def _contruct_ilang_model(self):
self.ilang_model = ModelRigidSSD('RigidSSD')
self.graph = ProbabilisticGraphicalModel(['target','source','transformation'])
self.graph.set_nodes_given(['target','source'],True)
self.graph.set_node_value('source',self.source.get_data())
self.graph.set_node_value('target',self.target.get_data())
self.graph.set_node_value('transformation',numpy.zeros((1,6)))
self.graph.add_dependence(self.ilang_model,{'target':'target','source':'source','transformation':'transformation'})
self.sampler = Sampler(self.graph)
def __make_graph(self):
self.ilang_model = SSD_ilang()
# The following is an alternative use of ilang models: define the log_p and gradient functions here in order to have access to
# the local variables.
self.ilang_model.log_conditional_probability_transformation = self.__P
self.ilang_model.log_conditional_probability_gradient_transformation = self.__G
self.ilang_graph = ProbabilisticGraphicalModel(
['source', 'target', 'transformation', 'sigma'])
self.ilang_graph.set_nodes_given(['sigma', 'source', 'target'], True)
self.ilang_graph.add_dependence(
self.ilang_model, {
'source': 'source',
'target': 'target',
'sigma': 'sigma',
'transformation': 'transformation'
})
self.ilang_sampler = Sampler(self.ilang_graph)
def _contruct_ilang_model(self):
self.ilang_model = ModelRigidSSD('RigidSSD')
self.graph = ProbabilisticGraphicalModel(['target','source','transformation'])
self.graph.set_nodes_given(['target','source'],True)
self.graph.set_node_value('source',self.source.get_data())
self.graph.set_node_value('target',self.target.get_data())
self.graph.set_node_value('transformation',numpy.zeros((1,6)))
self.graph.add_dependence(self.ilang_model,{'target':'target','source':'source','transformation':'transformation'})
self.sampler = Sampler(self.graph)
def __make_graph(self):
self.ilang_model = SSD_ilang()
# The following is an alternative use of ilang models: define the log_p and gradient functions here in order to have access to
# the local variables.
self.ilang_model.log_conditional_probability_transformation = self.__P
self.ilang_model.log_conditional_probability_gradient_transformation = self.__G
self.ilang_graph = ProbabilisticGraphicalModel(["source", "target", "transformation", "sigma"])
self.ilang_graph.set_nodes_given(["sigma", "source", "target"], True)
self.ilang_graph.add_dependence(
self.ilang_model,
{"source": "source", "target": "target", "sigma": "sigma", "transformation": "transformation"},
)
self.ilang_sampler = Sampler(self.ilang_graph)
def __make_graph(self):
self.ilang_graph = ProbabilisticGraphicalModel()
for i in range(len(self.__images)):
self.ilang_graph.add_node('im_%d' % i)
self.ilang_graph.set_nodes_given([
'im_%d' % i,
], True)
for i in range(len(self.__images) - 1):
self.ilang_graph.add_nodes(['sigma_%d' % i, 'T_%d' % i])
self.ilang_graph.set_nodes_given([
'sigma_%d' % i,
], True)
model = SSD_ilang('SSD_%d' % i)
self.ilang_graph.add_dependence(
model, {
'sigma': 'sigma_%d' % i,
'transformation': 'T_%d' % i,
'source': 'im_%d' % i,
'target': 'im_%d' % (i + 1)
})
self.ilang_sampler = Sampler(self.ilang_graph)
class Registration_Two_Images(object):
def __init__(self,
source=None,
target=None,
degrees_of_freedom=3,
sigma=DEFAULT_SIGMA,
initial_transformation=None):
self.set_source(source)
self.set_target(target)
self.set_sigma(sigma)
self.__make_graph()
if initial_transformation == None:
self.set_transformation(numpy.zeros(degrees_of_freedom))
else:
self.set_transformation(initial_transformation)
def __make_graph(self):
self.ilang_model = SSD_ilang()
# The following is an alternative use of ilang models: define the log_p and gradient functions here in order to have access to
# the local variables.
self.ilang_model.log_conditional_probability_transformation = self.__P
self.ilang_model.log_conditional_probability_gradient_transformation = self.__G
self.ilang_graph = ProbabilisticGraphicalModel(
['source', 'target', 'transformation', 'sigma'])
self.ilang_graph.set_nodes_given(['sigma', 'source', 'target'], True)
self.ilang_graph.add_dependence(
self.ilang_model, {
'source': 'source',
'target': 'target',
'sigma': 'sigma',
'transformation': 'transformation'
})
self.ilang_sampler = Sampler(self.ilang_graph)
def __set_space(self):
# make sure that the images are in the same space
if self.source.space != self.target.space:
raise "Source [%s] and Target [%s] must be in the same space" % (
source.space, target.space)
self.space = self.source.space
def set_transformation(self, tr):
self.__transformation = tr
def __get_transformation(self):
return self.__transformation
def set_cost_function(self, cost):
pass
def __initialize_registration(self,
optimization_method='QuasiNewton_L_BFGS_B',
n_points=DEFAULT_N_POINTS):
self.ilang_graph.set_node_value('source', self.source.data)
self.ilang_graph.set_node_value('target', self.target.data)
self.ilang_graph.set_node_value('sigma', self.sigma)
if n_points == None:
n_points = self.target.shape
self.grid = self.target.get_world_grid(n_points)
self.resampled_target = self.target.compute_resample_on_grid(self.grid)
self.ilang_sampler.set_node_sampling_method_manual(
'transformation', optimization_method)
def register(self,
optimization_method='GradientAscent',
iterations=DEFAULT_N_ITER,
n_points=DEFAULT_N_POINTS):
self.__initialize_registration(optimization_method, n_points)
self.ilang_graph.set_node_value('transformation', self.transformation)
self.ilang_sampler.sample(iterations)
self.transformation = self.ilang_graph.get_node_value('transformation')
def set_source(self, source):
self.source = source
if hasattr(self, 'target'):
self.__set_space()
def set_target(self, target):
self.target = target
if hasattr(self, 'source'):
self.__set_space()
def set_sigma(self, sigma):
self.sigma = sigma
def get_result(self):
result = self.source.copy()
result.transform(self.get_transformation_matrix())
return result
def get_transformation_matrix(self):
#rot = self.__transformation[0:3]
#tra = self.__transformation[3:6]
tra = self.__transformation
return Transform_Translation(
tra, self.space, self.space) #FIXME: rotation and translation
def display(self):
D = MultipleVolumesNiftyCore(
[self.source, self.target,
self.get_result()])
#D = MultipleVolumes([self.source, self.target])
return D
def __G(self, transformation):
# print "Transformation: ", transformation
#rot = transformation[0:3]
#tra = transformation[3:6]
tra = transformation
T = Transform_Translation(tra, self.space,
self.space) #FIXME: rotation and translation
re_s = self.source.compute_resample_on_grid(
self.grid, affine_grid_to_world=T) #FIXME: memoize
gr_s = self.source.compute_gradient_on_grid(self.grid,
affine_grid_to_world=T)
re_t = self.resampled_target
G_tra0 = (re_t.data - re_t.data * gr_s[0]).sum()
G_tra1 = (re_t.data - re_t.data * gr_s[1]).sum()
G_tra2 = (re_t.data - re_t.data * gr_s[2]).sum()
G_tra = [G_tra0, G_tra1, G_tra2]
G = numpy.asarray([G_tra[0], G_tra[1], G_tra[2]]) / self.sigma
# print "gradient: ", G
# print "log_likelihood: ", self.__P(transformation)
return G
def __P(self, transformation):
# FIXME: memoize (now image is resampled to compute log_p and the gradient)
#print transformation
#rot = transformation[0:3]
#tra = transformation[3:6]
tra = transformation
T = Transform_Translation(tra, self.space,
self.space) #FIXME: rotation and translation
resampled_source = self.source.compute_resample_on_grid(
self.grid, affine_grid_to_world=T)
P = -numpy.linalg.norm(resampled_source.data - self.resampled_target.
data) / self.sigma #FIXME: verify
# print "log_likelihood: ", P
return P
def _repr_html_(self):
#return self.ilang_graph._repr_html_()
return self.display()._repr_html_()
def __repr__(self):
return "Registration of 2 images."
transformation = property(__get_transformation, set_transformation)
class RigidTransformationSSD():
def __init__(self,target,source):
self.target = target
self.source = source
self._resampled_source_cache = None
self._resampled_source_need_update = True
self._contruct_ilang_model()
self._set_transformation(numpy.zeros((1,6)))
def _contruct_ilang_model(self):
self.ilang_model = ModelRigidSSD('RigidSSD')
self.graph = ProbabilisticGraphicalModel(['target','source','transformation'])
self.graph.set_nodes_given(['target','source'],True)
self.graph.set_node_value('source',self.source.get_data())
self.graph.set_node_value('target',self.target.get_data())
self.graph.set_node_value('transformation',numpy.zeros((1,6)))
self.graph.add_dependence(self.ilang_model,{'target':'target','source':'source','transformation':'transformation'})
self.sampler = Sampler(self.graph)
def _set_transformation(self,transformation):
self.graph.set_node_value('transformation',transformation)
self._resampled_source_need_update = True
def _get_transformation(self):
return self.graph.get_node_value('transformation')
def estimate_transformation(self,method=None,iterations=30000,trace=True,parameters=None, display_progress=True):
if method!=None:
self.sampler.set_node_sampling_method_manual('transformation',method)
else:
self.sampler.set_node_sampling_method_auto('transformation')
last_sample = self.sampler.sample_node('transformation',nsamples=iterations,trace=trace)
# This is not necessary, but it triggers the cache of the resampled source image:
self._set_transformation(last_sample)
return last_sample
def resample_in_target_space(self,image):
# FIXME
#print "Resampling .."
transformation = self._get_transformation()
return image
def _get_resampled_source(self):
#FIXME: remove the next few lines
if self.transformation.sum() != 0:
return self.target
if self._resampled_source_need_update:
resampled_source = self.resample_in_target_space(self.source)
self._resampled_source_cache = resampled_source
self._resampled_source_need_update = False
else:
resampled_source = self._resampled_source_cache
return resampled_source
def display_in_browser(self,axis=0,shrink=256,rotate=90,subsample_slices=4,scale_factors=None):
self.display(axis,shrink,rotate,scale_factors,open_browser=True)
def display(self,axis=0,shrink=256,rotate=90,subsample_slices=4,scale_factors=None,open_browser=False):
resampled = self.resampled_source
D = MultipleVolumes([resampled,self.target], axis, shrink, rotate, subsample_slices, scale_factors, open_browser)
return D.display()
def _repr_html_(self):
return self.display()._repr_html_()
transformation = property(_get_transformation, _set_transformation, None)
resampled_source = property(_get_resampled_source, None, None)
# Define the model components
observation = Poisson('SPECT')
prior_activity = Smoothness('Smoothing_Activity')
prior_attenuation = Smoothness('Smoothing_Attenuation')
# Build the graph
dag = ProbabilisticGraphicalModel(['activity','attenuation','counts','smoothing-activity','smoothing-attenuation'])
dag.set_nodes_given(['counts','smoothing-activity','smoothing-attenuation'], True)
dag.add_dependence(observation,{'lambda':'activity','alpha':'attenuation','z':'counts'})
dag.add_dependence(prior_activity,{'x':'activity','beta':'smoothing-activity'})
dag.add_dependence(prior_attenuation,{'x':'attenuation','beta':'smoothing-attenuation'})
# Initialize the nodes of the graph
dag.set_node_value('activity',numpy.ones((10,10)))
# Instantiate the sampler, tracer and display
sampler = Sampler(dag)
tracer = RamTracer(sampler)
display = MatplotlibDisplay(tracer)
# Sample
sampler.sample(1000,trace=False)
display.imagesc_node('activity')
if __name__=="__main__":
dag.webdisplay(background=False)
class RigidTransformationSSD():
def __init__(self, target, source):
self.target = target
self.source = source
self._resampled_source_cache = None
self._resampled_source_need_update = True
self._contruct_ilang_model()
self._set_transformation(numpy.zeros((1, 6)))
def _contruct_ilang_model(self):
self.ilang_model = ModelRigidSSD('RigidSSD')
self.graph = ProbabilisticGraphicalModel(
['target', 'source', 'transformation'])
self.graph.set_nodes_given(['target', 'source'], True)
self.graph.set_node_value('source', self.source.get_data())
self.graph.set_node_value('target', self.target.get_data())
self.graph.set_node_value('transformation', numpy.zeros((1, 6)))
self.graph.add_dependence(
self.ilang_model, {
'target': 'target',
'source': 'source',
'transformation': 'transformation'
})
self.sampler = Sampler(self.graph)
def _set_transformation(self, transformation):
self.graph.set_node_value('transformation', transformation)
self._resampled_source_need_update = True
def _get_transformation(self):
return self.graph.get_node_value('transformation')
def estimate_transformation(self,
method=None,
iterations=30000,
trace=True,
parameters=None,
display_progress=True):
if method != None:
self.sampler.set_node_sampling_method_manual(
'transformation', method)
else:
self.sampler.set_node_sampling_method_auto('transformation')
last_sample = self.sampler.sample_node('transformation',
nsamples=iterations,
trace=trace)
# This is not necessary, but it triggers the cache of the resampled source image:
self._set_transformation(last_sample)
return last_sample
def resample_in_target_space(self, image):
# FIXME
#print "Resampling .."
transformation = self._get_transformation()
return image
def _get_resampled_source(self):
#FIXME: remove the next few lines
if self.transformation.sum() != 0:
return self.target
if self._resampled_source_need_update:
resampled_source = self.resample_in_target_space(self.source)
self._resampled_source_cache = resampled_source
self._resampled_source_need_update = False
else:
resampled_source = self._resampled_source_cache
return resampled_source
def display_in_browser(self,
axis=0,
shrink=256,
rotate=90,
subsample_slices=4,
scale_factors=None):
self.display(axis, shrink, rotate, scale_factors, open_browser=True)
def display(self,
axis=0,
shrink=256,
rotate=90,
subsample_slices=4,
scale_factors=None,
open_browser=False):
resampled = self.resampled_source
D = MultipleVolumes([resampled, self.target], axis, shrink, rotate,
subsample_slices, scale_factors, open_browser)
return D.display()
def _repr_html_(self):
return self.display()._repr_html_()
transformation = property(_get_transformation, _set_transformation, None)
resampled_source = property(_get_resampled_source, None, None)
class Registration_Two_Images(object):
def __init__(
self, source=None, target=None, degrees_of_freedom=3, sigma=DEFAULT_SIGMA, initial_transformation=None
):
self.set_source(source)
self.set_target(target)
self.set_sigma(sigma)
self.__make_graph()
if initial_transformation == None:
self.set_transformation(numpy.zeros(degrees_of_freedom))
else:
self.set_transformation(initial_transformation)
def __make_graph(self):
self.ilang_model = SSD_ilang()
# The following is an alternative use of ilang models: define the log_p and gradient functions here in order to have access to
# the local variables.
self.ilang_model.log_conditional_probability_transformation = self.__P
self.ilang_model.log_conditional_probability_gradient_transformation = self.__G
self.ilang_graph = ProbabilisticGraphicalModel(["source", "target", "transformation", "sigma"])
self.ilang_graph.set_nodes_given(["sigma", "source", "target"], True)
self.ilang_graph.add_dependence(
self.ilang_model,
{"source": "source", "target": "target", "sigma": "sigma", "transformation": "transformation"},
)
self.ilang_sampler = Sampler(self.ilang_graph)
def __set_space(self):
# make sure that the images are in the same space
if self.source.space != self.target.space:
raise "Source [%s] and Target [%s] must be in the same space" % (source.space, target.space)
self.space = self.source.space
def set_transformation(self, tr):
self.__transformation = tr
def __get_transformation(self):
return self.__transformation
def set_cost_function(self, cost):
pass
def __initialize_registration(self, optimization_method="QuasiNewton_L_BFGS_B", n_points=DEFAULT_N_POINTS):
self.ilang_graph.set_node_value("source", self.source.data)
self.ilang_graph.set_node_value("target", self.target.data)
self.ilang_graph.set_node_value("sigma", self.sigma)
if n_points == None:
n_points = self.target.shape
self.grid = self.target.get_world_grid(n_points)
self.resampled_target = self.target.compute_resample_on_grid(self.grid)
self.ilang_sampler.set_node_sampling_method_manual("transformation", optimization_method)
def register(self, optimization_method="GradientAscent", iterations=DEFAULT_N_ITER, n_points=DEFAULT_N_POINTS):
self.__initialize_registration(optimization_method, n_points)
self.ilang_graph.set_node_value("transformation", self.transformation)
self.ilang_sampler.sample(iterations)
self.transformation = self.ilang_graph.get_node_value("transformation")
def set_source(self, source):
self.source = source
if hasattr(self, "target"):
self.__set_space()
def set_target(self, target):
self.target = target
if hasattr(self, "source"):
self.__set_space()
def set_sigma(self, sigma):
self.sigma = sigma
def get_result(self):
result = self.source.copy()
result.transform(self.get_transformation_matrix())
return result
def get_transformation_matrix(self):
# rot = self.__transformation[0:3]
# tra = self.__transformation[3:6]
tra = self.__transformation
return Transform_Translation(tra, self.space, self.space) # FIXME: rotation and translation
def display(self):
D = MultipleVolumesNiftyPy([self.source, self.target, self.get_result()])
# D = MultipleVolumes([self.source, self.target])
return D
def __G(self, transformation):
# print "Transformation: ", transformation
# rot = transformation[0:3]
# tra = transformation[3:6]
tra = transformation
T = Transform_Translation(tra, self.space, self.space) # FIXME: rotation and translation
re_s = self.source.compute_resample_on_grid(self.grid, affine_grid_to_world=T) # FIXME: memoize
gr_s = self.source.compute_gradient_on_grid(self.grid, affine_grid_to_world=T)
re_t = self.resampled_target
G_tra0 = (re_t.data - re_t.data * gr_s[0]).sum()
G_tra1 = (re_t.data - re_t.data * gr_s[1]).sum()
G_tra2 = (re_t.data - re_t.data * gr_s[2]).sum()
G_tra = [G_tra0, G_tra1, G_tra2]
G = numpy.asarray([G_tra[0], G_tra[1], G_tra[2]]) / self.sigma
# print "gradient: ", G
# print "log_likelihood: ", self.__P(transformation)
return G
def __P(self, transformation):
# FIXME: memoize (now image is resampled to compute log_p and the gradient)
# print transformation
# rot = transformation[0:3]
# tra = transformation[3:6]
tra = transformation
T = Transform_Translation(tra, self.space, self.space) # FIXME: rotation and translation
resampled_source = self.source.compute_resample_on_grid(self.grid, affine_grid_to_world=T)
P = -numpy.linalg.norm(resampled_source.data - self.resampled_target.data) / self.sigma # FIXME: verify
# print "log_likelihood: ", P
return P
def _repr_html_(self):
# return self.ilang_graph._repr_html_()
return self.display()._repr_html_()
def __repr__(self):
return "Registration of 2 images."
transformation = property(__get_transformation, set_transformation)
