def test_MultiShellDeconvModel():
gtab = get_3shell_gtab()
S0 = 1.
evals = np.array([.992, .254, .254]) * 1e-3
mevals = np.array([evals, evals])
angles = [(0, 0), (60, 0)]
S_wm, sticks = multi_tensor(gtab,
mevals,
S0,
angles=angles,
fractions=[30., 70.],
snr=None)
S_gm = np.exp(-gtab.bvals * gm_md)
S_csf = np.exp(-gtab.bvals * csf_md)
sh_order = 8
response = sim_response(sh_order, [0, 1000, 2000, 3500])
model = MultiShellDeconvModel(gtab, response)
vf = [1.3, .8, 1.9]
signal = sum(i * j for i, j in zip(vf, [S_csf, S_gm, S_wm]))
fit = model.fit(signal)
npt.assert_array_almost_equal(fit.volume_fractions, vf, 0)
S_pred = fit.predict()
npt.assert_array_almost_equal(S_pred, signal, 0)
def test_MSDeconvFit():
gtab = get_3shell_gtab()
mevals = np.array([wm_response[0, :3], wm_response[0, :3]])
angles = [(0, 0), (60, 0)]
S_wm, sticks = multi_tensor(gtab,
mevals,
wm_response[0, 3],
angles=angles,
fractions=[30., 70.],
snr=None)
S_gm = gm_response[0, 3] * np.exp(-gtab.bvals * gm_response[0, 0])
S_csf = csf_response[0, 3] * np.exp(-gtab.bvals * csf_response[0, 0])
sh_order = 8
response = multi_shell_fiber_response(sh_order, [0, 1000, 2000, 3500],
wm_response, gm_response,
csf_response)
model = MultiShellDeconvModel(gtab, response)
vf = [0.325, 0.2, 0.475]
signal = sum(i * j for i, j in zip(vf, [S_csf, S_gm, S_wm]))
fit = model.fit(signal)
# Testing volume fractions
npt.assert_array_almost_equal(fit.volume_fractions, vf, 1)
def test_mcsd_model_delta():
sh_order = 8
gtab = get_3shell_gtab()
shells = np.unique(gtab.bvals // 100.) * 100.
response = sim_response(sh_order, shells, evals_d)
model = MultiShellDeconvModel(gtab, response)
iso = response.iso
theta, phi = default_sphere.theta, default_sphere.phi
B = shm.real_sph_harm(response.m, response.n, theta[:, None], phi[:, None])
wm_delta = model.delta.copy()
# set isotropic components to zero
wm_delta[:iso] = 0.
wm_delta = _expand(model.m, iso, wm_delta)
for i, s in enumerate(shells):
g = GradientTable(default_sphere.vertices * s)
signal = model.predict(wm_delta, g)
expected = np.dot(response.response[i, iso:], B.T)
npt.assert_array_almost_equal(signal, expected)
signal = model.predict(wm_delta, gtab)
fit = model.fit(signal)
m = model.m
npt.assert_array_almost_equal(fit.shm_coeff[m != 0], 0., 2)
def test_mcsd_model_delta():
sh_order = 8
gtab = get_3shell_gtab()
response = multi_shell_fiber_response(sh_order, [0, 1000, 2000, 3500],
wm_response,
gm_response,
csf_response)
model = MultiShellDeconvModel(gtab, response)
iso = response.iso
theta, phi = default_sphere.theta, default_sphere.phi
B = shm.real_sh_descoteaux_from_index(
response.m, response.n, theta[:, None], phi[:, None])
wm_delta = model.delta.copy()
# set isotropic components to zero
wm_delta[:iso] = 0.
wm_delta = _expand(model.m, iso, wm_delta)
for i, s in enumerate([0, 1000, 2000, 3500]):
g = GradientTable(default_sphere.vertices * s)
signal = model.predict(wm_delta, g)
expected = np.dot(response.response[i, iso:], B.T)
npt.assert_array_almost_equal(signal, expected)
signal = model.predict(wm_delta, gtab)
fit = model.fit(signal)
m = model.m
npt.assert_array_almost_equal(fit.shm_coeff[m != 0], 0., 2)
def test_MSDeconvFit():
gtab = get_3shell_gtab()
mevals = np.array([wm_response[0, :3], wm_response[0, :3]])
angles = [(0, 0), (60, 0)]
S_wm, sticks = multi_tensor(gtab, mevals, wm_response[0, 3], angles=angles,
fractions=[30., 70.], snr=None)
S_gm = gm_response[0, 3] * np.exp(-gtab.bvals * gm_response[0, 0])
S_csf = csf_response[0, 3] * np.exp(-gtab.bvals * csf_response[0, 0])
sh_order = 8
with warnings.catch_warnings():
warnings.filterwarnings(
"ignore", message=shm.descoteaux07_legacy_msg,
category=PendingDeprecationWarning)
response = multi_shell_fiber_response(sh_order, [0, 1000, 2000, 3500],
wm_response,
gm_response,
csf_response)
model = MultiShellDeconvModel(gtab, response)
vf = [0.325, 0.2, 0.475]
signal = sum(i * j for i, j in zip(vf, [S_csf, S_gm, S_wm]))
fit = model.fit(signal)
# Testing volume fractions
npt.assert_array_almost_equal(fit.volume_fractions, vf, 1)
def create_mcsd_model(folder_name, data, gtab, labels, sh_order=8):
from dipy.reconst.mcsd import response_from_mask_msmt
from dipy.reconst.mcsd import MultiShellDeconvModel, multi_shell_fiber_response, MSDeconvFit
from dipy.core.gradients import unique_bvals_tolerance
bvals = gtab.bvals
wm = labels == 3
gm = labels == 2
csf = labels == 1
mask_wm = wm.astype(float)
mask_gm = gm.astype(float)
mask_csf = csf.astype(float)
response_wm, response_gm, response_csf = response_from_mask_msmt(
gtab, data, mask_wm, mask_gm, mask_csf)
ubvals = unique_bvals_tolerance(bvals)
response_mcsd = multi_shell_fiber_response(sh_order,
bvals=ubvals,
wm_rf=response_wm,
csf_rf=response_csf,
gm_rf=response_gm)
mcsd_model = MultiShellDeconvModel(gtab, response_mcsd)
mcsd_fit = mcsd_model.fit(data)
sh_coeff = mcsd_fit.all_shm_coeff
nan_count = len(np.argwhere(np.isnan(sh_coeff[..., 0])))
coeff = mcsd_fit.all_shm_coeff
n_vox = coeff.shape[0] * coeff.shape[1] * coeff.shape[2]
if nan_count > 0:
print(
f'{nan_count / n_vox} of the voxels did not complete fodf calculation, NaN values replaced with 0'
)
coeff = np.where(np.isnan(coeff), 0, coeff)
mcsd_fit = MSDeconvFit(mcsd_model, coeff, None)
np.save(folder_name + r'\coeff.npy', coeff)
return mcsd_fit
def _msmt_ft(self):
from dipy.reconst.mcsd import response_from_mask_msmt
from dipy.reconst.mcsd import MultiShellDeconvModel, multi_shell_fiber_response, MSDeconvFit
from dipy.core.gradients import unique_bvals_tolerance
bvals = self.gtab.bvals
wm = self.tissue_labels == 2
gm = self.tissue_labels == 1
csf = self.tissue_labels == 3
mask_wm = wm.astype(float)
mask_gm = gm.astype(float)
mask_csf = csf.astype(float)
response_wm, response_gm, response_csf = response_from_mask_msmt(
self.gtab, self.data, mask_wm, mask_gm, mask_csf)
ubvals = unique_bvals_tolerance(bvals)
response_mcsd = multi_shell_fiber_response(
self.parameters_dict['sh_order'],
bvals=ubvals,
wm_rf=response_wm,
csf_rf=response_csf,
gm_rf=response_gm)
mcsd_model = MultiShellDeconvModel(self.gtab, response_mcsd)
mcsd_fit = mcsd_model.fit(self.data)
sh_coeff = mcsd_fit.all_shm_coeff
nan_count = len(np.argwhere(np.isnan(sh_coeff[..., 0])))
coeff = mcsd_fit.all_shm_coeff
n_vox = coeff.shape[0] * coeff.shape[1] * coeff.shape[2]
if nan_count > 0:
print(
f'{nan_count / n_vox} of the voxels did not complete fodf calculation, NaN values replaced with 0'
)
coeff = np.where(np.isnan(coeff), 0, coeff)
mcsd_fit = MSDeconvFit(mcsd_model, coeff, None)
self.model_fit = mcsd_fit
mask_wm = wm.astype(float)
mask_gm = gm.astype(float)
mask_csf = csf.astype(float)
response_wm, response_gm, response_csf = response_from_mask_msmt(
gtab, data, mask_wm, mask_gm, mask_csf)
ubvals = unique_bvals_tolerance(bvals)
response_mcsd = multi_shell_fiber_response(sh_order=8,
bvals=ubvals,
wm_rf=response_wm,
csf_rf=response_csf,
gm_rf=response_gm)
mcsd_model = MultiShellDeconvModel(gtab, response_mcsd)
mcsd_fit = mcsd_model.fit(denoised_arr)
sh_coeff = mcsd_fit.all_shm_coeff
nan_count = len(np.argwhere(np.isnan(sh_coeff[..., 0])))
coeff = mcsd_fit.all_shm_coeff
n_vox = coeff.shape[0] * coeff.shape[1] * coeff.shape[2]
print(
f'{nan_count/n_vox} of the voxels did not complete fodf calculation, NaN values replaced with 0'
)
coeff = np.where(np.isnan(coeff), 0, coeff)
mcsd_fit = MSDeconvFit(mcsd_model, coeff, None)
np.save(folder_name + r'\coeff.npy', coeff)
gtab, data, affine, labels, white_matter, nii_file, bvec_file = load_dwi_files(
folder_name)
fa, classifier = create_fa_classifier(gtab, data, white_matter)
lab_labels_index = nodes_by_index_general(folder_name, atlas='yeo7_200')[0]
def mcsd_mod_est(gtab,
data,
B0_mask,
wm_in_dwi,
gm_in_dwi,
vent_csf_in_dwi,
sh_order=8,
roi_radii=10):
"""
Estimate a Constrained Spherical Deconvolution (CSD) model from dwi data.
Parameters
----------
gtab : Obj
DiPy object storing diffusion gradient information.
data : array
4D numpy array of diffusion image data.
B0_mask : str
File path to B0 brain mask.
sh_order : int
The order of the SH model. Default is 8.
Returns
-------
csd_mod : ndarray
Coefficients of the csd reconstruction.
model : obj
Fitted csd model.
References
----------
.. [1] Tournier, J.D., et al. NeuroImage 2007. Robust determination of
the fibre orientation distribution in diffusion MRI:
Non-negativity constrained super-resolved spherical
deconvolution
.. [2] Descoteaux, M., et al. IEEE TMI 2009. Deterministic and
Probabilistic Tractography Based on Complex Fibre Orientation
Distributions
.. [3] Côté, M-A., et al. Medical Image Analysis 2013. Tractometer:
Towards validation of tractography pipelines
.. [4] Tournier, J.D, et al. Imaging Systems and Technology
2012. MRtrix: Diffusion Tractography in Crossing Fiber Regions
"""
import dipy.reconst.dti as dti
from nilearn.image import math_img
from dipy.core.gradients import unique_bvals_tolerance
from dipy.reconst.mcsd import (mask_for_response_msmt,
response_from_mask_msmt,
multi_shell_fiber_response,
MultiShellDeconvModel)
print("Reconstructing using MCSD...")
B0_mask_data = np.nan_to_num(np.asarray(
nib.load(B0_mask).dataobj)).astype("bool")
# Load tissue maps and prepare tissue classifier
gm_mask_img = math_img("img > 0.10", img=gm_in_dwi)
gm_data = np.asarray(gm_mask_img.dataobj, dtype=np.float32)
wm_mask_img = math_img("img > 0.15", img=wm_in_dwi)
wm_data = np.asarray(wm_mask_img.dataobj, dtype=np.float32)
vent_csf_in_dwi_img = math_img("img > 0.50", img=vent_csf_in_dwi)
vent_csf_in_dwi_data = np.asarray(vent_csf_in_dwi_img.dataobj,
dtype=np.float32)
# Fit a simple DTI model
tenfit = dti.TensorModel(gtab).fit(data)
# Obtain the FA and MD metrics
FA = tenfit.fa
MD = tenfit.md
indices_csf = np.where(((FA < 0.2) & (vent_csf_in_dwi_data > 0.50)))
indices_gm = np.where(((FA < 0.2) & (gm_data > 0.10)))
indices_wm = np.where(((FA >= 0.2) & (wm_data > 0.15)))
selected_csf = np.zeros(FA.shape, dtype='bool')
selected_gm = np.zeros(FA.shape, dtype='bool')
selected_wm = np.zeros(FA.shape, dtype='bool')
selected_csf[indices_csf] = True
selected_gm[indices_gm] = True
selected_wm[indices_wm] = True
mask_wm, mask_gm, mask_csf = mask_for_response_msmt(
gtab,
data,
roi_radii=roi_radii,
wm_fa_thr=np.nanmean(FA[selected_wm]),
gm_fa_thr=np.nanmean(FA[selected_gm]),
csf_fa_thr=np.nanmean(FA[selected_csf]),
gm_md_thr=np.nanmean(MD[selected_gm]),
csf_md_thr=np.nanmean(MD[selected_csf]))
mask_wm *= wm_data.astype('int64')
mask_gm *= gm_data.astype('int64')
mask_csf *= vent_csf_in_dwi_data.astype('int64')
# nvoxels_wm = np.sum(mask_wm)
# nvoxels_gm = np.sum(mask_gm)
# nvoxels_csf = np.sum(mask_csf)
response_wm, response_gm, response_csf = response_from_mask_msmt(
gtab, data, mask_wm, mask_gm, mask_csf)
response_mcsd = multi_shell_fiber_response(sh_order=8,
bvals=unique_bvals_tolerance(
gtab.bvals),
wm_rf=response_wm,
gm_rf=response_gm,
csf_rf=response_csf)
model = MultiShellDeconvModel(gtab, response_mcsd, sh_order=sh_order)
mcsd_mod = model.fit(data, B0_mask_data).shm_coeff
mcsd_mod = np.clip(mcsd_mod, 0, np.max(mcsd_mod, -1)[..., None])
del response_mcsd, B0_mask_data
return mcsd_mod.astype("float32"), model
""" response = np.array([response_wm, response_gm, response_csf]) mcsd_model_simple_response = MultiShellDeconvModel(gtab, response, sh_order=8) """ Note that this technique only works for a 3 compartments model (wm, gm, csf). If one wants more compartments, a custom MultiShellResponse object must be used. It can be inspired by the ``multi_shell_fiber_response`` method. """ """ Now we build the MSMT-CSD model with the ``response_mcsd`` as input. We then call the ``fit`` function to fit one slice of the 3D data and visualize it. """ mcsd_model = MultiShellDeconvModel(gtab, response_mcsd) mcsd_fit = mcsd_model.fit(denoised_arr[:, :, 10:11]) """ The volume fractions of tissues for each voxel are also accessible, as well as the sh coefficients for all tissues. One can also get each sh tissue separately using ``all_shm_coeff`` for each compartment (isotropic) and ``shm_coeff`` for white matter. """ vf = mcsd_fit.volume_fractions sh_coeff = mcsd_fit.all_shm_coeff csf_sh_coeff = sh_coeff[..., 0] gm_sh_coeff = sh_coeff[..., 1] wm_sh_coeff = mcsd_fit.shm_coeff """ The model allows to predict a signal from sh coefficients. There are two ways of doing this.