def load_masks_cfl(filenames, image_shape=None):
"""Read masks from files."""
if image_shape is None:
# First find masks shape...
image_shape = [0, 0]
for f in filenames:
f_cfl = os.path.splitext(f)[0]
mask = np.squeeze(cfl.read(f_cfl))
shape_z = mask.shape[-2]
shape_y = mask.shape[-1]
if image_shape[-2] < shape_z:
image_shape[-2] = shape_z
if image_shape[-1] < shape_y:
image_shape[-1] = shape_y
masks = np.zeros([len(filenames)] + image_shape, dtype=np.complex64)
i_file = 0
for f in filenames:
f_cfl = os.path.splitext(f)[0]
tmp = np.squeeze(cfl.read(f_cfl))
tmp = recon.zeropad(tmp, image_shape)
masks[i_file, :, :] = tmp
i_file = i_file + 1
return masks
def bart_generate_mask(
shape,
acc,
variable_density=True,
shape_calib=10,
verbose=False,
tmp_file="mask.tmp",
):
"""Use bart poisson to generate masks."""
if verbose:
print("Generating sampling mask...")
random_seed = 1e6 * random.random()
flags = "-Z %d -Y %d -z %g -y %g -C %d -s %d" % (
shape[0],
shape[1],
acc[0],
acc[1],
shape_calib,
random_seed,
)
if variable_density:
flags = flags + " -v"
cmd = "bart poisson %s %s" % (flags, tmp_file)
if verbose:
print(" %s" % cmd)
subprocess.check_output(["bash", "-c", cmd])
mask = np.abs(np.squeeze(cfl.read(tmp_file)))
return mask
def bart_pics(
ks_input,
verbose=False,
sensemap=None,
shape_e=2,
do_cs=True,
do_imag_reg=False,
filename_ks_tmp="ks.tmp",
filename_map_tmp="map.tmp",
filename_im_tmp="im.tmp",
filename_ks_out_tmp="ks_out.tmp",
):
"""BART PICS reconstruction."""
if verbose:
print("PICS (l1-ESPIRiT) reconstruction...")
cfl.write(filename_ks_tmp, ks_input)
if sensemap is None:
cmd = "bart ecalib -m %d -c 1e-9 %s %s" % (
shape_e,
filename_ks_tmp,
filename_map_tmp,
)
if verbose:
print(" %s" % cmd)
subprocess.check_output(["bash", "-c", cmd])
else:
cfl.write(filename_map_tmp, sensemap)
if do_cs:
flags = "-l1 -r 1e-2"
else:
flags = "-l2 -r 1e-2"
if do_imag_reg:
flags = flags + " -R R1:7:1e-1"
cmd = "bart pics %s -S %s %s %s" % (
flags,
filename_ks_tmp,
filename_map_tmp,
filename_im_tmp,
)
if verbose:
print(" %s" % cmd)
subprocess.check_output(["bash", "-c", cmd])
cmd = "bart fakeksp -r %s %s %s %s" % (
filename_im_tmp,
filename_ks_tmp,
filename_map_tmp,
filename_ks_out_tmp,
)
if verbose:
print(" %s" % cmd)
subprocess.check_output(["bash", "-c", cmd])
ks_pics = np.squeeze(cfl.read(filename_ks_out_tmp))
ks_pics = np.expand_dims(ks_pics, axis=0)
return ks_pics
def load_recon(file, file_sensemap):
bart_recon = np.squeeze(cfl.read(file))
if bart_recon.ndim == 2:
bart_recon = np.transpose(bart_recon, [1, 0])
bart_recon = np.expand_dims(bart_recon, axis=0)
bart_recon = np.expand_dims(bart_recon, axis=-1)
if bart_recon.ndim == 3:
bart_recon = np.transpose(bart_recon, [2, 1, 0])
bart_recon = np.expand_dims(bart_recon, axis=-1)
return bart_recon
def load_recon(file, file_sensemap):
bart_recon = np.squeeze(cfl.read(file))
# print("bart recon")
# print(bart_recon.ndim)
# 18, 80, 180
if bart_recon.ndim == 2:
bart_recon = np.transpose(bart_recon, [1, 0])
bart_recon = np.expand_dims(bart_recon, axis=0)
bart_recon = np.expand_dims(bart_recon, axis=-1)
if bart_recon.ndim == 3:
bart_recon = np.transpose(bart_recon, [2, 1, 0])
bart_recon = np.expand_dims(bart_recon, axis=-1)
# print(bart_recon.shape)
return bart_recon
def bart_espirit(
ks_input,
shape=None,
verbose=False,
shape_e=2,
crop_value=None,
cal_size=None,
smooth=False,
filename_ks_tmp="ks.tmp",
filename_map_tmp="map.tmp",
):
"""Estimate sensitivity maps using BART ESPIRiT.
ks_input dimensions: [emaps, channels, kz, ky, kx]
"""
if verbose:
print("Estimating sensitivity maps...")
if shape is not None:
ks_input = recon.crop(ks_input, [-1, -1, shape[0], shape[1], -1])
ks_input = recon.zeropad(ks_input, [-1, -1, shape[0], shape[1], -1])
flags = ""
if crop_value is not None:
flags = flags + "-c %f " % crop_value
if cal_size is not None:
flags = flags + "-r %d " % cal_size
if smooth:
flags = flags + "-S "
cfl.write(filename_ks_tmp, ks_input)
cmd = "bart ecalib -m %d %s %s %s" % (
shape_e,
flags,
filename_ks_tmp,
filename_map_tmp,
)
if verbose:
print(" %s" % cmd)
time_start = timer()
subprocess.check_output(["bash", "-c", cmd])
time_end = timer()
sensemap = cfl.read(filename_map_tmp)
return sensemap, time_end - time_start
def setup_data_tfrecords(
dir_in_root,
dir_out,
data_divide=(0.75, 0.05, 0.2),
min_shape=[80, 180],
num_maps=1,
crop_maps=False,
verbose=False,
):
"""Setups training data as tfrecords.
prep_data.setup_data('/mnt/raid3/data/Studies_DCE/recon-ccomp6/',
'/mnt/raid3/jycheng/Project/deepspirit/data/train/', verbose=True)
"""
# Check for two echos in here
# Use glob to find if have echo01
if verbose:
print("Directory names:")
print(" Input root: %s" % dir_in_root)
print(" Output root: %s" % dir_out)
file_kspace = "kspace"
file_sensemap = "sensemap"
case_list = os.listdir(dir_in_root)
random.shuffle(case_list)
num_cases = len(case_list)
i_train_1 = np.round(data_divide[0] * num_cases).astype(int)
i_validate_0 = i_train_1 + 1
i_validate_1 = np.round(
data_divide[1] * num_cases).astype(int) + i_validate_0
if not os.path.exists(dir_out):
os.mkdir(dir_out)
if not os.path.exists(os.path.join(dir_out, "train")):
os.mkdir(os.path.join(dir_out, "train"))
if not os.path.exists(os.path.join(dir_out, "validate")):
os.mkdir(os.path.join(dir_out, "validate"))
if not os.path.exists(os.path.join(dir_out, "test")):
os.mkdir(os.path.join(dir_out, "test"))
i_case = 0
for case_name in case_list:
file_kspace_i = os.path.join(dir_in_root, case_name, file_kspace)
file_sensemap_i = os.path.join(dir_in_root, case_name, file_sensemap)
if i_case < i_train_1:
dir_out_i = os.path.join(dir_out, "train")
elif i_case < i_validate_1:
dir_out_i = os.path.join(dir_out, "validate")
else:
dir_out_i = os.path.join(dir_out, "test")
if verbose:
print("Processing [%d] %s..." % (i_case, case_name))
i_case = i_case + 1
kspace = np.squeeze(cfl.read(file_kspace_i))
if (min_shape is None) or (
min_shape[0] <= kspace.shape[1] and min_shape[1] <= kspace.shape[2]
):
if verbose:
print(" Slice shape: (%d, %d)" %
(kspace.shape[1], kspace.shape[2]))
print(" Num channels: %d" % kspace.shape[0])
shape_x = kspace.shape[-1]
kspace = fftc.ifftc(kspace, axis=-1)
kspace = kspace.astype(np.complex64)
# if shape_c_out < shape_c:
# if verbose:
# print(" applying coil compression (%d -> %d)..." %
# (shape_c, shape_c_out))
# shape_cal = 24
# ks_cal = recon.crop(ks, [-1, shape_cal, shape_cal, -1])
# ks_cal = np.reshape(ks_cal, [shape_c,
# shape_cal*shape_cal,
# shape_x])
# cc_mat = coilcomp.calc_gcc_weights_c(ks_cal, shape_c_out)
# ks_cc = np.reshape(ks, [shape_c, -1, shape_x])
# ks_cc = coilcomp.apply_gcc_weights_c(ks_cc, cc_mat)
# ks = np.reshape(ks_cc, [shape_c_out, shape_z, shape_y, shape_x])
cmd_flags = ""
if crop_maps:
cmd_flags = cmd_flags + " -c 1e-9"
cmd_flags = cmd_flags + (" -m %d" % num_maps)
cmd = "%s ecalib %s %s %s" % (
BIN_BART,
cmd_flags,
file_kspace_i,
file_sensemap_i,
)
if verbose:
print(" Estimating sensitivity maps (bart espirit)...")
print(" %s" % cmd)
subprocess.check_call(["bash", "-c", cmd])
sensemap = np.squeeze(cfl.read(file_sensemap_i))
sensemap = np.expand_dims(sensemap, axis=0)
sensemap = sensemap.astype(np.complex64)
if verbose:
print(" Creating tfrecords (%d)..." % shape_x)
for i_x in range(shape_x):
file_out = os.path.join(
dir_out_i, "%s_x%03d.tfrecords" % (case_name, i_x)
)
kspace_x = kspace[:, :, :, i_x]
sensemap_x = sensemap[:, :, :, :, i_x]
example = tf.train.Example(
features=tf.train.Features(
feature={
"name": _bytes_feature(str.encode(case_name)),
"xslice": _int64_feature(i_x),
"ks_shape_x": _int64_feature(kspace.shape[3]),
"ks_shape_y": _int64_feature(kspace.shape[2]),
"ks_shape_z": _int64_feature(kspace.shape[1]),
"ks_shape_c": _int64_feature(kspace.shape[0]),
"map_shape_x": _int64_feature(sensemap.shape[4]),
"map_shape_y": _int64_feature(sensemap.shape[3]),
"map_shape_z": _int64_feature(sensemap.shape[2]),
"map_shape_c": _int64_feature(sensemap.shape[1]),
"map_shape_m": _int64_feature(sensemap.shape[0]),
"ks": _bytes_feature(kspace_x.tostring()),
"map": _bytes_feature(sensemap_x.tostring()),
}
)
)
tf_writer = tf.python_io.TFRecordWriter(file_out)
tf_writer.write(example.SerializeToString())
tf_writer.close()
def setup_data_tfrecords(
dir_in_root,
dir_out,
data_divide=(0.75, 0.05, 0.2),
min_shape=[80, 180],
num_maps=1,
crop_maps=False,
verbose=False,
):
"""Setups training data as tfrecords.
prep_data.setup_data('/mnt/raid3/data/Studies_DCE/recon-ccomp6/',
'/mnt/raid3/jycheng/Project/deepspirit/data/train/', verbose=True)
"""
if verbose:
print("Directory names:")
print(" Input root: %s" % dir_in_root)
print(" Output root: %s" % dir_out)
file_kspace = "kspace"
file_sensemap = "sensemap"
case_list = os.listdir(dir_in_root)
random.shuffle(case_list)
num_cases = len(case_list)
i_train_1 = np.round(data_divide[0] * num_cases).astype(int)
i_validate_0 = i_train_1 + 1
i_validate_1 = np.round(
data_divide[1] * num_cases).astype(int) + i_validate_0
if not os.path.exists(dir_out):
os.mkdir(dir_out)
if not os.path.exists(os.path.join(dir_out, "train")):
os.mkdir(os.path.join(dir_out, "train"))
if not os.path.exists(os.path.join(dir_out, "validate")):
os.mkdir(os.path.join(dir_out, "validate"))
if not os.path.exists(os.path.join(dir_out, "test")):
os.mkdir(os.path.join(dir_out, "test"))
i_case = 0
for case_name in case_list:
file_kspace_i = os.path.join(dir_in_root, case_name, file_kspace)
file_sensemap_i = os.path.join(dir_in_root, case_name, file_sensemap)
if verbose:
print("Processing [%d] %s..." % (i_case, case_name))
if i_case < i_train_1:
dir_out_i = os.path.join(dir_out, "train")
elif i_case < i_validate_1:
dir_out_i = os.path.join(dir_out, "validate")
else:
dir_out_i = os.path.join(dir_out, "test")
i_case = i_case + 1
if not os.path.exists(file_kspace_i + ".hdr"):
print("skipping due to kspace not existing in this folder")
continue
kspace = np.squeeze(cfl.read(file_kspace_i))
print("original kspace shape")
print(kspace.shape)
shape_x = kspace.shape[3]
shape_y = kspace.shape[2]
shape_z = kspace.shape[1]
num_coils = kspace.shape[0]
if num_coils is not 32:
print("skipping due to incorrect number of coils")
continue
if min_shape[0] == kspace.shape[1] and min_shape[1] == kspace.shape[2]:
if verbose:
print(" Slice shape: (%d, %d)" %
(kspace.shape[1], kspace.shape[2]))
print(" Num channels: %d" % kspace.shape[0])
# shape_x = kspace.shape[-1]
# fix +1, -1 modulation along readout direction
# for n in range(shape_x):
# modulation = (-1)**n
# kspace[:,:,:,n] = kspace[:,:,:,n]*np.exp(-1j*modulation)
# print("kspace shape after modulation")
# print(kspace.shape)
# readout in kx
kspace = fftc.ifftc(kspace, axis=-1)
cmd_flags = ""
if crop_maps:
cmd_flags = cmd_flags + " -c 1e-9"
# smoothing flag
cmd_flags = cmd_flags + (" -S")
cmd_flags = cmd_flags + (" -m %d" % num_maps)
cmd = "%s ecalib %s %s %s" % (
BIN_BART,
cmd_flags,
file_kspace_i,
file_sensemap_i,
)
if verbose:
print(" Estimating sensitivity maps (bart espirit)...")
print(" %s" % cmd)
subprocess.check_call(["bash", "-c", cmd])
sensemap = np.squeeze(cfl.read(file_sensemap_i))
sensemap = np.expand_dims(sensemap, axis=0)
sensemap = sensemap.astype(np.complex64)
if verbose:
print(" Creating tfrecords (%d)..." % shape_x)
for i_x in range(shape_x):
file_out = os.path.join(
dir_out_i, "%s_x%03d.tfrecords" % (case_name, i_x))
kspace_x = kspace[:, :, :, i_x]
sensemap_x = sensemap[:, :, :, :, i_x]
example = tf.train.Example(features=tf.train.Features(
feature={
"name": _bytes_feature(str.encode(case_name)),
"xslice": _int64_feature(i_x),
"ks_shape_x": _int64_feature(kspace.shape[3]),
"ks_shape_y": _int64_feature(kspace.shape[2]),
"ks_shape_z": _int64_feature(kspace.shape[1]),
"ks_shape_c": _int64_feature(kspace.shape[0]),
"map_shape_x": _int64_feature(sensemap.shape[4]),
"map_shape_y": _int64_feature(sensemap.shape[3]),
"map_shape_z": _int64_feature(sensemap.shape[2]),
"map_shape_c": _int64_feature(sensemap.shape[1]),
"map_shape_m": _int64_feature(sensemap.shape[0]),
"ks": _bytes_feature(kspace_x.tostring()),
"map": _bytes_feature(sensemap_x.tostring()),
}))
tf_writer = tf.python_io.TFRecordWriter(file_out)
tf_writer.write(example.SerializeToString())
tf_writer.close()
else:
print("skipping due to wrong slice dimensions")
def setup_data_tfrecords_3d(
dir_in_root,
dir_out,
data_divide=(0.8, 0.1, 0.2),
min_shape=[80, 180],
num_maps=1,
crop_maps=False,
verbose=False,
shuffle=True,
):
"""Setups training data as tfrecords.
prep_data.setup_data('/mnt/raid3/data/Studies_DCE/recon-ccomp6/',
'/mnt/raid3/jycheng/Project/deepspirit/data/train/', verbose=True)
"""
# Check for two echos in here
# Use glob to find if have echo01
if verbose:
print("Directory names:")
print(" Input root: %s" % dir_in_root)
print(" Output root: %s" % dir_out)
# edits for /mnt/dense/data/MFAST_DCE and /home_local/ekcole/MFAST_DCE
dir_kspace = "sort-ccomp6"
dir_map = "recon-ccomp6"
dir_cases = os.path.join(dir_in_root, dir_kspace)
file_kspace = "ks_sorted"
file_sensemap = "map"
case_list = os.listdir(dir_cases)
if shuffle is True:
random.shuffle(case_list)
else:
print("don't shuffle dataset")
num_cases = len(case_list)
i_train_1 = np.round(data_divide[0] * num_cases).astype(int)
i_validate_0 = i_train_1 + 1
i_validate_1 = np.round(
data_divide[1] * num_cases).astype(int) + i_validate_0
if not os.path.exists(dir_out):
os.mkdir(dir_out)
if not os.path.exists(os.path.join(dir_out, "train")):
os.mkdir(os.path.join(dir_out, "train"))
if not os.path.exists(os.path.join(dir_out, "validate")):
os.mkdir(os.path.join(dir_out, "validate"))
if not os.path.exists(os.path.join(dir_out, "test")):
os.mkdir(os.path.join(dir_out, "test"))
i_case = 0
for case_name in case_list:
file_kspace_i = os.path.join(dir_in_root, dir_kspace, case_name,
file_kspace)
file_sensemap_i = os.path.join(dir_in_root, dir_map, case_name,
file_sensemap)
file_sensemap_i_check = os.path.join(dir_in_root, dir_map, case_name,
file_sensemap + ".cfl")
if i_case < i_train_1:
dir_out_i = os.path.join(dir_out, "train")
elif i_case < i_validate_1:
dir_out_i = os.path.join(dir_out, "validate")
else:
dir_out_i = os.path.join(dir_out, "test")
print("dir out")
print(dir_out_i)
if verbose:
print("Processing [%d] %s..." % (i_case, case_name))
i_case = i_case + 1
# if no map, skip this case
# do nothing
if not path.exists(file_sensemap_i_check):
print("Sensitivity map does not exist")
continue
# get dims from .hdr
h = open(file_kspace_i + ".hdr", "r")
h.readline() # skip
l = h.readline()
h.close()
dims = [int(i) for i in l.split()]
print(dims)
ky = dims[1]
kz = dims[2]
if ky != 180 or kz != 80:
print("wrong dimensions")
continue
kspace = np.squeeze(cfl.read(file_kspace_i))
# it wants coils x y z frames
kspace = np.transpose(kspace, [1, -1, -2, 2, 0])
if verbose:
print(" Slice shape: (%d, %d)" %
(kspace.shape[2], kspace.shape[3]))
print(" Num channels: %d" % kspace.shape[0])
print(" Num frames: %d" % kspace.shape[-1])
# number of frames
shape_f = kspace.shape[-1]
# number of slices in x direction
num_slices = kspace.shape[1]
kspace = fftc.ifftc(kspace, axis=1)
kspace = kspace.astype(np.complex64)
print("Exists")
sensemap = np.squeeze(cfl.read(file_sensemap_i))
sensemap = sensemap[0, :, :, :, :]
# it has coils z y x
# 6, 80, 156, 192
# print(sensemap.shape)
# we want coils x y z
sensemap = np.transpose(sensemap, [0, -1, 2, 1])
sensemap = np.expand_dims(sensemap, axis=0)
sensemap = sensemap.astype(np.complex64)
if verbose:
print(" Creating tfrecords (%d)..." % num_slices)
# for 2D plus time, only iterate over slices, not time frames
for i_slice in range(num_slices):
# normalization across time frames
kspace_x = kspace[:, i_slice, :, :, :]
file_out = os.path.join(
dir_out_i, "%s_x%03d.tfrecords" % (case_name, i_slice))
# kspace_x = kspace[:, i_x, :, :, i_f]/max_frames
sensemap_x = sensemap[:, :, i_slice, :, :]
example = tf.train.Example(features=tf.train.Features(
feature={
"name": _bytes_feature(str.encode(case_name)),
"slice": _int64_feature(i_slice),
"ks_shape_x": _int64_feature(kspace.shape[1]),
"ks_shape_y": _int64_feature(kspace.shape[2]),
"ks_shape_z": _int64_feature(kspace.shape[3]),
"ks_shape_t": _int64_feature(kspace.shape[4]),
"ks_shape_c": _int64_feature(kspace.shape[0]),
"map_shape_x": _int64_feature(sensemap.shape[2]),
"map_shape_y": _int64_feature(sensemap.shape[3]),
"map_shape_z": _int64_feature(sensemap.shape[4]),
"map_shape_c": _int64_feature(sensemap.shape[1]),
"map_shape_m": _int64_feature(sensemap.shape[0]),
"ks": _bytes_feature(kspace_x.tostring()),
"map": _bytes_feature(sensemap_x.tostring()),
}))
tf_writer = tf.python_io.TFRecordWriter(file_out)
tf_writer.write(example.SerializeToString())
tf_writer.close()
def setup_data_tfrecords_DCE(
dir_in_root,
dir_out,
data_divide=(0.75, 0.05, 0.2),
min_shape=[80, 180],
num_maps=1,
crop_maps=False,
verbose=False,
shuffle=True,
):
"""Setups training data as tfrecords.
prep_data.setup_data('/mnt/raid3/data/Studies_DCE/recon-ccomp6/',
'/mnt/raid3/jycheng/Project/deepspirit/data/train/', verbose=True)
"""
if verbose:
print("Directory names:")
print(" Input root: %s" % dir_in_root)
print(" Output root: %s" % dir_out)
# edits for /mnt/dense/data/MFAST_DCE and /home_local/ekcole/MFAST_DCE
dir_kspace = "sort-ccomp6"
dir_map = "recon-ccomp6"
# dir_kspace = "dce-ccomp6"
dir_cases = os.path.join(dir_in_root, dir_kspace)
file_kspace = "ks_sorted"
file_sensemap = "map"
case_list = os.listdir(dir_cases)
# shuffle cases (patients)
if shuffle is True:
random.shuffle(case_list)
else:
print("don't shuffle dataset")
num_cases = len(case_list)
i_train_1 = np.round(data_divide[0] * num_cases).astype(int)
i_validate_0 = i_train_1 + 1
i_validate_1 = np.round(
data_divide[1] * num_cases).astype(int) + i_validate_0
if not os.path.exists(dir_out):
os.mkdir(dir_out)
if not os.path.exists(os.path.join(dir_out, "train")):
os.mkdir(os.path.join(dir_out, "train"))
if not os.path.exists(os.path.join(dir_out, "validate")):
os.mkdir(os.path.join(dir_out, "validate"))
if not os.path.exists(os.path.join(dir_out, "test")):
os.mkdir(os.path.join(dir_out, "test"))
i_case = 0
for case_name in case_list:
file_kspace_i = os.path.join(dir_in_root, dir_kspace, case_name,
file_kspace)
file_sensemap_i = os.path.join(dir_in_root, dir_map, case_name,
file_sensemap)
file_sensemap_i_check = os.path.join(dir_in_root, dir_map, case_name,
file_sensemap + ".cfl")
# if no map, skip this case and do nothing
if not path.exists(file_sensemap_i_check):
print("Does not exist")
continue
# get dims from .hdr
h = open(file_kspace_i + ".hdr", "r")
h.readline() # skip
l = h.readline()
h.close()
dims = [int(i) for i in l.split()]
print(dims)
ky = dims[1]
kz = dims[2]
if ky != 180 or kz != 80:
print("wrong dimensions")
continue
if i_case < i_train_1:
dir_out_i = os.path.join(dir_out, "train")
elif i_case < i_validate_1:
dir_out_i = os.path.join(dir_out, "validate")
else:
dir_out_i = os.path.join(dir_out, "test")
if verbose:
print("Processing [%d] %s..." % (i_case, case_name))
i_case = i_case + 1
# if(i_case >= 50):
# break
kspace = np.squeeze(cfl.read(file_kspace_i))
# it wants coils x y z frames
kspace = np.transpose(kspace, [1, -1, -2, 2, 0])
if verbose:
print(" Slice shape: (%d, %d)" %
(kspace.shape[2], kspace.shape[3]))
print(" Num channels: %d" % kspace.shape[0])
print(" Num frames: %d" % kspace.shape[-1])
# number of frames
shape_f = kspace.shape[-1]
# number of slices in x direction
shape_x = kspace.shape[1]
kspace = fftc.ifftc(kspace, axis=1)
kspace = kspace.astype(np.complex64)
sensemap = np.squeeze(cfl.read(file_sensemap_i))
sensemap = sensemap[0, :, :, :, :]
# we want coils x y z
sensemap = np.transpose(sensemap, [0, -1, 2, 1])
sensemap = np.expand_dims(sensemap, axis=0)
sensemap = sensemap.astype(np.complex64)
if verbose:
print(" Creating tfrecords (%d)..." % shape_x)
# Need to iterate over both z and frames
for i_x in range(shape_x):
# normalization across time frames
kspace_x = kspace[:, i_x, :, :, :]
max_frames = np.max(np.abs(kspace_x))
# print(max_frames)
for i_f in range(shape_f):
file_out = os.path.join(
dir_out_i,
"%s_x%03d_f%03d.tfrecords" % (case_name, i_x, i_f))
kspace_x = kspace[:, i_x, :, :, i_f] / max_frames
sensemap_x = sensemap[:, :, i_x, :, :]
# #save images as pngs to check if time frames shuffling is done here
# # ks = np.squeeze(kspace_x)
# ks = kspace_x
# print(ks.shape)
# ks = np.transpose(ks, [1,2,0])
# # ks = np.expand_dims(ks, 0)
# ks = tf.convert_to_tensor(ks)
# print("ks")
# print(ks)
# sense = np.squeeze(sensemap_x)
# print(sense.shape)
# sense = np.transpose(sense, [1,2,0])
# sense = np.expand_dims(sense, -2)
# sense = tf.convert_to_tensor(sense)
# print("sensemap")
# print(sense)
# image_x = tf_util.model_transpose(ks, sense)
# sess = tf.Session()
# # Evaluate the tensor `c`.
# image_x = sess.run(image_x)
# filename = dir_out_i + '/images/case' + str(i_x) + '_f' + str(i_f) + '.png'
# print(filename)
# scipy.misc.imsave(filename, np.squeeze(np.abs(image_x)))
# at this stage, the images were not shuffled
example = tf.train.Example(features=tf.train.Features(
feature={
"name": _bytes_feature(str.encode(case_name)),
"xslice": _int64_feature(i_x),
"ks_shape_x": _int64_feature(kspace.shape[1]),
"ks_shape_y": _int64_feature(kspace.shape[2]),
"ks_shape_z": _int64_feature(kspace.shape[3]),
"ks_shape_c": _int64_feature(kspace.shape[0]),
"map_shape_x": _int64_feature(sensemap.shape[2]),
"map_shape_y": _int64_feature(sensemap.shape[3]),
"map_shape_z": _int64_feature(sensemap.shape[4]),
"map_shape_c": _int64_feature(sensemap.shape[1]),
"map_shape_m": _int64_feature(sensemap.shape[0]),
"ks": _bytes_feature(kspace_x.tostring()),
"map": _bytes_feature(sensemap_x.tostring()),
}))
tf_writer = tf.python_io.TFRecordWriter(file_out)
tf_writer.write(example.SerializeToString())
tf_writer.close()
