def run_expmt():
for file_id in file_id_list:
f, ksp_orig = load_h5(file_id)
ksp_orig = torch.from_numpy(ksp_orig)
mask = get_mask(ksp_orig)
net, net_input, ksp_orig_, _ = init_convdecoder(ksp_orig, mask)
ksp_masked = SCALE_FAC * ksp_orig_ * mask
img_masked = ifft_2d(ksp_masked)
net, mse_wrt_ksp, mse_wrt_img = fit(ksp_masked=ksp_masked,
img_masked=img_masked,
net=net,
net_input=net_input,
mask2d=mask,
num_iter=NUM_ITER)
img_out, _ = net(net_input.type(dtype))
img_out = reshape_adj_channels_to_complex_vals(img_out[0])
ksp_est = fft_2d(img_out)
ksp_dc = torch.where(mask, ksp_masked, ksp_est)
img_est = crop_center(
root_sum_squares(ifft_2d(ksp_est)).detach(), dim, dim)
img_dc = crop_center(
root_sum_squares(ifft_2d(ksp_dc)).detach(), dim, dim)
img_gt = crop_center(root_sum_squares(ifft_2d(ksp_orig)), dim, dim)
np.save('{}{}_est.npy'.format(path_out, file_id), img_est)
np.save('{}{}_dc.npy'.format(path_out, file_id), img_dc)
np.save('{}{}_gt.npy'.format(path_out, file_id), img_gt)
def run_expmt(file_id_list):
for file_id in file_id_list:
if os.path.exists('{}{}_dc.npy'.format(path_out, file_id)):
continue
f, ksp_orig = load_h5_fastmri(file_id)
ksp_orig = torch.from_numpy(ksp_orig)
mask = get_mask(ksp_orig)
net, net_input, ksp_orig_ = init_convdecoder(ksp_orig, mask)
ksp_masked = 0.1 * ksp_orig_ * mask
net = fit(ksp_masked, net, net_input, mask)
img_out = net(net_input.type(dtype))
img_out = reshape_adj_channels_to_complex_vals(img_out[0])
ksp_est = fft_2d(img_out)
ksp_dc = torch.where(mask, ksp_masked, ksp_est)
#img_est = crop_center(root_sum_squares(ifft_2d(ksp_est)).detach(), dim, dim)
img_dc = crop_center(root_sum_squares(ifft_2d(ksp_dc)).detach(), dim, dim)
img_gt = crop_center(root_sum_squares(ifft_2d(ksp_orig)), dim, dim)
# note: use unscaled ksp_orig to make gt -- different from original processing
np.save('{}{}_dc.npy'.format(path_out, file_id), img_dc)
np.save('{}{}_gt.npy'.format(path_out, file_id), img_gt)
def run_demo():
ksp_orig = load_h5_fastmri(file_id=None, demo=True)
mask = get_mask(ksp_orig)
net, net_input, ksp_orig_ = init_convdecoder(ksp_orig)
ksp_masked = 0.1 * ksp_orig_ * mask
net = fit(ksp_masked, net, net_input, mask)
img_out = net(net_input.type(dtype))
img_out = reshape_adj_channels_to_complex_vals(img_out[0])
ksp_est = fft_2d(img_out)
ksp_dc = torch.where(mask, ksp_masked, ksp_est)
img_dc = crop_center(root_sum_squares(ifft_2d(ksp_dc)).detach(), dim, dim)
img_gt = crop_center(root_sum_squares(ifft_2d(ksp_orig)), dim, dim)
img_zf = crop_center(root_sum_squares(ifft_2d(ksp_masked)), dim, dim)
np.save('data/out.npy', img_dc)
np.save('data/gt.npy', img_gt)
np.save('data/zf.npy', img_zf)
def run_expmt(args):
for file_id in args.file_id_list:
ksp_orig = load_qdess(
file_id, idx_kx=None) # default central slice in kx (axial)
for accel in args.accel_list:
# manage paths for input/output
path_base = '/bmrNAS/people/dvv/out_qdess/accel_{}x/'.format(accel)
path_out = '{}{}/'.format(path_base, args.dir_out)
args.path_gt = path_base + 'gt/'
if os.path.exists('{}MTR_{}_e1.npy'.format(path_out, file_id)):
continue
if not os.path.exists(path_out):
os.makedirs(path_out)
if not os.path.exists(args.path_gt):
os.makedirs(args.path_gt)
# initialize network
net1, net_input1, ksp_orig_ = init_convdecoder(
ksp_orig, fix_random_seed=False)
net2, net_input2, _ = init_convdecoder(ksp_orig,
fix_random_seed=False)
# apply mask after rescaling k-space. want complex tensors dim (nc, ky, kz)
ksp_masked, mask = apply_mask(ksp_orig_,
accel,
calib=args.calib,
expmt=True)
# fit network, get net output - default 10k iterations, lam_tv=1e-8
net1, net2 = fit(ksp_masked=ksp_masked,
net1=net1,
net_input1=net_input1,
net2=net2,
net_input2=net_input2,
mask=mask)
im_out1 = net1(
net_input1.type(dtype)) # real tensor dim (2*nc, kx, ky)
im_out2 = net2(
net_input2.type(dtype)) # real tensor dim (2*nc, kx, ky)
im_out = torch.mean(torch.stack([im_out1, im_out2]), dim=0)
im_out = reshape_adj_channels_to_complex_vals(
im_out[0]) # complex tensor dim (nc, kx, ky)
# perform dc step
ksp_est = fft_2d(im_out)
ksp_dc = torch.where(mask, ksp_masked, ksp_est)
np.save('{}/MTR_{}_ksp_dc.npy'.format(path_out, file_id),
ksp_dc.detach().numpy())
# create data-consistent, ground-truth images from k-space
im_1_dc = root_sum_squares(ifft_2d(ksp_dc[:8])).detach()
im_2_dc = root_sum_squares(ifft_2d(ksp_dc[8:])).detach()
np.save('{}MTR_{}_e1.npy'.format(path_out, file_id), im_1_dc)
np.save('{}MTR_{}_e2.npy'.format(path_out, file_id), im_2_dc)
# save gt w proper array scaling if dne
if not os.path.exists('{}MTR_{}_e1_gt.npy'.format(
args.path_gt, file_id)):
im_1_gt = root_sum_squares(ifft_2d(ksp_orig[:8]))
im_2_gt = root_sum_squares(ifft_2d(ksp_orig[8:]))
np.save('{}MTR_{}_e1_gt.npy'.format(args.path_gt, file_id),
im_1_gt)
np.save('{}MTR_{}_e2_gt.npy'.format(args.path_gt, file_id),
im_2_gt)
print('recon {}'.format(file_id))
return
def run_expmt():
path_in = '/bmrNAS/people/arjun/data/qdess_knee_2020/files_recon_calib-16/'
#files = [f for f in listdir(path_in) if isfile(join(path_in, f))]
#files.sort()
#NUM_SAMPS = 10 # number of samples to recon
NUM_ITER = 10000
ACCEL_LIST = [8] # 4, 6, 8]
for fn in test_set: #files[:NUM_SAMPS]:
# load data
f = h5py.File(path_in + fn, 'r')
try:
ksp = torch.from_numpy(f['kspace'][()])
except KeyError:
print('No kspace in file {} w keys {}'.format(fn, f.keys()))
f.close()
continue
f.close()
# NOTE: if change to echo2, must manually change path nomenclature
ksp_vol = ksp[:,:,:,0,:].permute(3,0,1,2) # get echo1, reshape to be (nc, kx, ky, kz)
# get central slice in kx, i.e. axial plane b/c we undersample in (ky, kz)
idx_kx = ksp_vol.shape[1] // 2
ksp_orig = ksp_vol[:, idx_kx, :, :]
for ACCEL in ACCEL_LIST:
path_out = '/bmrNAS/people/dvv/out_qdess/accel_{}x/echo1/'.format(ACCEL)
# original masks created w central region 32x32 forced to 1's
mask = torch.from_numpy(np.load('ipynb/masks/mask_poisson_disc_{}x.npy'.format(ACCEL)))
# initialize network
net, net_input, ksp_orig_, _ = init_convdecoder(ksp_orig, mask)
# apply mask after rescaling k-space. want complex tensors dim (nc, ky, kz)
ksp_masked = ksp_orig_ * mask
img_masked = ifft_2d(ksp_masked)
# fit network, get net output
net, mse_wrt_ksp, mse_wrt_img = fit(
ksp_masked=ksp_masked, img_masked=img_masked,
net=net, net_input=net_input, mask2d=mask, num_iter=NUM_ITER)
img_out, _ = net(net_input.type(dtype)) # real tensor dim (2*nc, kx, ky)
img_out = reshape_adj_channels_to_complex_vals(img_out[0]) # complex tensor dim (nc, kx, ky)
# perform dc step
ksp_est = fft_2d(img_out)
ksp_dc = torch.where(mask, ksp_masked, ksp_est)
# create data-consistent, ground-truth images from k-space
img_dc = root_sum_squares(ifft_2d(ksp_dc)).detach()
img_gt = root_sum_squares(ifft_2d(ksp_orig))
# save results
samp = fn.split('.h5')[0] #+ '_echo2'
np.save('{}{}_dc.npy'.format(path_out, samp), img_dc)
np.save('{}{}_gt.npy'.format(path_out, samp), img_gt)
print('recon {}'.format(samp))
return
