def get_random_geometry(dict_of_g):
g = SpatialInfo(
shape=np.random.randint(0, 8, size=[3]),
origin=np.random.uniform(-4, 6, size=[3]),
spacing=np.random.uniform(0.5, 1.5, size=[3]),
)
return g
def test_transform_resampling_random_dependent(self):
"""
Use a functors to randomly generate a resampling geometry.
All volumes MUST use the same geometry.
"""
batch = {
'v1': torch.arange(10 * 9 * 8).reshape((1, 1, 10, 9, 8)),
'v2': torch.arange(10 * 9 * 8).reshape((1, 1, 10, 9, 8)),
'other': 'other_value'
}
get_spatial_info = functools.partial(
TestTransformResample.get_spatial_info_generic,
geometry=SpatialInfo(origin=(0, 0, 0),
spacing=(1, 1, 1),
shape=(10, 9, 8)))
transform = trw.transforms.TransformResample(
get_spatial_info_from_batch_name=get_spatial_info,
resampling_geometry=TestTransformResample.get_random_geometry,
)
batch_transformed = transform(batch)
assert len(batch_transformed) == 3
assert batch_transformed['other'] == 'other_value'
assert batch_transformed['v1'].shape == batch_transformed['v2'].shape
def test_spatial_info_coordinate_mapping(self):
origin = torch.tensor([10, 11, 12], dtype=torch.float32)
spacing = torch.tensor([2, 3, 4], dtype=torch.float32)
pst = affine_transformation_translation(origin).mm(
affine_transformation_scale(spacing))
si = SpatialInfo(shape=[20, 21, 22], patient_scale_transform=pst)
origin_flipped = torch.flip(origin, (0, ))
spacing_flipped = torch.flip(spacing, (0, ))
si_2 = SpatialInfo(shape=[20, 21, 22],
origin=origin_flipped,
spacing=spacing_flipped)
assert (si.patient_scale_transform -
si_2.patient_scale_transform).abs().max() < 1e-5
# index (0, 0, 0) is origin!
p = si.index_to_position(index_zyx=torch.tensor([0, 0, 0]))
assert len(p.shape) == 1
assert p.shape[0] == 3
assert (p - torch.tensor([12, 11, 10],
dtype=torch.float32)).abs().max() < 1e-5
assert (si.position_to_index(position_zyx=p) - torch.tensor(
[0, 0, 0], dtype=torch.float32)).abs().max() < 1e-5
# move in one direction from the origin
for n in range(20):
dp = torch.empty(3).uniform_(0, 10).type(torch.float32)
p = origin_flipped + dp * spacing_flipped
i = si.position_to_index(position_zyx=p)
assert (i - dp).abs().max() < 1e-5
p_back = si.index_to_position(index_zyx=i)
assert (p - p_back).abs().max() < 1e-5
def test_spatial_info_pst(self):
pst = affine_transformation_translation([10, 11, 12]).mm(
affine_transformation_rotation_3d_x(0.3)).mm(
affine_transformation_scale([2, 3, 4]))
si = SpatialInfo(shape=[20, 21, 22], patient_scale_transform=pst)
assert (si.spacing == np.asarray([4, 3, 2])).all()
assert (si.origin == np.asarray([12, 11, 10])).all()
def test_sub_geometry(self):
pst = affine_transformation_translation([10, 11, 12]).mm(
affine_transformation_rotation_3d_x(0.3)).mm(
affine_transformation_scale([2, 3, 4]))
si = SpatialInfo(shape=[20, 21, 22], patient_scale_transform=pst)
start_zyx = torch.tensor([5, 6, 7])
end_zyx = torch.tensor([8, 10, 13])
si_sub = si.sub_geometry(start_index_zyx=start_zyx,
end_index_zyx_inclusive=end_zyx)
o = si_sub.index_to_position(index_zyx=torch.tensor([0, 0, 0]))
expected_o = si.index_to_position(index_zyx=start_zyx)
assert (o - expected_o).abs().max() == 0
index_e = end_zyx - start_zyx
pos_e = si_sub.index_to_position(index_zyx=index_e)
pos_e_expected = si.index_to_position(index_zyx=end_zyx)
assert (pos_e - pos_e_expected).abs().max() <= 1e-5
def test_transform_background_volume_dependent(self):
batch = {
'v1': torch.arange(10 * 9 * 8).reshape((1, 1, 10, 9, 8)),
'v2': torch.arange(10 * 9 * 8).reshape((1, 1, 10, 9, 8)),
'other': 'other_value'
}
get_spatial_info = functools.partial(
TestTransformResample.get_spatial_info_generic,
geometry=SpatialInfo(origin=(0, 0, 0),
spacing=(1, 1, 1),
shape=(10, 9, 8)))
transform = trw.transforms.TransformResample(
get_spatial_info_from_batch_name=get_spatial_info,
resampling_geometry=SpatialInfo(origin=(0, 0, 0),
spacing=(1, 1, 1),
shape=(1, 9, 18)),
constant_background_value={
'v1': 42,
'v2': 43
})
batch_transformed = transform(batch)
assert len(batch_transformed) == 3
assert batch_transformed['other'] == 'other_value'
assert batch_transformed['v1'].shape == (1, 1, 1, 9, 18)
assert batch_transformed['v2'].shape == (1, 1, 1, 9, 18)
# voxel within FoV of the volumes
assert (batch_transformed['v1'][0,
0, :, :, :8] == batch['v1'][0, 0,
0:1, :, :8]
).all()
assert (batch_transformed['v2'][0,
0, :, :, :8] == batch['v2'][0, 0,
0:1, :, :8]
).all()
# voxels in the background
assert (batch_transformed['v1'][0, 0, 0:1, :, 8:] == 42).all()
assert (batch_transformed['v2'][0, 0, 0:1, :, 8:] == 43).all()
def get_spatial_info_type(batch: Batch, name: str) -> SpatialInfo:
v = batch[name]
assert len(v.shape) == 5
assert v.shape[0] == 1
assert v.shape[1] == 1
affine_matrix = batch[name.replace('_voxels', '_affine')]
assert len(affine_matrix.shape) == 3 and affine_matrix.shape[0] == 1
affine_matrix = affine_matrix[0]
spacing = get_spacing_from_4x4(affine_matrix)
origin = get_translation_from_4x4(affine_matrix)
return SpatialInfo(origin=origin, spacing=spacing, shape=v.shape[2:])
def test_transform_resampling_random_fixed(self):
batch = {
'v1': torch.arange(10 * 9 * 8).reshape((1, 1, 10, 9, 8)),
'v2': torch.arange(1 * 1 * 1).reshape((1, 1, 1, 1, 1)),
}
get_spatial_info = functools.partial(
TestTransformResample.get_spatial_info_generic,
geometry=SpatialInfo(origin=(0, 0, 0),
spacing=(1, 1, 1),
shape=(10, 9, 8)))
transform = trw.transforms.TransformResample(
get_spatial_info_from_batch_name=get_spatial_info,
resampling_geometry=functools.partial(
random_fixed_geometry_within_geometries,
fixed_geometry_shape=(3, 3, 3),
fixed_geometry_spacing=(1, 1, 1)),
)
batch_transformed = transform(batch)
assert len(batch_transformed) == 2
assert batch_transformed['v1'].shape == batch_transformed['v2'].shape
def test_random_volumes(self):
def fill_volume(v, nb):
for n in range(nb):
min_index = torch.tensor([
torch.randint(0, v.shape[0] - 5, size=(1, )),
torch.randint(0, v.shape[1] - 5, size=(1, )),
torch.randint(0, v.shape[2] - 5, size=(1, )),
])
shape = torch.randint(10, 15, size=(3, ))
value = torch.randint(255, size=(1, ))
max_index = min_index + shape
max_index = trw.utils.clamp_n(max_index,
torch.tensor([0, 0, 0]),
torch.tensor(v.shape)).squeeze(0)
trw.utils.sub_tensor(v, min_index, max_index)[:] = value
np.random.seed(0)
torch.manual_seed(0)
nb_points = 10000
all_avg_errors = []
for e in range(5):
tfm = torch.from_numpy(
np.asarray([
[0.95, 0, 0, -0.5],
[0, 0.9, 0, -5],
[0, 0, 1, 1],
[0, 0, 0, 1],
],
dtype=np.float32))
shape_moving_zyx = (42, 30, 40)
shape_fixed_zyx = (37, 31, 42)
moving_geometry = SpatialInfo(shape=shape_moving_zyx,
origin=(10, 15, 20),
spacing=(0.9, 1.1, 1.3))
fixed_geometry = SpatialInfo(shape=shape_fixed_zyx,
origin=(15, 5, 15),
spacing=(1.3, 1.4, 1.1))
moving = torch.zeros(shape_moving_zyx,
dtype=torch.float32).unsqueeze(0).unsqueeze(0)
fill_volume(moving[0, 0], 5)
fixed = resample_spatial_info(geometry_moving=moving_geometry,
moving_volume=moving,
geometry_fixed=fixed_geometry,
tfm=tfm)
error = 0.0
nb_voxels = 0
shape_moving_zyx = torch.tensor(shape_fixed_zyx)
for _ in range(nb_points):
index_fixed = torch.tensor([
torch.randint(shape_fixed_zyx[0], size=(1, )),
torch.randint(shape_fixed_zyx[1], size=(1, )),
torch.randint(shape_fixed_zyx[2], size=(1, ))
],
dtype=torch.float32)
index_fixed_i = index_fixed.type(torch.long)
# transform: index (fixed) -> position in world space
p = fixed_geometry.index_to_position(index_zyx=index_fixed)
# apply moving transform
p = apply_homogeneous_affine_transform_zyx(tfm, p)
# transform position -> index (moving)
index_moving = moving_geometry.position_to_index(
position_zyx=p)
index_moving_rounded = index_moving.round().type(torch.long)
if (index_moving_rounded >= 0).all() and \
((index_moving_rounded - shape_moving_zyx) < 0).all():
value_resampled = fixed[0, 0][index_fixed_i[0],
index_fixed_i[1],
index_fixed_i[2]]
if value_resampled <= 2:
# discard background values
continue
value_fixed = moving[0, 0][index_moving_rounded[0],
index_moving_rounded[1],
index_moving_rounded[2]]
error += (value_resampled - value_fixed).abs()
nb_voxels += 1
avg_error = error / nb_voxels
print('avg=', error / nb_voxels, 'nb_voxels=', nb_voxels)
# empirical value
assert avg_error <= 25.0
all_avg_errors.append(avg_error)
assert np.mean(all_avg_errors) < 15