def test_ssd_2d_gauss_newton():
r""" Test 2D SyN with SSD metric, Gauss-Newton optimizer
Classical Circle-To-C experiment for 2D monomodal registration. We
verify that the final registration is of good quality.
"""
fname_moving = get_fnames('reg_o')
fname_static = get_fnames('reg_c')
moving = np.load(fname_moving)
static = np.load(fname_static)
moving = np.array(moving, dtype=floating)
static = np.array(static, dtype=floating)
moving = (moving - moving.min()) / (moving.max() - moving.min())
static = (static - static.min()) / (static.max() - static.min())
# Create the SSD metric
smooth = 4
inner_iter = 5
step_type = 'gauss_newton'
similarity_metric = metrics.SSDMetric(2, smooth, inner_iter, step_type)
# Configure and run the Optimizer
level_iters = [200, 100, 50, 25]
step_length = 0.5
opt_tol = 1e-4
inv_iter = 40
inv_tol = 1e-3
ss_sigma_factor = 0.2
optimizer = imwarp.SymmetricDiffeomorphicRegistration(
similarity_metric, level_iters, step_length, ss_sigma_factor, opt_tol,
inv_iter, inv_tol)
# test callback not being called
optimizer.INIT_START_CALLED = 0
optimizer.INIT_END_CALLED = 0
optimizer.OPT_START_CALLED = 0
optimizer.OPT_END_CALLED = 0
optimizer.SCALE_START_CALLED = 0
optimizer.SCALE_END_CALLED = 0
optimizer.ITER_START_CALLED = 0
optimizer.ITER_END_CALLED = 0
optimizer.verbosity = VerbosityLevels.DEBUG
id = np.eye(3)
mapping = optimizer.optimize(static, moving, id, id, id)
m = optimizer.get_map()
assert_equal(mapping, m)
warped = mapping.transform(moving)
starting_energy = np.sum((static - moving)**2)
final_energy = np.sum((static - warped)**2)
reduced = 1.0 - final_energy / starting_energy
assert (reduced > 0.9)
assert_equal(optimizer.OPT_START_CALLED, 0)
assert_equal(optimizer.OPT_END_CALLED, 0)
assert_equal(optimizer.SCALE_START_CALLED, 0)
assert_equal(optimizer.SCALE_END_CALLED, 0)
assert_equal(optimizer.ITER_START_CALLED, 0)
assert_equal(optimizer.ITER_END_CALLED, 0)
def test_ssd_3d_gauss_newton():
r'''
Register a stack of circles ('cylinder') before and after warping them with
a synthetic diffeomorphism. This test is intended to detect regressions
only: we saved the energy profile (the sequence of energy values at each
iteration) of a working version of SSD in 3D using the Gauss-Newton step,
and this test checks that the current energy profile matches the saved
one. The validation of the "working version" was
done by registering the 18 manually annotated T1 brain MRI database IBSR
with each other and computing the jaccard index for all 31 common
anatomical regions.
'''
moving, static = get_synthetic_warped_circle(35)
moving[...,:10] = 0
moving[...,-1:-11:-1] = 0
static[...,:10] = 0
static[...,-1:-11:-1] = 0
#Create the SSD metric
smooth = 4
inner_iter = 5
step_type = 'gauss_newton'
similarity_metric = metrics.SSDMetric(3, smooth, inner_iter, step_type)
#Create the optimizer
level_iters = [10, 5]
step_length = 0.25
opt_tol = 1e-4
inv_iter = 20
inv_tol = 1e-3
ss_sigma_factor = 0.5
optimizer = imwarp.SymmetricDiffeomorphicRegistration(similarity_metric,
level_iters, step_length, ss_sigma_factor, opt_tol, inv_iter, inv_tol)
optimizer.verbosity = VerbosityLevels.DEBUG
mapping = optimizer.optimize(static, moving, None)
m = optimizer.get_map()
assert_equal(mapping, m)
energy_profile = subsample_profile(
optimizer.full_energy_profile, 10)
print(energy_profile)
if USING_VC_SSE2:
expected_profile = \
np.array([348.3204721, 143.480757, 44.30003405, 8.73624842,
3.13227203, 14.70806563, 6.48360268, 23.52491883,
17.25669088, 48.99709064])
elif USING_GCC_SSE2:
expected_profile = \
np.array([348.3204721, 143.48075646, 44.30003413, 8.73624841,
3.13227181, 14.70806845, 6.48360884, 23.52499421,
17.25667176, 48.997691])
assert_array_almost_equal(energy_profile, expected_profile, decimal=4)
def test_ssd_3d_gauss_newton():
r'''
Register a stack of circles ('cylinder') before and after warping them with
a synthetic diffeomorphism. This test is intended to detect regressions
only: we saved the energy profile (the sequence of energy values at each
iteration) of a working version of SSD in 3D using the Gauss-Newton step,
and this test checks that the current energy profile matches the saved
one. The validation of the "working version" was
done by registering the 18 manually annotated T1 brain MRI database IBSR
with each other and computing the jaccard index for all 31 common anatomical
regions.
'''
moving, static = get_synthetic_warped_circle(35)
moving[...,:10] = 0
moving[...,-1:-11:-1] = 0
static[...,:10] = 0
static[...,-1:-11:-1] = 0
#Create the SSD metric
smooth = 4
inner_iter = 5
step_type = 'gauss_newton'
similarity_metric = metrics.SSDMetric(3, smooth, inner_iter, step_type)
#Create the optimizer
level_iters = [10, 5]
step_length = 0.25
opt_tol = 1e-4
inv_iter = 20
inv_tol = 1e-3
ss_sigma_factor = 0.5
optimizer = imwarp.SymmetricDiffeomorphicRegistration(similarity_metric,
level_iters, step_length, ss_sigma_factor, opt_tol, inv_iter, inv_tol)
optimizer.verbosity = VerbosityLevels.DEBUG
mapping = optimizer.optimize(static, moving, None)
m = optimizer.get_map()
assert_equal(mapping, m)
energy_profile = np.array(optimizer.full_energy_profile)
if floating is np.float32:
expected_profile = \
np.array([348.32047209927373, 143.49111863234222, 44.328151771258206,
8.759564367010988, 3.1378191742723662, 14.846951961939153,
6.405154727081836, 20.437036950018083, 17.399044912417597,
49.072929423423496, 269.2553956858318, 80.72079256138973,
200.242742072974, 68.21238489882822, 208.28730597575378])
else:
expected_profile = \
np.array([348.32049916992855, 143.49111631974688, 44.328145727486174,
8.759562612294948, 3.137819214539283, 14.846929648490525,
6.405154001052728, 20.437123731745928, 17.39892098642616,
49.07339619667625, 269.2533380585498, 80.72162511785703,
200.24505477294477, 68.21183884286609, 208.29025925025073])
assert_array_almost_equal(energy_profile, expected_profile, decimal=6)
def test_ssd_3d_demons():
r'''
Register a stack of circles ('cylinder') before and after warping them
with a synthetic diffeomorphism. This test is intended to detect
regressions only: we saved the energy profile (the sequence of energy
values at each iteration) of a working version of SSD in 3D using the
Demons step, and this test checks that the current energy profile matches
the saved one. The validation of the "working version" was done by
registering the 18 manually annotated T1 brain MRI database IBSR with each
other and computing the jaccard index for all 31 common anatomical regions.
'''
moving, static = get_synthetic_warped_circle(30)
moving[..., :8] = 0
moving[..., -1:-9:-1] = 0
static[..., :8] = 0
static[..., -1:-9:-1] = 0
#Create the SSD metric
smooth = 4
step_type = 'demons'
similarity_metric = metrics.SSDMetric(3,
smooth=smooth,
step_type=step_type)
#Create the optimizer
level_iters = [10, 5]
step_length = 0.25
opt_tol = 1e-4
inv_iter = 20
inv_tol = 1e-3
ss_sigma_factor = 0.5
optimizer = imwarp.SymmetricDiffeomorphicRegistration(
similarity_metric, level_iters, step_length, ss_sigma_factor, opt_tol,
inv_iter, inv_tol)
optimizer.verbosity = VerbosityLevels.DEBUG
mapping = optimizer.optimize(static, moving, None)
m = optimizer.get_map()
assert_equal(mapping, m)
energy_profile = subsample_profile(optimizer.full_energy_profile, 10)
print(energy_profile)
if USING_VC_SSE2:
expected_profile = \
np.array([312.22706987, 154.65556884, 53.88543188, 9.11484007,
36.46592407, 13.20522299, 48.65663399, 14.91579802,
49.82954704, 14.92646254])
elif USING_GCC_SSE2:
expected_profile = \
np.array([312.22706987, 154.65556885, 53.88455398, 9.11770682,
36.48642824, 13.21706748, 48.67710635, 14.91782047,
49.84142899, 14.92531294])
assert_array_almost_equal(energy_profile, expected_profile, decimal=4)
def test_optimizer_exceptions():
#An arbitrary valid metric
metric = metrics.SSDMetric(2)
# The metric must not be None
assert_raises(ValueError, imwarp.SymmetricDiffeomorphicRegistration, None)
# The iterations list must not be empty
assert_raises(ValueError, imwarp.SymmetricDiffeomorphicRegistration, metric, [])
optimizer = imwarp.SymmetricDiffeomorphicRegistration(metric, None)
#Verify the default iterations list
assert_array_equal(optimizer.level_iters, [100,100,25])
#Verify exception thrown when attepting to fit the energy profile without enough data
assert_raises(ValueError, optimizer._get_energy_derivative)
def test_ssd_3d_demons():
r'''
Register a stack of circles ('cylinder') before and after warping them with
a synthetic diffeomorphism. This test is intended to detect regressions
only: we saved the energy profile (the sequence of energy values at each
iteration) of a working version of SSD in 3D using the Demons step, and this
test checks that the current energy profile matches the saved one. The
validation of the "working version" was done by registering the 18 manually
annotated T1 brain MRI database IBSR with each other and computing the
jaccard index for all 31 common anatomical regions.
'''
moving, static = get_synthetic_warped_circle(30)
moving[...,:8] = 0
moving[...,-1:-9:-1] = 0
static[...,:8] = 0
static[...,-1:-9:-1] = 0
#Create the SSD metric
smooth = 4
step_type = 'demons'
similarity_metric = metrics.SSDMetric(3, smooth=smooth, step_type=step_type)
#Create the optimizer
level_iters = [10, 5]
step_length = 0.25
opt_tol = 1e-4
inv_iter = 20
inv_tol = 1e-3
ss_sigma_factor = 0.5
optimizer = imwarp.SymmetricDiffeomorphicRegistration(similarity_metric,
level_iters, step_length, ss_sigma_factor, opt_tol, inv_iter, inv_tol)
optimizer.verbosity = VerbosityLevels.DEBUG
mapping = optimizer.optimize(static, moving, None)
m = optimizer.get_map()
assert_equal(mapping, m)
energy_profile = np.array(optimizer.full_energy_profile)
if floating is np.float32:
expected_profile = \
np.array([312.22706987, 154.65556885, 53.88594419, 9.22626825,
36.50370933, 13.54829978, 49.57619437, 15.71122527,
53.45897119, 15.62018739, 521.95785712, 158.16217928,
182.49116432, 144.91081752, 176.6810387])
else:
expected_profile = \
np.array([312.22709468, 154.65706498, 53.8856324, 8.90160898,
34.91911552, 12.66043296, 49.61341791, 15.14198327,
52.25467529, 18.88243845, 490.48088231, 149.29027701,
192.26219053, 137.5291187, 187.2795753])
assert_array_almost_equal(energy_profile, expected_profile, decimal=6)
def test_ssd_3d_gauss_newton():
r""" Test 3D SyN with SSD metric, Gauss-Newton optimizer
Register a stack of circles ('cylinder') before and after warping them
with a synthetic diffeomorphism. We verify that the final registration
is of good quality.
"""
moving, static = get_synthetic_warped_circle(35)
moving[..., :10] = 0
moving[..., -1:-11:-1] = 0
static[..., :10] = 0
static[..., -1:-11:-1] = 0
# Create the SSD metric
smooth = 4
inner_iter = 5
step_type = 'gauss_newton'
similarity_metric = metrics.SSDMetric(3, smooth, inner_iter, step_type)
# Create the optimizer
level_iters = [10, 10]
step_length = 0.1
opt_tol = 1e-4
inv_iter = 20
inv_tol = 1e-3
ss_sigma_factor = 0.5
optimizer = imwarp.SymmetricDiffeomorphicRegistration(
similarity_metric,
level_iters,
step_length,
ss_sigma_factor,
opt_tol,
inv_iter,
inv_tol)
optimizer.verbosity = VerbosityLevels.DEBUG
mapping = optimizer.optimize(static, moving, None)
m = optimizer.get_map()
assert_equal(mapping, m)
warped = mapping.transform(moving)
starting_energy = np.sum((static - moving)**2)
final_energy = np.sum((static - warped)**2)
reduced = 1.0 - final_energy / starting_energy
assert(reduced > 0.9)
def test_ssd_2d_gauss_newton():
r'''
Classical Circle-To-C experiment for 2D Monomodal registration. This test
is intended to detect regressions only: we saved the energy profile (the
sequence of energy values at each iteration) of a working version of SSD
in 2D using the Gauss Newton step, and this test checks that the current
energy profile matches the saved one.
'''
fname_moving = get_data('reg_o')
fname_static = get_data('reg_c')
moving = np.load(fname_moving)
static = np.load(fname_static)
moving = np.array(moving, dtype=floating)
static = np.array(static, dtype=floating)
moving = (moving-moving.min())/(moving.max() - moving.min())
static = (static-static.min())/(static.max() - static.min())
#Create the SSD metric
smooth = 4
inner_iter = 5
step_type = 'gauss_newton'
similarity_metric = metrics.SSDMetric(2, smooth, inner_iter, step_type)
#Configure and run the Optimizer
level_iters = [200, 100, 50, 25]
step_length = 0.5
opt_tol = 1e-4
inv_iter = 40
inv_tol = 1e-3
ss_sigma_factor = 0.2
optimizer = imwarp.SymmetricDiffeomorphicRegistration(similarity_metric,
level_iters, step_length, ss_sigma_factor, opt_tol, inv_iter, inv_tol)
#test callback not being called
optimizer.INIT_START_CALLED = 0
optimizer.INIT_END_CALLED = 0
optimizer.OPT_START_CALLED = 0
optimizer.OPT_END_CALLED = 0
optimizer.SCALE_START_CALLED = 0
optimizer.SCALE_END_CALLED = 0
optimizer.ITER_START_CALLED = 0
optimizer.ITER_END_CALLED = 0
optimizer.verbosity = VerbosityLevels.DEBUG
id = np.eye(3)
mapping = optimizer.optimize(static, moving, id, id, id)
m = optimizer.get_map()
assert_equal(mapping, m)
subsampled_energy_profile = subsample_profile(
optimizer.full_energy_profile, 10)
print(subsampled_energy_profile)
if USING_VC_SSE2:
expected_profile = \
np.array([312.68133316, 70.17782995, 21.38508088, 96.41054776,
49.990781, 43.11867579, 24.53952718, 51.0786643,
143.24848252, 150.48349573])
elif USING_GCC_SSE2:
expected_profile = \
np.array([312.68133316, 70.17782938, 21.26798507, 96.51765054,
51.1495088, 37.86204803, 21.62425293, 49.44868302,
121.6643917, 137.91427228])
assert_array_almost_equal(subsampled_energy_profile, expected_profile,
decimal = 5)
assert_equal(optimizer.OPT_START_CALLED, 0)
assert_equal(optimizer.OPT_END_CALLED, 0)
assert_equal(optimizer.SCALE_START_CALLED, 0)
assert_equal(optimizer.SCALE_END_CALLED, 0)
assert_equal(optimizer.ITER_START_CALLED, 0)
assert_equal(optimizer.ITER_END_CALLED, 0)
def test_ssd_2d_demons():
r'''
Classical Circle-To-C experiment for 2D Monomodal registration. This test
is intended to detect regressions only: we saved the energy profile (the
sequence of energy values at each iteration) of a working version of SSD in
2D using the Demons step, and this test checks that the current energy
profile matches the saved one.
'''
fname_moving = get_data('reg_o')
fname_static = get_data('reg_c')
moving = np.load(fname_moving)
static = np.load(fname_static)
moving = np.array(moving, dtype=floating)
static = np.array(static, dtype=floating)
moving = (moving-moving.min())/(moving.max() - moving.min())
static = (static-static.min())/(static.max() - static.min())
#Create the SSD metric
smooth = 4
step_type = 'demons'
similarity_metric = metrics.SSDMetric(2, smooth=smooth, step_type=step_type)
#Configure and run the Optimizer
level_iters = [200, 100, 50, 25]
step_length = 0.25
opt_tol = 1e-4
inv_iter = 40
inv_tol = 1e-3
ss_sigma_factor = 0.2
optimizer = imwarp.SymmetricDiffeomorphicRegistration(similarity_metric,
level_iters, step_length, ss_sigma_factor, opt_tol, inv_iter, inv_tol)
#test callback being called
optimizer.INIT_START_CALLED = 0
optimizer.INIT_END_CALLED = 0
optimizer.OPT_START_CALLED = 0
optimizer.OPT_END_CALLED = 0
optimizer.SCALE_START_CALLED = 0
optimizer.SCALE_END_CALLED = 0
optimizer.ITER_START_CALLED = 0
optimizer.ITER_END_CALLED = 0
optimizer.callback_counter_test = 0
optimizer.callback = simple_callback
optimizer.verbosity = VerbosityLevels.DEBUG
mapping = optimizer.optimize(static, moving, None)
m = optimizer.get_map()
assert_equal(mapping, m)
subsampled_energy_profile = subsample_profile(
optimizer.full_energy_profile, 10)
print(subsampled_energy_profile)
if USING_VC_SSE2:
expected_profile = \
np.array([312.6813333, 80.74625551, 49.43591374, 34.08871301,
25.18286981, 17.78955273, 25.91334939, 20.16932281,
43.86083145, 79.0966558 ])
elif USING_GCC_SSE2:
expected_profile = \
np.array([312.6813333, 98.17321941, 60.98300837, 47.75387157,
34.11067498, 122.91901409, 19.75599298, 14.28763847,
36.33599718, 88.62426913])
assert_array_almost_equal(subsampled_energy_profile,
expected_profile, decimal=5)
assert_equal(optimizer.OPT_START_CALLED, 1)
assert_equal(optimizer.OPT_END_CALLED, 1)
assert_equal(optimizer.SCALE_START_CALLED, 1)
assert_equal(optimizer.SCALE_END_CALLED, 1)
assert_equal(optimizer.ITER_START_CALLED, 1)
assert_equal(optimizer.ITER_END_CALLED, 1)
def test_ssd_2d_gauss_newton():
r'''
Classical Circle-To-C experiment for 2D Monomodal registration. This test
is intended to detect regressions only: we saved the energy profile (the
sequence of energy values at each iteration) of a working version of SSD in
2D using the Gauss Newton step, and this test checks that the current energy
profile matches the saved one.
'''
fname_moving = get_data('reg_o')
fname_static = get_data('reg_c')
moving = np.load(fname_moving)
static = np.load(fname_static)
moving = np.array(moving, dtype=floating)
static = np.array(static, dtype=floating)
moving = (moving-moving.min())/(moving.max() - moving.min())
static = (static-static.min())/(static.max() - static.min())
#Create the SSD metric
smooth = 4
inner_iter = 5
step_type = 'gauss_newton'
similarity_metric = metrics.SSDMetric(2, smooth, inner_iter, step_type)
#Configure and run the Optimizer
level_iters = [200, 100, 50, 25]
step_length = 0.5
opt_tol = 1e-4
inv_iter = 40
inv_tol = 1e-3
ss_sigma_factor = 0.2
optimizer = imwarp.SymmetricDiffeomorphicRegistration(similarity_metric,
level_iters, step_length, ss_sigma_factor, opt_tol, inv_iter, inv_tol)
#test callback not being called
optimizer.INIT_START_CALLED = 0
optimizer.INIT_END_CALLED = 0
optimizer.OPT_START_CALLED = 0
optimizer.OPT_END_CALLED = 0
optimizer.SCALE_START_CALLED = 0
optimizer.SCALE_END_CALLED = 0
optimizer.ITER_START_CALLED = 0
optimizer.ITER_END_CALLED = 0
optimizer.verbosity = VerbosityLevels.DEBUG
mapping = optimizer.optimize(static, moving, np.eye(3), np.eye(3), np.eye(3))
m = optimizer.get_map()
assert_equal(mapping, m)
subsampled_energy_profile = np.array(optimizer.full_energy_profile[::10])
if floating is np.float32:
expected_profile = \
np.array([312.68133316, 79.40404517, 23.3715698, 125.02700267,
59.79982213, 34.64971733, 23.37131446, 171.28250576,
62.22266377, 125.24392168])
else:
expected_profile = \
np.array([312.68133361, 79.40404354, 23.34588446, 124.3247997,
61.69601973, 38.15047181, 23.53315113, 80.0791295,
57.21700113, 143.73270476])
assert_array_almost_equal(subsampled_energy_profile, expected_profile)
assert_equal(optimizer.OPT_START_CALLED, 0)
assert_equal(optimizer.OPT_END_CALLED, 0)
assert_equal(optimizer.SCALE_START_CALLED, 0)
assert_equal(optimizer.SCALE_END_CALLED, 0)
assert_equal(optimizer.ITER_START_CALLED, 0)
assert_equal(optimizer.ITER_END_CALLED, 0)
def test_ssd_2d_demons():
r'''
Classical Circle-To-C experiment for 2D Monomodal registration. This test
is intended to detect regressions only: we saved the energy profile (the
sequence of energy values at each iteration) of a working version of SSD in
2D using the Demons step, and this test checks that the current energy
profile matches the saved one.
'''
fname_moving = get_data('reg_o')
fname_static = get_data('reg_c')
moving = np.load(fname_moving)
static = np.load(fname_static)
moving = np.array(moving, dtype=floating)
static = np.array(static, dtype=floating)
moving = (moving-moving.min())/(moving.max() - moving.min())
static = (static-static.min())/(static.max() - static.min())
#Create the SSD metric
smooth = 4
step_type = 'demons'
similarity_metric = metrics.SSDMetric(2, smooth=smooth, step_type=step_type)
#Configure and run the Optimizer
level_iters = [200, 100, 50, 25]
step_length = 0.25
opt_tol = 1e-4
inv_iter = 40
inv_tol = 1e-3
ss_sigma_factor = 0.2
optimizer = imwarp.SymmetricDiffeomorphicRegistration(similarity_metric,
level_iters, step_length, ss_sigma_factor, opt_tol, inv_iter, inv_tol)
#test callback being called
optimizer.INIT_START_CALLED = 0
optimizer.INIT_END_CALLED = 0
optimizer.OPT_START_CALLED = 0
optimizer.OPT_END_CALLED = 0
optimizer.SCALE_START_CALLED = 0
optimizer.SCALE_END_CALLED = 0
optimizer.ITER_START_CALLED = 0
optimizer.ITER_END_CALLED = 0
optimizer.callback_counter_test = 0
optimizer.callback = simple_callback
optimizer.verbosity = VerbosityLevels.DEBUG
mapping = optimizer.optimize(static, moving, None)
m = optimizer.get_map()
assert_equal(mapping, m)
subsampled_energy_profile = np.array(optimizer.full_energy_profile[::10])
if floating is np.float32:
expected_profile = \
np.array([312.6813333, 162.57756447, 99.2766679, 77.38698935,
61.75415204, 55.37420428, 46.36872571, 41.81811505,
36.38683617, 33.03952963, 30.91409901, 54.41447237,
23.40232241, 12.75092466, 10.19231733, 9.21058037,
57.4636143, 38.94004856, 36.26093212, 108.0136453,
81.35521049, 74.61956833])
else:
expected_profile = \
np.array([312.68133361, 162.57744066, 99.27669798, 77.38683186,
61.75391429, 55.3740711, 46.36870776, 41.81809239,
36.3898153, 32.78365961, 30.69843811, 53.67073767,
21.74630524, 11.98102583, 11.51086685, 55.30707781,
39.88467545, 34.29444978, 33.10822964, 122.64743831,
84.18144073, 75.60088687])
assert_array_almost_equal(subsampled_energy_profile, expected_profile)
assert_equal(optimizer.OPT_START_CALLED, 1)
assert_equal(optimizer.OPT_END_CALLED, 1)
assert_equal(optimizer.SCALE_START_CALLED, 1)
assert_equal(optimizer.SCALE_END_CALLED, 1)
assert_equal(optimizer.ITER_START_CALLED, 1)
assert_equal(optimizer.ITER_END_CALLED, 1)
