def data_generators(self, image_post_processing_np):
"""
Returns the data generators that process one input. See datasources() for dict values.
:param image_post_processing_np: The np postprocessing function for the image data generator.
:return: A dict of data generators.
"""
image_generator = ImageGenerator(
self.dim,
self.image_size,
post_processing_np=image_post_processing_np,
interpolator='linear',
resample_default_pixel_value=0,
data_format=self.data_format,
resample_sitk_pixel_type=sitk.sitkFloat32,
np_pixel_type=np.float32)
if self.downsampling_factor == 1:
heatmap_post_transformation = None
else:
heatmap_post_transformation = scale.Fixed(self.dim,
self.downsampling_factor)
landmark_generator = LandmarkGeneratorHeatmap(
self.dim,
self.heatmap_size, [1] * self.dim,
self.sigma,
scale_factor=1.0,
normalize_center=True,
data_format=self.data_format,
post_transformation=heatmap_post_transformation)
return {'image': image_generator, 'landmarks': landmark_generator}
def spatial_transformation_volumetric_augmented(self, image):
"""
The spatial image transformation with random augmentation.
:return: The transformation.
"""
dim = 3
transformations_list = [translation.InputCenterToOrigin(dim),
scale.FitFixedAr(dim, self.image_size + [None], ignore_dim=[2])]
if self.scale_factor[0] == 1.0 and self.scale_factor[1] == 1.0:
# if no scale_factor, randomly shift by certain value
transformations_list.append(translation.Random(dim, [self.image_size[0] * 0.35, self.image_size[1] * 0.35, 0.0]))
else:
# else, randomly move in imag size
move_factor = [(1.0 - self.scale_factor[i]) * 0.5 for i in [0, 1]]
transformations_list.append(translation.Random(dim, [self.image_size[0] * move_factor[0], self.image_size[1] * move_factor[1], 0]))
transformations_list.append(scale.Fixed(dim, self.scale_factor + [1.0]))
transformations_list.extend([flip.Random(dim, [0.5, 0.5, 0.0]),
rotation.Random(dim, [0., 0., math.pi]),
scale.RandomUniform(dim, 0.25, ignore_dim=[2]),
scale.Random(dim, [0.25, 0.25, 0.0]),
translation.OriginToOutputCenter(dim, self.image_size + [self.num_frames]),
deformation.Output(dim, [6, 6, 4], [10, 10, 0], self.image_size + [self.num_frames])])
comp = composite.Composite(dim, transformations_list,
name='image_transformation_comp',
kwparents={'image': image})
return LambdaNode(lambda comp: sitk.DisplacementFieldTransform(sitk.TransformToDisplacementField(comp, sitk.sitkVectorFloat64, size=self.image_size + [self.num_frames])),
name='image_transformation',
kwparents={'comp': comp})
def spatial_transformation_augmented(self, image):
"""
The spatial image transformation with random augmentation.
:return: The transformation.
"""
# bring image to center and fit to AR
transformations_list = [translation.InputCenterToOrigin(self.dim),
scale.FitFixedAr(self.dim, self.image_size)]
if self.scale_factor[0] == 1.0 and self.scale_factor[1] == 1.0:
# if no scale_factor, randomly shift by certain value
transformations_list.append(translation.Random(self.dim, [self.image_size[0] * 0.35, self.image_size[1] * 0.35]))
else:
# else, randomly move in imag size
move_factor = [(1.0 - self.scale_factor[i]) * 0.5 for i in [0, 1]]
transformations_list.append(translation.Random(self.dim, [self.image_size[0] * move_factor[0], self.image_size[1] * move_factor[1]]))
transformations_list.append(scale.Fixed(self.dim, self.scale_factor))
transformations_list.extend([flip.Random(self.dim, [0.5, 0.5]),
rotation.Random(self.dim, [math.pi]),
scale.RandomUniform(self.dim, 0.25),
scale.Random(self.dim, [0.25, 0.25]),
translation.OriginToOutputCenter(self.dim, self.image_size),
deformation.Output(self.dim, [8, 8], 10, self.image_size)])
comp = composite.Composite(self.dim, transformations_list,
name='image_transformation_comp',
kwparents={'image': image})
return LambdaNode(lambda comp: sitk.DisplacementFieldTransform(sitk.TransformToDisplacementField(comp, sitk.sitkVectorFloat64, size=self.image_size)),
name='image_transformation',
kwparents={'comp': comp})
def image_transformation_val(self):
return composite.Composite(self.dim, [
translation.InputCenterToOrigin(self.dim),
scale.Fit(self.dim, self.image_size),
scale.Fixed(self.dim, self.additional_scale),
translation.OriginToOutputCenter(self.dim, self.image_size)
])
def spatial_transformation(self, image):
"""
The spatial image transformation without random augmentation.
:return: The transformation.
"""
return composite.Composite(self.dim,
[translation.InputCenterToOrigin(self.dim),
scale.Fit(self.dim, self.image_size),
scale.Fixed(self.dim, self.additional_scale),
translation.OriginToOutputCenter(self.dim, self.image_size)],
name='image',
kwparents={'image': image})
def data_generators(self, data_sources, transformation,
image_post_processing_np):
"""
Returns the data generators that process one input. See datasources() for dict values.
:param data_sources: The data_sources dictionary.
:param transformation: The used transformation.
:param image_post_processing_np: The np postprocessing function for the image data generator.
:return: A dict of data generators.
"""
image_generator = ImageGenerator(
self.dim,
self.image_size,
self.image_spacing,
post_processing_np=image_post_processing_np,
interpolator='linear',
resample_default_pixel_value=0,
data_format=self.data_format,
resample_sitk_pixel_type=sitk.sitkFloat32,
np_pixel_type=np.float32,
parents=[data_sources['image_datasource'], transformation],
name='image')
if self.downsampling_factor == 1:
heatmap_post_transformation = None
else:
heatmap_post_transformation = scale.Fixed(self.dim,
self.downsampling_factor)
landmark_generator = LandmarkGenerator(
self.dim,
self.heatmap_size,
self.image_spacing,
#sigma=1.0,
#scale_factor=1.0,
#normalize_center=True,
data_format=self.data_format,
post_transformation=heatmap_post_transformation,
min_max_transformation_distance=30,
parents=[data_sources['landmark_datasource'], transformation],
name='landmarks')
return {'image': image_generator, 'landmarks': landmark_generator}