysuMediaWmhFlairT1Model" + str(i),
antsxnet_cache_directory=antsxnet_cache_directory)
unet_model = create_sysu_media_unet_model_2d(
(*image_size, number_of_channels))
unet_loss = binary_dice_coefficient(smoothing_factor=1.)
unet_model.compile(optimizer=keras.optimizers.Adam(learning_rate=2e-4),
loss=unet_loss)
unet_model.load_weights(weights_file_name)
unet_models.append(unet_model)
################################
#
# Extract slices
#
################################
dimensions_to_predict = [2]
total_number_of_slices = 0
for d in range(len(dimensions_to_predict)):
total_number_of_slices += flair_preprocessed_warped.shape[
dimensions_to_predict[d]]
batchX = np.zeros(
(total_number_of_slices, *image_size, number_of_channels))
slice_count = 0
for d in range(len(dimensions_to_predict)):
number_of_slices = flair_preprocessed_warped.shape[
dimensions_to_predict[d]]
if verbose == True:
print("Extracting slices for dimension ", dimensions_to_predict[d],
".")
for i in range(number_of_slices):
flair_slice = pad_or_crop_image_to_size(
ants.slice_image(flair_preprocessed_warped,
dimensions_to_predict[d], i), image_size)
batchX[slice_count, :, :, 0] = flair_slice.numpy()
if number_of_channels == 2:
t1_slice = pad_or_crop_image_to_size(
ants.slice_image(t1_preprocessed_warped,
dimensions_to_predict[d], i), image_size)
batchX[slice_count, :, :, 1] = t1_slice.numpy()
slice_count += 1
################################
#
# Do prediction and then restack into the image
#
################################
if verbose == True:
print("Prediction.")
prediction = unet_models[0].predict(np.transpose(batchX,
axes=(0, 2, 1, 3)),
verbose=verbose)
if number_of_models > 1:
for i in range(1, number_of_models, 1):
prediction += unet_models[i].predict(np.transpose(batchX,
axes=(0, 2, 1,
3)),
verbose=verbose)
prediction /= number_of_models
prediction = np.transpose(prediction, axes=(0, 2, 1, 3))
permutations = list()
permutations.append((0, 1, 2))
permutations.append((1, 0, 2))
permutations.append((1, 2, 0))
prediction_image_average = ants.image_clone(flair_preprocessed_warped) * 0
current_start_slice = 0
for d in range(len(dimensions_to_predict)):
current_end_slice = current_start_slice + flair_preprocessed_warped.shape[
dimensions_to_predict[d]]
which_batch_slices = range(current_start_slice, current_end_slice)
prediction_per_dimension = prediction[which_batch_slices, :, :, :]
prediction_array = np.transpose(np.squeeze(prediction_per_dimension),
permutations[dimensions_to_predict[d]])
prediction_image = ants.copy_image_info(
flair_preprocessed_warped,
pad_or_crop_image_to_size(ants.from_numpy(prediction_array),
flair_preprocessed_warped.shape))
prediction_image_average = prediction_image_average + (
prediction_image - prediction_image_average) / (d + 1)
current_start_slice = current_end_slice
probability_image = ants.apply_ants_transform_to_image(
ants.invert_ants_transform(xfrm), prediction_image_average,
flair_preprocessed)
probability_image = ants.copy_image_info(flair, probability_image)
return (probability_image)
domain_image=mask,
stride_length=stride_length,
domain_image_is_mask=True)
probability_image = ants.iMath_normalize(probability_image)
probability_images_array.append(probability_image)
end_time = time.time()
elapsed_time = end_time - start_time
print(" (elapsed time: ", elapsed_time, " seconds)")
for i in range(number_of_classification_labels - 1):
print("Writing", classes[i])
start_time = time.time()
ants.image_write(
probability_images_array[i],
output_file_name_prefix + classes[i] + "Segmentation.nii.gz")
end_time = time.time()
elapsed_time = end_time - start_time
print(" (elapsed time: ", elapsed_time, " seconds)")
probability_images_matrix = ants.image_list_to_matrix(probability_images_array,
mask)
segmentation_vector = np.argmax(probability_images_matrix, axis=0) + 1
segmentation_image = ants.make_image(mask, segmentation_vector)
ants.image_write(segmentation_image,
output_file_name_prefix + "Segmentation.nii.gz")
end_time_total = time.time()
elapsed_time_total = end_time_total - start_time_total
print("Total elapsed time: ", elapsed_time_total, "seconds")
