def test_dti_xval():
"""
Test k-fold cross-validation
"""
data = nib.load(fdata).get_data()
gtab = gt.gradient_table(fbval, fbvec)
dm = dti.TensorModel(gtab, "LS")
# The data has 102 directions, so will not divide neatly into 10 bits
npt.assert_raises(ValueError, xval.kfold_xval, dm, data, 10)
# But we can do this with 2 folds:
kf_xval = xval.kfold_xval(dm, data, 2)
# In simulation with no noise, COD should be perfect:
psphere = dpd.get_sphere("symmetric362")
bvecs = np.concatenate(([[0, 0, 0]], psphere.vertices))
bvals = np.zeros(len(bvecs)) + 1000
bvals[0] = 0
gtab = gt.gradient_table(bvals, bvecs)
mevals = np.array(([0.0015, 0.0003, 0.0001], [0.0015, 0.0003, 0.0003]))
mevecs = [np.array([[1, 0, 0], [0, 1, 0], [0, 0, 1]]), np.array([[0, 0, 1], [0, 1, 0], [1, 0, 0]])]
S = sims.single_tensor(gtab, 100, mevals[0], mevecs[0], snr=None)
dm = dti.TensorModel(gtab, "LS")
kf_xval = xval.kfold_xval(dm, S, 2)
cod = xval.coeff_of_determination(S, kf_xval)
npt.assert_array_almost_equal(cod, np.ones(kf_xval.shape[:-1]) * 100)
# Test with 2D data for use of a mask
S = np.array([[S, S], [S, S]])
mask = np.ones(S.shape[:-1], dtype=bool)
mask[1, 1] = 0
kf_xval = xval.kfold_xval(dm, S, 2, mask=mask)
cod2d = xval.coeff_of_determination(S, kf_xval)
npt.assert_array_almost_equal(np.round(cod2d[0, 0]), cod)
def test_dti_xval():
"""
Test k-fold cross-validation
"""
data = nib.load(fdata).get_data()
gtab = gt.gradient_table(fbval, fbvec)
dm = dti.TensorModel(gtab, 'LS')
# The data has 102 directions, so will not divide neatly into 10 bits
npt.assert_raises(ValueError, xval.kfold_xval, dm, data, 10)
# But we can do this with 2 folds:
kf_xval = xval.kfold_xval(dm, data, 2)
# In simulation with no noise, COD should be perfect:
psphere = dpd.get_sphere('symmetric362')
bvecs = np.concatenate(([[0, 0, 0]], psphere.vertices))
bvals = np.zeros(len(bvecs)) + 1000
bvals[0] = 0
gtab = gt.gradient_table(bvals, bvecs)
mevals = np.array(([0.0015, 0.0003, 0.0001], [0.0015, 0.0003, 0.0003]))
mevecs = [
np.array([[1, 0, 0], [0, 1, 0], [0, 0, 1]]),
np.array([[0, 0, 1], [0, 1, 0], [1, 0, 0]])
]
S = sims.single_tensor(gtab, 100, mevals[0], mevecs[0], snr=None)
dm = dti.TensorModel(gtab, 'LS')
dmfit = dm.fit(S)
kf_xval = xval.kfold_xval(dm, S, 2)
cod = xval.coeff_of_determination(S, kf_xval)
npt.assert_array_almost_equal(cod, np.ones(kf_xval.shape[:-1]) * 100)
def test_csd_xval():
# First, let's see that it works with some data:
data = nib.load(fdata).get_data()[1:3, 1:3, 1:3] # Make it *small*
gtab = gt.gradient_table(fbval, fbvec)
S0 = np.mean(data[..., gtab.b0s_mask])
response = ([0.0015, 0.0003, 0.0001], S0)
csdm = csd.ConstrainedSphericalDeconvModel(gtab, response)
kf_xval = xval.kfold_xval(csdm, data, 2, response, sh_order=2)
# In simulation, it should work rather well (high COD):
psphere = dpd.get_sphere('symmetric362')
bvecs = np.concatenate(([[0, 0, 0]], psphere.vertices))
bvals = np.zeros(len(bvecs)) + 1000
bvals[0] = 0
gtab = gt.gradient_table(bvals, bvecs)
mevals = np.array(([0.0015, 0.0003, 0.0001], [0.0015, 0.0003, 0.0003]))
mevecs = [
np.array([[1, 0, 0], [0, 1, 0], [0, 0, 1]]),
np.array([[0, 0, 1], [0, 1, 0], [1, 0, 0]])
]
S0 = 100
S = sims.single_tensor(gtab, S0, mevals[0], mevecs[0], snr=None)
sm = csd.ConstrainedSphericalDeconvModel(gtab, response)
smfit = sm.fit(S)
np.random.seed(12345)
response = ([0.0015, 0.0003, 0.0001], S0)
kf_xval = xval.kfold_xval(sm, S, 2, response, sh_order=2)
# Because of the regularization, COD is not going to be perfect here:
cod = xval.coeff_of_determination(S, kf_xval)
# We'll just test for regressions:
csd_cod = 97 # pre-computed by hand for this random seed
# We're going to be really lenient here:
npt.assert_array_almost_equal(np.round(cod), csd_cod)
def test_csd_xval():
# First, let's see that it works with some data:
data = nib.load(fdata).get_data()[1:3, 1:3, 1:3] # Make it *small*
gtab = gt.gradient_table(fbval, fbvec)
S0 = np.mean(data[..., gtab.b0s_mask])
response = ([0.0015, 0.0003, 0.0001], S0)
csdm = csd.ConstrainedSphericalDeconvModel(gtab, response)
kf_xval = xval.kfold_xval(csdm, data, 2, response, sh_order=2)
# In simulation, it should work rather well (high COD):
psphere = dpd.get_sphere('symmetric362')
bvecs = np.concatenate(([[0, 0, 0]], psphere.vertices))
bvals = np.zeros(len(bvecs)) + 1000
bvals[0] = 0
gtab = gt.gradient_table(bvals, bvecs)
mevals = np.array(([0.0015, 0.0003, 0.0001], [0.0015, 0.0003, 0.0003]))
mevecs = [ np.array( [ [1, 0, 0], [0, 1, 0], [0, 0, 1] ] ),
np.array( [ [0, 0, 1], [0, 1, 0], [1, 0, 0] ] ) ]
S0 = 100
S = sims.single_tensor( gtab, S0, mevals[0], mevecs[0], snr=None )
sm = csd.ConstrainedSphericalDeconvModel(gtab, response)
smfit = sm.fit(S)
np.random.seed(12345)
response = ([0.0015, 0.0003, 0.0001], S0)
kf_xval = xval.kfold_xval(sm, S, 2, response, sh_order=2)
# Because of the regularization, COD is not going to be perfect here:
cod = xval.coeff_of_determination(S, kf_xval)
# We'll just test for regressions:
csd_cod = 97 # pre-computed by hand for this random seed
# We're going to be really lenient here:
npt.assert_array_almost_equal(np.round(cod), csd_cod)
def test_dti_xval():
data = load_nifti_data(fdata)
gtab = gt.gradient_table(fbval, fbvec)
dm = dti.TensorModel(gtab, 'LS')
# The data has 102 directions, so will not divide neatly into 10 bits
npt.assert_raises(ValueError, xval.kfold_xval, dm, data, 10)
# In simulation with no noise, COD should be perfect:
psphere = dpd.get_sphere('symmetric362')
bvecs = np.concatenate(([[0, 0, 0]], psphere.vertices))
bvals = np.zeros(len(bvecs)) + 1000
bvals[0] = 0
gtab = gt.gradient_table(bvals, bvecs)
mevals = np.array(([0.0015, 0.0003, 0.0001], [0.0015, 0.0003, 0.0003]))
mevecs = [
np.array([[1, 0, 0], [0, 1, 0], [0, 0, 1]]),
np.array([[0, 0, 1], [0, 1, 0], [1, 0, 0]])
]
S = sims.single_tensor(gtab, 100, mevals[0], mevecs[0], snr=None)
dm = dti.TensorModel(gtab, 'LS')
kf_xval = xval.kfold_xval(dm, S, 2)
cod = xval.coeff_of_determination(S, kf_xval)
npt.assert_array_almost_equal(cod, np.ones(kf_xval.shape[:-1]) * 100)
# Test with 2D data for use of a mask
S = np.array([[S, S], [S, S]])
mask = np.ones(S.shape[:-1], dtype=bool)
mask[1, 1] = 0
kf_xval = xval.kfold_xval(dm, S, 2, mask=mask)
cod2d = xval.coeff_of_determination(S, kf_xval)
npt.assert_array_almost_equal(np.round(cod2d[0, 0]), cod)
def test_csd_xval():
# First, let's see that it works with some data:
data = load_nifti_data(fdata)[1:3, 1:3, 1:3] # Make it *small*
gtab = gt.gradient_table(fbval, fbvec)
S0 = np.mean(data[..., gtab.b0s_mask])
response = ([0.0015, 0.0003, 0.0001], S0)
# In simulation, it should work rather well (high COD):
psphere = dpd.get_sphere('symmetric362')
bvecs = np.concatenate(([[0, 0, 0]], psphere.vertices))
bvals = np.zeros(len(bvecs)) + 1000
bvals[0] = 0
gtab = gt.gradient_table(bvals, bvecs)
mevals = np.array(([0.0015, 0.0003, 0.0001], [0.0015, 0.0003, 0.0003]))
mevecs = [
np.array([[1, 0, 0], [0, 1, 0], [0, 0, 1]]),
np.array([[0, 0, 1], [0, 1, 0], [1, 0, 0]])
]
S0 = 100
S = sims.single_tensor(gtab, S0, mevals[0], mevecs[0], snr=None)
with warnings.catch_warnings():
warnings.filterwarnings("ignore",
message=descoteaux07_legacy_msg,
category=PendingDeprecationWarning)
sm = csd.ConstrainedSphericalDeconvModel(gtab, response)
np.random.seed(12345)
response = ([0.0015, 0.0003, 0.0001], S0)
with warnings.catch_warnings():
warnings.filterwarnings("ignore",
message=descoteaux07_legacy_msg,
category=PendingDeprecationWarning)
kf_xval = xval.kfold_xval(sm, S, 2, response, sh_order=2)
# Because of the regularization, COD is not going to be perfect here:
cod = xval.coeff_of_determination(S, kf_xval)
# We'll just test for regressions:
csd_cod = 97 # pre-computed by hand for this random seed
# We're going to be really lenient here:
npt.assert_array_almost_equal(np.round(cod), csd_cod)
# Test for sD data with more than one voxel for use of a mask:
S = np.array([[S, S], [S, S]])
mask = np.ones(S.shape[:-1], dtype=bool)
mask[1, 1] = 0
with warnings.catch_warnings():
warnings.filterwarnings("ignore",
message=descoteaux07_legacy_msg,
category=PendingDeprecationWarning)
kf_xval = xval.kfold_xval(sm, S, 2, response, sh_order=2, mask=mask)
cod = xval.coeff_of_determination(S, kf_xval)
npt.assert_array_almost_equal(np.round(cod[0]), csd_cod)
dti_model = dti.TensorModel(gtab) response, ratio = csd.auto_response(gtab, data, roi_radius=10, fa_thr=0.7) csd_model = csd.ConstrainedSphericalDeconvModel(gtab, response) """ Next, we perform cross-validation for each kind of model, comparing model predictions to the diffusion MRI data in each one of these voxels. Note that we use 2-fold cross-validation, which means that in each iteration, the model will be fit to half of the data, and used to predict the other half. """ dti_cc = xval.kfold_xval(dti_model, cc_vox, 2) csd_cc = xval.kfold_xval(csd_model, cc_vox, 2, response) dti_cso = xval.kfold_xval(dti_model, cso_vox, 2) csd_cso = xval.kfold_xval(csd_model, cso_vox, 2, response) """ We plot a scatter plot of the data with the model predictions in each of these voxels, focusing only on the diffusion-weighted measurements (each point corresponds to a different gradient direction). The two models are compared in each sub-plot (blue=DTI, red=CSD). """ fig, ax = plt.subplots(1,2) fig.set_size_inches([12,6])
""" dti_model = dti.TensorModel(gtab) response, ratio = csd.auto_response(gtab, data, roi_radius=10, fa_thr=0.7) csd_model = csd.ConstrainedSphericalDeconvModel(gtab, response) """ Next, we perform cross-validation for each kind of model, comparing model predictions to the diffusion MRI data in each one of these voxels. Note that we use 2-fold cross-validation, which means that in each iteration, the model will be fit to half of the data, and used to predict the other half. """ dti_cc = xval.kfold_xval(dti_model, cc_vox, 2) csd_cc = xval.kfold_xval(csd_model, cc_vox, 2, response) dti_cso = xval.kfold_xval(dti_model, cso_vox, 2) csd_cso = xval.kfold_xval(csd_model, cso_vox, 2, response) """ We plot a scatter plot of the data with the model predictions in each of these voxels, focusing only on the diffusion-weighted measurements (each point corresponds to a different gradient direction). The two models are compared in each sub-plot (blue=DTI, red=CSD). """ fig, ax = plt.subplots(1, 2) fig.set_size_inches([12, 6]) ax[0].plot(cc_vox[gtab.b0s_mask == 0],
def compare_models(subject):
s3 = boto3.resource('s3')
boto3.setup_default_session(profile_name='cirrus')
bucket = s3.Bucket('hcp-dki')
path_dki = '%s/%s_cod_dki.nii.gz' % (subject, subject)
path_dti = '%s/%s_cod_dti.nii.gz' % (subject, subject)
if not (exists(path_dki, bucket.name) and exists(path_dti, bucket.name)):
print("Files don't exist - going ahead")
with tempfile.TemporaryDirectory() as tempdir:
try:
bucket = setup_boto()
dwi_file = op.join(tempdir, 'data.nii.gz')
bvec_file = op.join(tempdir, 'data.bvec')
bval_file = op.join(tempdir, 'data.bval')
data_files = {}
data_files[dwi_file] = \
'HCP_900/%s/T1w/Diffusion/data.nii.gz' % subject
data_files[bvec_file] = \
'HCP_900/%s/T1w/Diffusion/bvecs' % subject
data_files[bval_file] = \
'HCP_900/%s/T1w/Diffusion/bvals' % subject
for k in data_files.keys():
if not op.exists(k):
bucket.download_file(data_files[k], k)
wm_file = op.join(tempdir, 'wm.nii.gz')
boto3.setup_default_session(profile_name='cirrus')
s3 = boto3.resource('s3')
s3.meta.client.download_file(
'hcp-dki',
'%s/%s_white_matter_mask.nii.gz' % (subject, subject),
wm_file)
wm_mask = nib.load(wm_file).get_data().astype(bool)
dwi_img = nib.load(dwi_file)
data = dwi_img.get_data()
bvals = np.loadtxt(bval_file)
bvecs = np.loadtxt(bvec_file)
gtab = dpg.gradient_table(bvals, bvecs,
b0_threshold=10)
s3 = boto3.resource('s3')
boto3.setup_default_session(profile_name='cirrus')
bucket = s3.Bucket('hcp-dki')
for model_object, method in zip([dti.TensorModel,
dki.DiffusionKurtosisModel],
['dti', 'dki']):
path_method = '%s/%s_cod_%s.nii.gz' % (subject,
subject,
method)
if not (exists(path_method, bucket.name)):
print("No %s file - fitting" % method)
print("1")
model = model_object(gtab)
print("2")
pred = xval.kfold_xval(model, data, 5, mask=wm_mask)
print("3")
cod = xval.coeff_of_determination(pred, data)
cod_file = op.join(tempdir, 'cod_%s.nii.gz' % method)
print("4")
nib.save(nib.Nifti1Image(cod, dwi_img.affine),
cod_file)
print("5")
s3.meta.client.upload_file(
cod_file,
'hcp-dki',
path_method)
return subject, True
except Exception as err:
return subject, err.args
else:
return subject, True
def calc_cod1000(subject):
s3 = boto3.resource('s3')
boto3.setup_default_session(profile_name='cirrus')
bucket = s3.Bucket('hcp-dki')
path_dti = '%s/%s_cod_dti_1000.nii.gz' % (subject, subject)
path_dki = '%s/%s_cod_dki_1000.nii.gz' % (subject, subject)
if not (exists(path_dti, bucket.name) and exists(path_dki, bucket.name)):
print("File doesn't exist - going ahead")
with tempfile.TemporaryDirectory() as tempdir:
try:
bucket = setup_boto()
dwi_file = op.join(tempdir, 'data.nii.gz')
bvec_file = op.join(tempdir, 'data.bvec')
bval_file = op.join(tempdir, 'data.bval')
data_files = {}
data_files[dwi_file] = \
'HCP_900/%s/T1w/Diffusion/data.nii.gz' % subject
data_files[bvec_file] = \
'HCP_900/%s/T1w/Diffusion/bvecs' % subject
data_files[bval_file] = \
'HCP_900/%s/T1w/Diffusion/bvals' % subject
for k in data_files.keys():
if not op.exists(k):
bucket.download_file(data_files[k], k)
wm_file = op.join(tempdir, 'wm.nii.gz')
boto3.setup_default_session(profile_name='cirrus')
s3 = boto3.resource('s3')
bucket = s3.Bucket('hcp-dki')
s3.meta.client.download_file(
'hcp-dki',
'%s/%s_white_matter_mask.nii.gz' % (subject, subject),
wm_file)
wm_mask = nib.load(wm_file).get_data().astype(bool)
dwi_img = nib.load(dwi_file)
data = dwi_img.get_data()
bvals = np.loadtxt(bval_file)
bvecs = np.loadtxt(bvec_file)
idx = bvals < 1985
if not exists(path_dki, bucket.name):
gtab = dpg.gradient_table(bvals, bvecs, b0_threshold=10)
dki_model = dki.DiffusionKurtosisModel(gtab)
# Use all the data to calculate the mode
pred = xval.kfold_xval(dki_model, data, 5, mask=wm_mask)
# But compare only on the b=1000 shell (same as DTI):
cod = xval.coeff_of_determination(pred[..., idx],
data[..., idx])
cod_file = op.join(tempdir, 'cod_dki_1000.nii.gz')
nib.save(nib.Nifti1Image(cod, dwi_img.affine),
cod_file)
s3.meta.client.upload_file(
cod_file,
'hcp-dki',
path_dki)
if not exists(path_dti, bucket.name):
data = data[..., idx]
gtab = dpg.gradient_table(bvals[idx],
bvecs[:, idx].squeeze(),
b0_threshold=10)
model = dti.TensorModel(gtab)
pred = xval.kfold_xval(model, data, 5, mask=wm_mask)
cod = xval.coeff_of_determination(pred, data)
cod_file = op.join(tempdir, 'cod_dti_1000.nii.gz')
nib.save(nib.Nifti1Image(cod, dwi_img.affine),
cod_file)
s3.meta.client.upload_file(
cod_file,
'hcp-dki',
path_dti)
return subject, True
except Exception as err:
return subject, err.args
else:
return subject, True
def compare_models(subject):
s3 = boto3.resource('s3')
boto3.setup_default_session(profile_name='cirrus')
bucket = s3.Bucket('hcp-dki')
path_dki = '%s/%s_cod_dki.nii.gz' % (subject, subject)
path_dti = '%s/%s_cod_dti.nii.gz' % (subject, subject)
if not (exists(path_dki, bucket.name) and exists(path_dti, bucket.name)):
print("Files don't exist - going ahead")
with tempfile.TemporaryDirectory() as tempdir:
try:
bucket = setup_boto()
dwi_file = op.join(tempdir, 'data.nii.gz')
bvec_file = op.join(tempdir, 'data.bvec')
bval_file = op.join(tempdir, 'data.bval')
data_files = {}
data_files[dwi_file] = \
'HCP_900/%s/T1w/Diffusion/data.nii.gz' % subject
data_files[bvec_file] = \
'HCP_900/%s/T1w/Diffusion/bvecs' % subject
data_files[bval_file] = \
'HCP_900/%s/T1w/Diffusion/bvals' % subject
for k in data_files.keys():
if not op.exists(k):
bucket.download_file(data_files[k], k)
wm_file = op.join(tempdir, 'wm.nii.gz')
boto3.setup_default_session(profile_name='cirrus')
s3 = boto3.resource('s3')
s3.meta.client.download_file(
'hcp-dki',
'%s/%s_white_matter_mask.nii.gz' % (subject, subject),
wm_file)
wm_mask = nib.load(wm_file).get_data().astype(bool)
dwi_img = nib.load(dwi_file)
data = dwi_img.get_data()
bvals = np.loadtxt(bval_file)
bvecs = np.loadtxt(bvec_file)
gtab = dpg.gradient_table(bvals, bvecs, b0_threshold=10)
s3 = boto3.resource('s3')
boto3.setup_default_session(profile_name='cirrus')
bucket = s3.Bucket('hcp-dki')
for model_object, method in zip(
[dti.TensorModel, dki.DiffusionKurtosisModel],
['dti', 'dki']):
path_method = '%s/%s_cod_%s.nii.gz' % (subject, subject,
method)
if not (exists(path_method, bucket.name)):
print("No %s file - fitting" % method)
print("1")
model = model_object(gtab)
print("2")
pred = xval.kfold_xval(model, data, 5, mask=wm_mask)
print("3")
cod = xval.coeff_of_determination(pred, data)
cod_file = op.join(tempdir, 'cod_%s.nii.gz' % method)
print("4")
nib.save(nib.Nifti1Image(cod, dwi_img.affine),
cod_file)
print("5")
s3.meta.client.upload_file(cod_file, 'hcp-dki',
path_method)
return subject, True
except Exception as err:
return subject, err.args
else:
return subject, True
def calc_cod1000(subject):
s3 = boto3.resource('s3')
boto3.setup_default_session(profile_name='cirrus')
bucket = s3.Bucket('hcp-dki')
path_dti = '%s/%s_cod_dti_1000.nii.gz' % (subject, subject)
path_dki = '%s/%s_cod_dki_1000.nii.gz' % (subject, subject)
if not (exists(path_dti, bucket.name) and exists(path_dki, bucket.name)):
print("File doesn't exist - going ahead")
with tempfile.TemporaryDirectory() as tempdir:
try:
bucket = setup_boto()
dwi_file = op.join(tempdir, 'data.nii.gz')
bvec_file = op.join(tempdir, 'data.bvec')
bval_file = op.join(tempdir, 'data.bval')
data_files = {}
data_files[dwi_file] = \
'HCP_900/%s/T1w/Diffusion/data.nii.gz' % subject
data_files[bvec_file] = \
'HCP_900/%s/T1w/Diffusion/bvecs' % subject
data_files[bval_file] = \
'HCP_900/%s/T1w/Diffusion/bvals' % subject
for k in data_files.keys():
if not op.exists(k):
bucket.download_file(data_files[k], k)
wm_file = op.join(tempdir, 'wm.nii.gz')
boto3.setup_default_session(profile_name='cirrus')
s3 = boto3.resource('s3')
bucket = s3.Bucket('hcp-dki')
s3.meta.client.download_file(
'hcp-dki',
'%s/%s_white_matter_mask.nii.gz' % (subject, subject),
wm_file)
wm_mask = nib.load(wm_file).get_data().astype(bool)
dwi_img = nib.load(dwi_file)
data = dwi_img.get_data()
bvals = np.loadtxt(bval_file)
bvecs = np.loadtxt(bvec_file)
idx = bvals < 1985
if not exists(path_dki, bucket.name):
gtab = dpg.gradient_table(bvals, bvecs, b0_threshold=10)
dki_model = dki.DiffusionKurtosisModel(gtab)
# Use all the data to calculate the mode
pred = xval.kfold_xval(dki_model, data, 5, mask=wm_mask)
# But compare only on the b=1000 shell (same as DTI):
cod = xval.coeff_of_determination(pred[..., idx],
data[..., idx])
cod_file = op.join(tempdir, 'cod_dki_1000.nii.gz')
nib.save(nib.Nifti1Image(cod, dwi_img.affine), cod_file)
s3.meta.client.upload_file(cod_file, 'hcp-dki', path_dki)
if not exists(path_dti, bucket.name):
data = data[..., idx]
gtab = dpg.gradient_table(bvals[idx],
bvecs[:, idx].squeeze(),
b0_threshold=10)
model = dti.TensorModel(gtab)
pred = xval.kfold_xval(model, data, 5, mask=wm_mask)
cod = xval.coeff_of_determination(pred, data)
cod_file = op.join(tempdir, 'cod_dti_1000.nii.gz')
nib.save(nib.Nifti1Image(cod, dwi_img.affine), cod_file)
s3.meta.client.upload_file(cod_file, 'hcp-dki', path_dti)
return subject, True
except Exception as err:
return subject, err.args
else:
return subject, True