def __call__(self, listmode: Listmode) -> Image:
if self.mode == 'numba':
sub_length = int(len(listmode) // self.n_sub)
x = self.emap.update(
data=np.ones(self.emap.shape, dtype=np.float32))
for _ in tqdm(range(self.n_iter)):
index = np.random.randint(0,
high=len(listmode),
size=sub_length)
_listmode = Project(self.mode)(x, listmode.lors[index, :])
_bp = BackProject(self.mode)(listmode[index] / _listmode, x)
x *= _bp / self.emap
return x
elif self.mode == 'tf-eager':
raise NotImplementedError
x = self.emap.update(
data=np.ones(self.emap.shape, dtype=np.float32))
x_tf = x.update(data=tf.constant(x.data))
emap_tf = self.emap.update(data=tf.constant(self.emap.data))
lors_tf = listmode.lors.update(
data=tf.constant(listmode.lors.data))
listmode_tf = listmode.update(data=tf.constant(listmode.data),
lors=lors_tf)
_listmode_tf = Project(self.mode)(x_tf, lors_tf)
_bp = BackProject(self.mode)(listmode_tf / _listmode_tf, emap_tf)
_mlem_iter = tf.assign(x_tf.data, (x_tf * _bp / emap_tf).data)
with tf.compat.v1.Session('') as sess:
sess.run(tf.compat.v1.global_variables_initializer())
for _ in tqdm(range(self.n_iter)):
sess.run(_mlem_iter)
return x_tf.update(data=sess.run(x_tf.data))
elif self.mode == 'tf':
raise NotImplementedError
def __call__(self, listmodes: List(Listmode), dvfs: List(Dvf)) -> Image:
declare_eager_execution()
x = self.emap.update(data=np.ones(self.emap.shape, dtype=np.float32))
x_tf = x.update(data=tf.Variable(x.data))
emap_tf = self.emap.update(data=tf.constant(self.emap.data))
for _ in tqdm(range(self.n_iter)):
for listmode, dvf in zip(listmodes, dvfs):
dvf_x_tf = tf.constant(dvf.dvf_x.data)
dvf_y_tf = tf.constant(dvf.dvf_y.data)
dvf_z_tf = tf.constant(dvf.dvf_z.data)
lors_tf = listmode.lors.update(
data=tf.constant(listmode.lors.data))
listmode_tf = listmode.update(data=tf.constant(listmode.data),
lors=lors_tf)
_x_tf = x_tf.update(data=self._deform_invert_tf(
x_tf.data, dvf_x_tf, dvf_y_tf, dvf_z_tf))
_listmode_tf = Project('tf')(_x_tf, lors_tf)
_listmode_tf = _listmode_tf + 1e-8
_bp = BackProject('tf')(listmode_tf / _listmode_tf, emap_tf)
emap_tf2 = emap_tf.update(data=self._deform_tf(
emap_tf.data, dvf_x_tf, dvf_y_tf, dvf_z_tf)) + 1e-8
_bp2 = _bp.update(data=self._deform_tf(_bp.data, dvf_x_tf,
dvf_y_tf, dvf_z_tf))
x_tf = x_tf * _bp2 / emap_tf2
return x_tf.update(data=x_tf.data.numpy())
def listmode_from_gate_out(path, scanner, nb_sub):
full_data = None
for i in tqdm(range(nb_sub)):
path_ = path.replace('?', str(i))
data = np.load(path_)[:, :7]
if full_data is None:
full_data = np.array(data)
else:
full_data = np.vstack((full_data, data))
if isinstance(scanner, nef.PetEcatScanner):
lors = Lors(full_data)
listmode = Listmode(np.ones((lors.length, ), dtype=np.float32), lors)
row = nef.EcatCrystalPosToIndex(scanner)(lors.data[:, :3])
col = nef.EcatCrystalPosToIndex(scanner)(lors.data[:, 3:])
lors_data1 = nef.EcatIndexToCrystalPos(scanner)(row)
lors_data2 = nef.EcatIndexToCrystalPos(scanner)(col)
lors = np.hstack((lors_data1, lors_data2))
return Listmode(np.ones((lors.shape[0], ), dtype=np.float32), lors)
elif isinstance(scanner, nef.PetCylindricalScanner):
lors = Lors(full_data)
listmode = Listmode(np.ones((lors.length, ), dtype=np.float32), lors)
ind1 = nef.CylindricalCrystalPosToIndex(scanner)(lors.data[:, :3])
ind2 = nef.CylindricalCrystalPosToIndex(scanner)(lors.data[:, 3:])
lors_data1 = nef.CylindricalIndexToCrystalPos(scanner)(ind1)
lors_data2 = nef.CylindricalIndexToCrystalPos(scanner)(ind2)
lors_data = np.hstack((lors_data1, lors_data2))
return listmode.update(lors=nef.Lors(lors_data))
else:
raise NotImplementedError
def __call__(self) -> Image:
declare_eager_execution()
if self.emap is None:
self.generate_emap()
if self.atten_corr is not None:
listmode_ = self.atten_corr(self.listmode)
else:
listmode_ = self.listmode
x_tf = Image(data=tf.Variable(
np.ones(self.image_config.shape, dtype=np.float32)),
center=self.image_config.center,
size=self.image_config.size)
emap_data_n0_zero = copy(self.emap.data)
emap_data_n0_zero[emap_data_n0_zero == 0.0] = 1e8
emap_tf = self.emap.update(data=tf.constant(emap_data_n0_zero))
lors_tf = self.listmode.lors.update(
data=tf.constant(self.listmode.lors.data))
listmode_tf = self.listmode.update(data=tf.constant(listmode_.data),
lors=lors_tf)
for _ in tqdm(range(self.n_iter)):
_listmode_tf = Project('tf-eager')(x_tf, lors_tf)
listmode_div = tf.div_no_nan(listmode_tf.data, _listmode_tf.data)
_bp = BackProject('tf-eager')(
listmode_tf.update(data=listmode_div), emap_tf)
x_tf = x_tf * _bp / emap_tf
x = x_tf.update(data=x_tf.data.numpy())
# if self.scatter_corr is not None:
# if self.atten_corr is not None:
# listmode_ = self.scatter_corr(x, self.atten_corr.u_map, self.scanner, self.listmode)
# x_tf = Image(data = tf.Variable(np.ones(self.image_config.shape, dtype = np.float32)),
# center = self.image_config.center,
# size = self.image_config.size)
# emap_data_n0_zero = copy(self.emap.data)
# emap_data_n0_zero[emap_data_n0_zero == 0.0] = 1e8
# emap_tf = self.emap.update(data = tf.constant(emap_data_n0_zero))
# lors_tf = self.listmode.lors.update(data = tf.constant(self.listmode.lors.data))
# listmode_tf = self.listmode.update(data = tf.constant(listmode_.data), lors = lors_tf)
# for _ in tqdm(range(self.n_iter)):
# _listmode_tf = Project('tf-eager')(x_tf, lors_tf)
# listmode_div = tf.div_no_nan(listmode_tf.data, _listmode_tf.data)
# _bp = BackProject('tf-eager')(listmode_tf.update(data = listmode_div), emap_tf)
# x_tf = x_tf * _bp / emap_tf
# x = x_tf.update(data = x_tf.data.numpy())
if self.psf_corr is not None:
image_ = self.psf_corr(x)
else:
image_ = x
return image_
def listmode_tof_from_gate_out(path, scanner, nb_sub):
full_data = None
for i in tqdm(range(nb_sub)):
path_ = path.replace('?', str(i))
data = np.load(path_)[:, :7]
if full_data is None:
full_data = np.array(data)
else:
full_data = np.vstack((full_data, data))
lors = Lors(full_data)
listmode = Listmode(np.ones((lors.length, ), dtype=np.float32), lors)
row, col = ListmodeToId()(listmode, scanner)
lors = IdToListmode()(row, col, scanner)
print(lors.data.shape)
print(full_data.shape)
lors_tof = lors.update(
data=np.append(lors.data, full_data[:, -1].reshape(-1, 1), axis=1))
return Listmode(np.ones((lors.shape[0], ), dtype=np.float32), lors_tof)
def __call__(self, image: Image):
if self.kernel_xy is None:
raise ValueError('Please do make kernel first')
from srfnef.utils import declare_eager_execution
declare_eager_execution()
x = np.ones((image.shape[0] * image.shape[1], image.shape[2]),
dtype=np.float32)
x_tf = tf.Variable(x)
d = image.data.reshape((-1, image.shape[2]))
d_tf = tf.constant(d)
kernel_xy_tf = tf.sparse.SparseTensor(indices=list(
zip(self.kernel_xy.row, self.kernel_xy.col)),
values=self.kernel_xy.data,
dense_shape=self.kernel_xy.shape)
kernel_z_tf = tf.sparse.SparseTensor(indices=list(
zip(self.kernel_z.row, self.kernel_z.col)),
values=self.kernel_z.data,
dense_shape=self.kernel_z.shape)
if self.kernel_xy.nnz * image.shape[2] > 2**31:
raise ValueError(
'Cannot use GPU when output.shape[1] * nnz(a) > 2^31 [Op:SparseTensorDenseMatMul]'
)
for _ in tqdm(range(self.n_iter)):
c_tf = tf.transpose(
tf.sparse.sparse_dense_matmul(kernel_z_tf,
x_tf,
adjoint_b=True))
c_tf = tf.sparse.sparse_dense_matmul(kernel_xy_tf, c_tf) + 1e-16
c_tf = d_tf / c_tf
c_tf = tf.sparse.sparse_dense_matmul(kernel_xy_tf,
c_tf,
adjoint_a=True)
c_tf = tf.transpose(
tf.sparse.sparse_dense_matmul(kernel_z_tf,
c_tf,
adjoint_a=True,
adjoint_b=True))
x_tf = x_tf * c_tf
image = image.update(data=x_tf.numpy().reshape(image.shape))
return image
def __call__(self, listmode: Listmode) -> Image:
declare_eager_execution()
x_tf = Image(data=tf.Variable(
np.ones(self.emap.shape, dtype=np.float32)),
center=self.emap.center,
size=self.emap.size)
emap_data_n0_zero = copy(self.emap.data)
emap_data_n0_zero[emap_data_n0_zero == 0.0] = 1e8
emap_tf = self.emap.update(data=tf.constant(emap_data_n0_zero))
lors_tf = listmode.lors.update(data=tf.constant(listmode.lors.data))
listmode_tf = listmode.update(data=tf.constant(listmode.data),
lors=lors_tf)
for _ in tqdm(range(self.n_iter)):
_listmode_tf = Project('tf-eager')(x_tf, lors_tf)
_listmode_tf = _listmode_tf + np.mean(listmode.data) * 1e-8
_bp = BackProject('tf-eager')(listmode_tf / _listmode_tf, emap_tf)
x_tf = x_tf * _bp / emap_tf
return x_tf.update(data=x_tf.data.numpy())
def listmode_from_gate_out_multi_bed(path, scanner, nb_sub):
full_data = {}
for i in tqdm(range(nb_sub)):
path_ = path.replace('?', str(i))
data = np.load(path_)[:, :7]
bed_id = np.load(path_)[:, -1]
for i_bed in set(bed_id):
if i_bed not in full_data:
full_data[i_bed] = np.array(data[bed_id == i_bed, :])
else:
full_data[i_bed] = np.vstack(
(full_data[i_bed], data[bed_id == i_bed, :]))
listmode_out = {}
for key, values in full_data.items():
lors = Lors(values)
listmode = Listmode(np.ones((lors.length, ), dtype=np.float32), lors)
row, col = ListmodeToId()(listmode, scanner)
lors = IdToListmode()(row, col, scanner)
listmode_out[key] = Listmode(
np.ones((lors.shape[0], ), dtype=np.float32), lors)
return listmode_out
def __call__(self, image: Image):
from srfnef import EcatIndexToCrystalPos
if self.mode == 'full':
declare_eager_execution()
ind2pos = EcatIndexToCrystalPos(self.scanner)
ind = np.arange(self.scanner.nb_crystals)
pos1 = pos2 = ind2pos(ind)
pos1_ = np.kron(pos1, [1] * pos2.size)
pos2_ = np.kron(pos2, [[1]] * pos1.size).reshape(-1, 3)
lors_data = np.hstack((pos1_, pos2_))
listmode = LorsToListmode()(nef.Lors(lors_data))
return Emap(**BackProject(
mode='tf-eager')(listmode, image).asdict())
elif self.mode == 'block':
declare_eager_execution()
single_block_scanner = self.scanner.update(nb_blocks_per_ring=1)
ind2pos = EcatIndexToCrystalPos(single_block_scanner)
ind = np.arange(self.scanner.nb_crystals_per_block *
self.scanner.nb_rings)
pos1 = pos2 = ind2pos(ind)
pos1_x = np.kron(pos1[:, 0], [1] * ind.size)
pos1_y = np.kron(pos1[:, 1], [1] * ind.size)
pos1_z = np.kron(pos1[:, 2], [1] * ind.size)
pos1_ = np.vstack((pos1_x, pos1_y, pos1_z)).transpose()
emap_data = np.zeros(image.shape, np.float32)
emap_tf = Emap(data=tf.Variable(emap_data),
center=image.center,
size=image.size)
for d in tqdm(range(self.scanner.nb_blocks_per_ring)):
angle = d * self.scanner.angle_per_block
print(angle)
pos2_x = np.kron(pos2[:, 0], [[1]] * ind.size).ravel()
pos2_y = np.kron(pos2[:, 1], [[1]] * ind.size).ravel()
pos2_z = np.kron(pos2[:, 2], [[1]] * ind.size).ravel()
pos2_ = np.vstack(
(pos2_x * np.cos(angle) - pos2_y * np.sin(angle),
pos2_x * np.sin(angle) + pos2_y * np.cos(angle),
pos2_z)).transpose()
lors_data = np.hstack((pos1_, pos2_)).astype(np.float32)
listmode = LorsToListmode()(nef.Lors(lors_data))
listmode_tf = listmode.update(data=tf.Variable(listmode.data),
lors=nef.Lors(
tf.Variable(lors_data)))
_emap = BackProject(mode='tf')(listmode_tf, emap_tf)
for i in range(self.scanner.nb_blocks_per_ring):
_emap_rotate_data = self._rotate_tf(
_emap.data, i * self.scanner.angle_per_block)
tf.compat.v1.assign_add(emap_tf.data, _emap_rotate_data)
emap_data = emap_tf.data.numpy()
return emap_tf.update(data=emap_data,
center=image.center,
size=image.size)
elif self.mode == 'block-full':
declare_eager_execution()
single_block_scanner = self.scanner.update(nb_blocks_per_ring=1)
ind2pos = EcatIndexToCrystalPos(single_block_scanner)
ind = np.arange(self.scanner.nb_crystals_per_block *
self.scanner.nb_rings)
pos1 = pos2 = ind2pos(ind)
emap_data = np.zeros(image.shape, np.float32)
emap_tf = Emap(data=tf.Variable(emap_data),
center=image.center,
size=image.size)
for i in tqdm(range(self.scanner.nb_blocks_per_ring)):
angle1 = i * self.scanner.angle_per_block
pos1_x = np.kron(pos1[:, 0], [1] * ind.size)
pos1_y = np.kron(pos1[:, 1], [1] * ind.size)
pos1_z = np.kron(pos1[:, 2], [1] * ind.size)
pos1_ = np.vstack(
(pos1_x * np.cos(angle1) - pos1_y * np.sin(angle1),
pos1_x * np.sin(angle1) + pos1_y * np.cos(angle1),
pos1_z)).transpose()
for j in range(self.scanner.nb_blocks_per_ring):
angle2 = j * self.scanner.angle_per_block
pos2_x = np.kron(pos2[:, 0], [[1]] * ind.size).ravel()
pos2_y = np.kron(pos2[:, 1], [[1]] * ind.size).ravel()
pos2_z = np.kron(pos2[:, 2], [[1]] * ind.size).ravel()
pos2_ = np.vstack(
(pos2_x * np.cos(angle2) - pos2_y * np.sin(angle2),
pos2_x * np.sin(angle2) + pos2_y * np.cos(angle2),
pos2_z)).transpose()
lors_data = np.hstack((pos1_, pos2_)).astype(np.float32)
listmode = LorsToListmode()(nef.Lors(lors_data))
listmode_tf = listmode.update(
data=tf.Variable(listmode.data),
lors=nef.Lors(tf.Variable(lors_data)))
_emap = BackProject(mode='tf')(listmode_tf, emap_tf)
tf.compat.v1.assign_add(emap_tf.data, _emap.data)
emap_data = emap_tf.data.numpy()
return emap_tf.update(data=emap_data,
center=image.center,
size=image.size)
elif self.mode == 'rsector':
return self.update(mode='block')(image)
elif self.mode == 'rsector-full':
return self.update(mode='block-full')(image)
else:
raise NotImplementedError