def nyu_dataset(filename):
hf = h5py.File(filename)
print('Keys:', list(hf.keys()))
print('Attrs:', dict(hf.attrs))
volume_kspace = hf['kspace']
print(volume_kspace[20].shape
) # Using slice number 20 of the volume of specific MRI scans
slice_kspace = volume_kspace
show_slices((slice_kspace), [20], cmap='gray')
plt.imshow(np.log(np.abs(slice_kspace[20])))
plt.show()
# Data Wrangling to add dimension of 1 for number of coils as needed for following methods
test = volume_kspace[20]
test2 = test[np.newaxis, :, :]
# Generating sensitivity maps using ESPIRiT method
mps = mr.app.EspiritCalib(test2).run()
pl.ImagePlot(mps, title='Sensitivity Maps Estimated by ESPIRiT')
# Running SENSE reconstruction algorithm on partial k-space image
lamda = 0.01
img_sense = mr.app.SenseRecon(test2, mps, lamda=lamda).run()
pl.ImagePlot(img_sense, title='SENSE Reconstruction')
# Running L1 Wavelet Regularized Reconstruction on partial k-space, images inconclusive
lamda = 0.005
img_l1wav = mr.app.L1WaveletRecon(test2, mps, lamda).run()
pl.ImagePlot(img_l1wav, title='L1 Wavelet Regularized Reconstruction')
import sigpy as sp import sigpy.mri as mr import sigpy.mri.rf as rf import numpy as np import sigpy.plot as pl import matplotlib dim = 32 Nc = 8 img_shape = [dim, dim] sens_shape = [Nc, dim, dim] sens = mr.birdcage_maps(sens_shape) pl.ImagePlot(sens) fov = 0.55 # FOV in m N = dim # matrix size gts = 6.4e-6 # hardware dwell time, s gslew = 150 # gradient slew rate in mT/m/ms gamp = 30 # maximum gradient amplitude in mT/m densamp = 10000 # duration of full density sampling (in samples) dentrans = 10000 # duration of transition from low-high density (in samples) R = 1 / 2 # degree of undersampling of outer region of trajectory- let's oversample by a factor of 2 dx = 0.025 # in m rewinder = False # construct a trajectory g, k, t, s = rf.spiral_arch(fov / R, dx, gts, gslew, gamp) #Note that this trajectory is a spiral-out trajectory. #We will simply time-reverse it to create a spiral-in. k = np.flipud(k) g = np.flipud(g)
import numpy as np
import sigpy.plot as pl
name="../data/cg_img.npy"
name="./cg_img.npy"
img=np.load(name)
print(img.shape)
for t in range(len(img)):
s = img[t, ..., ::-1]
pl.ImagePlot(img[t, ..., ::-1], interpolation='lanczos')
print(f'K-space dtype: {ksp.dtype}')
print(f'K-space (min, max): ({np.abs(ksp).min()}, {np.abs(ksp).max()})')
print(f'Coord shape: {coord.shape}') # (na, ns, 2)
print(f'Coord shape: {coord.dtype}')
print(f'Coord (min, max): ({coord.min()}, {coord.max()})')
plt.ion()
f, ax = plt.subplots(1, 1)
ax.scatter(coord[:15, :, -1], coord[:15, :, -2])
# Use JSENSE to estimate sensitivity maps
mps = mr.app.JsenseRecon(ksp, coord=coord, device=device).run()
print(f'Shape of coil sensitivity maps: {mps.shape}')
pl.ImagePlot(mps)
# Primal dual hybrid gradient reconstruction
pdhg_app = mr.app.TotalVariationRecon(ksp,
mps,
lamda=lamda,
coord=coord,
max_iter=max_iter,
device=device,
save_objective_values=True)
print(f'Name of solver: {pdhg_app.alg_name}')
pdhg_img = pdhg_app.run()
print(f'Image shape: {pdhg_img.shape}')
print(f'Image dtype: {pdhg_img.dtype}')
import sigpy as sp
import sigpy.mri as mr
import sigpy.plot as pl
from wshfl import WaveShuffling
# -------------------------------------------------------------------------------------------------------------------------- #
rdr = np.load('data/rdr.npy')
tbl = np.load('data/tbl.npy')
mps = np.load('data/mps.npy').T
psf = np.load('data/psf.npy').T
phi = np.load('data/phi.npy').T
mit = 300
sparse_repr = 'W'
# -------------------------------------------------------------------------------------------------------------------------- #
start = time.time()
Waffle = WaveShuffling(rdr, tbl, mps, psf, phi, spr=sparse_repr, lmb=1e-6, mit=mit, dev=0)
Waffle.run()
end = time.time()
print("Device used: " + str(Waffle.device) + ". Reconstruction took " + str(end - start) + " seconds.")
# -------------------------------------------------------------------------------------------------------------------------- #
pl.ImagePlot(Waffle.S.H(Waffle.res).squeeze(), x=1, y=2, z=0, hide_axes=False)
# -------------------------------------------------------------------------------------------------------------------------- #
coord = xp.load(coord_file)
def show_data_info(data, name):
print("{}: shape={}, dtype={}".format(name, data.shape, data.dtype))
dcf = (coord[..., 0]**2 + coord[..., 1]**2)**0.5
pl.ScatterPlot(coord, dcf, title='Density compensation')
show_data_info(ksp, "ksp")
show_data_info(coord, "coord")
show_data_info(dcf, "dcf")
img_grid = sp.nufft_adjoint(ksp * dcf, coord)
pl.ImagePlot(img_grid, z=0, title='Multi-channel Gridding')
#%% md
## Estimate sensitivity maps using JSENSE
# Here we use [JSENSE](https://onlinelibrary.wiley.com/doi/full/10.1002/mrm.21245) to estimate sensitivity maps.
#%%
mps = mr.app.JsenseRecon(ksp, coord=coord, device=device).run()
#%% md
## CG
## FISTA
#%%
fista_app = mr.app.L1WaveletRecon(ksp,
mps,
lamda=lamda,
coord=coord,
device=device,
max_iter=max_iter,
save_objective_values=True)
fista_img = fista_app.run()
pl.ImagePlot(fista_img)
#%% md
## ADMM
#%%
'''
admm_app = mr.app.L1WaveletRecon(
ksp, mps, solver='ADMM', lamda=lamda, coord=coord, device=device,
max_iter=max_iter // max_cg_iter, max_cg_iter=max_cg_iter, save_objective_values=True)
admm_img = admm_app.run()
pl.ImagePlot(admm_img)
'''
if print_cost:
toc = time.time()
print('EPOCH = ', epoch, 'COST = ', minibatch_loss,
'Elapsed time = ', (toc - tic))
if epoch % 200 == 0:
save_path = saver.save(sess,
"model/model_maniflod_spiral.ckpt")
print("Model saved in file: %s" % save_path)
Y_opt = np.array(
sess.run(DECONV, feed_dict={
X: X_train,
Y: Y_train
}))
#Y_opt = Y_opt.eval(session = sess)
sess.close()
return Y_opt
Y_test = forward_model(
X_train,
Y_train,
learning_rate=0.00001,
num_epochs=1000,
minibatch_size=2, # should be < than the number of input examples
print_cost=True)
pl.ImagePlot(Y_test)
np.save('data_spiral.npy', Y_test)
# Y_test = manifold_net(X_train).eval()
W = sp.linop.Wavelet(img_shape)
wav = W * S.H * F.H * ksp
#pl.ImagePlot(wav**0.1, title=r'$W S^H F^H y$')
print(np.amax(np.abs(wav)))
print(np.amin(np.abs(wav)))
print(np.shape(wav))
plt.figure(1)
lala = ksp[0, :, :, 160]
print(np.shape(lala))
plt.imshow(np.abs(wav[:, :, 160]))
plt.clim(0.0001, 0.001)
plt.show()
pl.ImagePlot(wav, title=r'$W S^H F^H y$')
A = P * F * S * W.H
## Prox
print("Define Prox")
lamda = 0.005
proxg = sp.prox.L1Reg(wav.shape, lamda)
alpha = 1
wav_thresh = proxg(alpha, wav)
pl.ImagePlot(wav_thresh**0.1)
## Alg
print("Define Alg")
#coord = np.load(dir+'coord.npy').transpose((1,0,2))
ksp = np.load(dir+'ksp.npy')
coord = np.load(dir+'coord.npy')
ksp = ksp.transpose((2,1,0))
coord = coord.transpose((1,0,2))*96
print("estimate shape=", estimate_shape(coord))
#dcf = (coord[..., 0]**2 + coord[..., 1]**2+ coord[..., 2]**2)**0.5
show_data_info(ksp, "ksp")
show_data_info(coord, "coord")
#show_data_info(dcf, "dcf")
#ksp = np.stack((ksp.real, ksp.imag), axis=-1)
#ksp = np.stack((ksp.real, ksp.imag), axis=-1).astype(np.double)
mps = mr.app.JsenseRecon(ksp, coord=coord, device=device).run()
#mps = mr.app.JsenseRecon(ksp, device=device).run()
cg_app = mr.app.SenseRecon(
ksp, mps, coord=coord, device=device, lamda=lamda,
max_iter=max_iter, save_objective_values=True)
cg_img = cg_app.run()
np.save("cg_img.npy", cg_img)
pl.ImagePlot(cg_img)
#%%
mps = mr.app.JsenseRecon(ksp, coord=coord, device=device).run()
#%% md
## ADMM
#%%
admm_app = mr.app.TotalVariationRecon(
ksp, mps, lamda=lamda, coord=coord, max_iter=max_iter // max_cg_iter,
solver='ADMM', max_cg_iter=max_cg_iter, device=device, save_objective_values=True)
admm_img = admm_app.run()
pl.ImagePlot(admm_img)
#%% md
## ADMM with circulant preconditioner
#%%
rho = 1
circ_precond = mr.circulant_precond(mps, coord=coord, device=device, lamda=rho)
img_shape = mps.shape[1:]
G = sp.linop.FiniteDifference(img_shape)
g = G.H * G * sp.dirac(img_shape)
g = sp.fft(g)
g = sp.to_device(g, device=device)
mesh[:, :, 1] = m2
return mesh.astype(np.float)
name = 'img.jpg'
image = Image.open(name).convert('L')
arr = np.array(image) + 1j
traj = cartisian2D(arr.shape, [1, 1], 1)
plt.ScatterPlot(traj, title='Trajectory')
image.close()
arr = arr / np.max(arr[...])
print(traj.shape)
kspaceNUFFT = sp.nufft(arr, traj)
plt.ImagePlot(np.log(kspaceNUFFT), title='k-space data from NUFFT')
kspaceFFT = sp.fft(arr)
plt.ImagePlot(np.log(kspaceFFT), title='k-space data from FFT')
print(kspaceFFT.shape)
print(kspaceNUFFT.shape)
sumNUFFT = np.sum(kspaceNUFFT)
sumFFT = np.sum(kspaceFFT)
if (np.allclose(kspaceNUFFT, kspaceFFT, rtol=10, atol=10)
and np.isclose(sumNUFFT, sumFFT, rtol=50, atol=50)):
print('Outputs are similar!')
else:
print('Outputs are NOT similar!')