def training_step(self, batch, batch_idx):
subsampled_kspace, _, _, _, _, _, mask1 = batch
mask2 = -(mask1-1)
subsampled_kspace1 = subsampled_kspace * mask1 + 0.0
subsampled_kspace2 = subsampled_kspace * mask2 + 0.0
image1 = fastmri.ifft2c(subsampled_kspace1)
image1 = fastmri.complex_abs(image1)
image2 = fastmri.ifft2c(subsampled_kspace2)
image2 = fastmri.complex_abs(image2)
image = torch.vstack((image1, image2))
output_image = self(image)
output_image1 = output_image[0,:,:]
output_image2 = output_image[1,:,:]
output_kspace1 = torch.fft.fft2(output_image1)
output_kspace1 = torch.stack((output_kspace1.real, output_kspace1.imag), axis=-1)
output_kspace1 = output_kspace1 * mask2 + 0.0
output_kspace2 = torch.fft.fft2(output_image2)
output_kspace2 = torch.stack((output_kspace2.real, output_kspace2.imag), axis=-1)
output_kspace2 = output_kspace2 * mask1 + 0.0
loss = l1_l2_loss(output_kspace1, subsampled_kspace2) \
+ l1_l2_loss(output_kspace2, subsampled_kspace1)
self.log("loss", loss.detach())
return loss
def forward(self, kspace_pred: torch.Tensor, ref_kspace: torch.Tensor):
"""
Compute data consistency loss in kspace and
total variation loss in image space.
Inputs:
- kspace_pred: PyTorch tensor of shape (N, H, W, 2) holding predicted kspace.
- ref_kspace : input masked kspace
Returns:
- loss: PyTorch Variable holding a scalar giving the total variation loss
for img.
"""
output = fastmri.complex_abs(fastmri.ifft2c(kspace_pred))
gt = fastmri.complex_abs(fastmri.ifft2c(ref_kspace))
# energy_loss = torch.abs(torch.sum(output) - torch.sum(gt))
hdiff_pred = (output[:, :, :-1] - output[:, :, 1:]).view(-1)
hdiff_gt = (gt[:, :, :-1] - gt[:, :, 1:]).view(-1)
vdiff_pred = (output[:, :-1, :] - output[:, 1:, :]).view(-1)
vdiff_gt = (gt[:, :-1, :] - gt[:, 1:, :]).view(-1)
hdiff_var_loss = torch.sqrt(
torch.var(hdiff_pred)) - 1.25 * torch.sqrt(torch.var(hdiff_gt))
vdiff_var_loss = torch.sqrt(
torch.var(vdiff_pred)) - 1.25 * torch.sqrt(torch.var(vdiff_gt))
tv_loss = (torch.abs(hdiff_var_loss) + torch.abs(vdiff_var_loss))
return self.tv_weight * tv_loss
def hist_loss(current_kspace: torch.Tensor,
masked_kspace: torch.Tensor,
bins: int = 5):
"""
Inputs:
- kspace_pred: PyTorch tensor of shape (N, H, W, 2) holding predicted kspace.
- ref_kspace : input masked kspace
- mask: the subsampling mask
Returns:
- loss: PyTorch Variable holding a scalar giving the total variation loss
for img.
"""
output = fastmri.complex_abs(fastmri.ifft2c(current_kspace))
gt = fastmri.complex_abs(fastmri.ifft2c(masked_kspace))
hdiff_pred = (output[:, :, :-1] - output[:, :, 1:]).view(-1)
hdiff_gt = (gt[:, :, :-1] - gt[:, :, 1:]).view(-1)
hmin_pred, hmax_pred = hdiff_pred.min().item(), hdiff_pred.max().item()
hmin_gt, hmax_gt = hdiff_gt.min().item(), hdiff_gt.max().item()
hist_x = differentiable_histogram(hdiff_pred,
bins=bins,
min=hmin_pred,
max=hmax_pred)
hist_y = differentiable_histogram(hdiff_gt,
bins=bins,
min=hmin_gt,
max=hmax_gt)
hdiff_hist_loss = (hist_x - hist_y) / len(hdiff_pred)
hdiff_hist_loss = torch.norm(hdiff_hist_loss)
vdiff_pred = (output[:, :-1, :] - output[:, 1:, :]).view(-1)
vdiff_gt = (gt[:, :-1, :] - gt[:, 1:, :]).view(-1)
vmin_pred, vmax_pred = vdiff_pred.min().item(), vdiff_pred.max().item()
vmin_gt, vmax_gt = vdiff_gt.min().item(), vdiff_gt.max().item()
hist_x = differentiable_histogram(vdiff_pred,
bins=bins,
min=vmin_pred,
max=vmax_pred)
hist_y = differentiable_histogram(vdiff_gt,
bins=bins,
min=vmin_gt,
max=vmax_gt)
vdiff_hist_loss = (hist_x - hist_y) / len(vdiff_pred)
vdiff_hist_loss = torch.norm(vdiff_hist_loss)
output = output.view(-1)
gt = gt.view(-1)
gt_min, gt_max = gt.min().item(), gt.max().item()
hist_x = differentiable_histogram(output,
bins=bins,
min=gt_min,
max=gt_max)
hist_y = differentiable_histogram(gt, bins=bins, min=gt_min, max=gt_max)
gt_hist_loss = (hist_x - hist_y) / len(output)
gt_hist_loss = torch.norm(gt_hist_loss)
return (hdiff_hist_loss + vdiff_hist_loss + gt_hist_loss)
def load_data(file_dir_path):
file_path = get_files(file_dir_path)
file_num = len(file_path)
total_image_list = []
total_sampled_image_list = []
for h5_num in range(file_num):
total_kspace, slices_num = load_dataset(file_path[0])
image_list = []
slice_kspace_tensor_list = []
for i in range(slices_num):
slice_kspace = total_kspace[i]
slice_kspace_tensor = T.to_tensor(
slice_kspace) # convert numpy to tensor
slice_image = fastmri.ifft2c(
slice_kspace_tensor) # inverse fast FT
slice_image_abs = fastmri.complex_abs(
slice_image) # compute the absolute value to get a real image
image_list.append(slice_image_abs)
slice_kspace_tensor_list.append(
slice_kspace_tensor) # 35* torch[640, 368])
image_list_tensor = torch.stack(image_list,
dim=0) # torch.Size([35, 640, 368])
total_image_list.append(image_list_tensor)
mask_func = RandomMaskFunc(
center_fractions=[0.08],
accelerations=[4]) # create the mask function object
sampled_image_list = []
for i in range(slices_num):
slice_kspace_tensor = slice_kspace_tensor_list[i]
masked_kspace, mask = T.apply_mask(slice_kspace_tensor, mask_func)
Ny, Nx, _ = slice_kspace_tensor.shape
mask = mask.repeat(Ny, 1, 1).squeeze()
# functions.show_slice(mask, cmap='gray')
# functions.show_slice(image_list[10], cmap='gray')
sampled_image = fastmri.ifft2c(
masked_kspace) # inverse fast FT to get the complex image
sampled_image_abs = fastmri.complex_abs(sampled_image)
sampled_image_list.append(sampled_image_abs)
sampled_image_list_tensor = torch.stack(
sampled_image_list, dim=0) # torch.Size([35, 640, 368])
total_sampled_image_list.append(sampled_image_list_tensor)
# total_image_tensor = torch.cat(total_image_list, dim=0) # torch.Size([6965, 640, 368])
# total_sampled_image_tensor = torch.cat(total_sampled_image_list, dim=0) # torch.Size([6965, 640, 368])
total_image_tensor = torch.stack(total_image_list,
dim=0) # torch.Size([199, 35, 640, 368])
total_sampled_image_tensor = torch.stack(
total_sampled_image_list, dim=0) # torch.Size([199, 35, 640, 368])
print(total_image_tensor.shape)
print(total_sampled_image_tensor.shape)
return total_image_tensor, total_sampled_image_tensor
def to_cropped_image(masked_kspace, target, attrs):
# inverse Fourier transform to get zero filled solution
image = fastmri.ifft2c(masked_kspace)
# crop input to correct size
if target is not None:
crop_size = (target.shape[-2], target.shape[-1])
else:
crop_size = (attrs["recon_size"][0], attrs["recon_size"][1])
# check for FLAIR 203
if image.shape[-2] < crop_size[1]:
crop_size = (image.shape[-2], image.shape[-2])
image = T.complex_center_crop(image, crop_size)
# absolute value
image = fastmri.complex_abs(image)
# normalize input
image, mean, std = T.normalize_instance(image, eps=1e-11)
image = image.clamp(-6, 6)
# normalize target
if target is not None:
if isinstance(target, np.ndarray):
target = T.to_tensor(target)
target = T.center_crop(target, crop_size)
target = T.normalize(target, mean, std, eps=1e-11)
target = target.clamp(-6, 6)
else:
target = torch.Tensor([0])
return image, target
def data_consistency_kspace(self, prediction, k_space_slice, mask):
"""
Args:
- prediction: net (or block) predicted real image in complex domain
- k_space_slice: initially sampled elements in k-space
- mask: corresponding nonzero location in kspace
Res:
image in k space where:
- masked entries of initial slice are replaced with entries predicted by output
- non masked entries of initial slice stay the same
"""
prediction = prediction[:, :, 0:320, 0:320]
prediction = prediction.permute(
0, 2, 3, 1) # prediction from 1 x 2 x h x w to 1 x h x w x 2
prediction = self.proper_padding(
prediction, k_space_slice) # pad prediction to be 640 x 372 x 2
k_space_prediction = fastmri.fft2c(
prediction) # transform prediction to kspace domain
k_space_out = (
1 - mask) * k_space_prediction + mask * k_space_slice # apply mask
prediction = fastmri.ifft2c(k_space_out) # back to cplx image
prediction = transforms.complex_center_crop(
prediction, (320, 320)) # crop image to 320 x 320
prediction = prediction.permute(0, 3, 1, 2) # back to 1 x 2 x h x w
return prediction
def save_zero_filled(data_dir, out_dir, which_challenge):
reconstructions = {}
for f in data_dir.iterdir():
with h5py.File(f, "r") as hf:
enc = ismrmrd.xsd.CreateFromDocument(hf["ismrmrd_header"][()]).encoding[0]
masked_kspace = transforms.to_tensor(hf["kspace"][()])
# extract target image width, height from ismrmrd header
crop_size = (enc.reconSpace.matrixSize.x, enc.reconSpace.matrixSize.y)
# inverse Fourier Transform to get zero filled solution
image = fastmri.ifft2c(masked_kspace)
# check for FLAIR 203
if image.shape[-2] < crop_size[1]:
crop_size = (image.shape[-2], image.shape[-2])
# crop input image
image = transforms.complex_center_crop(image, crop_size)
# absolute value
image = fastmri.complex_abs(image)
# apply Root-Sum-of-Squares if multicoil data
if which_challenge == "multicoil":
image = fastmri.rss(image, dim=1)
reconstructions[f.name] = image
fastmri.save_reconstructions(reconstructions, out_dir)
def save_zero_filled(data_dir, out_dir, which_challenge):
reconstructions = {}
for fname in tqdm(list(data_dir.glob("*.h5"))):
with h5py.File(fname, "r") as hf:
et_root = etree.fromstring(hf["ismrmrd_header"][()])
masked_kspace = transforms.to_tensor(hf["kspace"][()])
# extract target image width, height from ismrmrd header
enc = ["encoding", "encodedSpace", "matrixSize"]
crop_size = (
int(et_query(et_root, enc + ["x"])),
int(et_query(et_root, enc + ["y"])),
)
# inverse Fourier Transform to get zero filled solution
image = fastmri.ifft2c(masked_kspace)
# check for FLAIR 203
if image.shape[-2] < crop_size[1]:
crop_size = (image.shape[-2], image.shape[-2])
# crop input image
image = transforms.complex_center_crop(image, crop_size)
# absolute value
image = fastmri.complex_abs(image)
# apply Root-Sum-of-Squares if multicoil data
if which_challenge == "multicoil":
image = fastmri.rss(image, dim=1)
reconstructions[fname.name] = image
fastmri.save_reconstructions(reconstructions, out_dir)
def __call__(self, target, fname, slice_num=0, attrs=None, seed=None):
# Preprocess the data here
# target shape: [H, W, 1] or [H, W, 3]
img = target
if target.shape[2] != 2:
img = np.concatenate((target, np.zeros_like(target)), axis=2)
assert img.shape[-1] == 2
img = to_tensor(img)
kspace = fastmri.fft2c(img)
center_kspace, _ = apply_mask(kspace,
self.mask_func,
hamming=True,
seed=seed)
img_LF = fastmri.complex_abs(fastmri.ifft2c(center_kspace))
img_LF = img_LF.unsqueeze(0)
image, mean, std = normalize_instance(img_LF, eps=1e-11)
image = image.clamp(-6, 6)
# img_LF tensor should have shape [H, W, ?]
target = to_tensor(np.transpose(target,
(2, 0, 1))) # target shape [1, H, W]
target = normalize(target, mean, std, eps=1e-11)
target = target.clamp(-6, 6)
target = target.squeeze(0)
# check for max value
max_value = 0.0
# print('traget shape', target.shape)
# print('image shape', image.shape)
return image, target, mean, std, fname, slice_num, max_value
def load_data_from_pathlist(path):
file_num = len(path)
use_num = file_num // 3
total_target_list = []
total_sampled_image_list = []
for h5_num in range(use_num):
total_kspace, slices_num, target = load_dataset(path[h5_num])
image_list = []
slice_kspace_tensor_list = []
target_image_list = []
for i in range(slices_num):
slice_kspace = total_kspace[i]
#target_image = target[i]
slice_kspace_tensor = T.to_tensor(
slice_kspace) # convert numpy to tensor
slice_kspace_tensor = slice_kspace_tensor.float()
#print(slice_kspace_tensor.shape)
slice_kspace_tensor_list.append(
slice_kspace_tensor) # 35* torch[640, 368])
#target = target_image_list.append(target_image)
#image_list_tensor = torch.stack(image_list, dim=0) # torch.Size([35, 640, 368])
#total_image_list.append(image_list_tensor)
mask_func = RandomMaskFunc(
center_fractions=[0.08],
accelerations=[4]) # create the mask function object
sampled_image_list = []
target_list = []
for i in range(slices_num):
slice_kspace_tensor = slice_kspace_tensor_list[i]
masked_kspace, mask = T.apply_mask(slice_kspace_tensor, mask_func)
Ny, Nx, _ = slice_kspace_tensor.shape
mask = mask.repeat(Ny, 1, 1).squeeze()
# functions.show_slice(mask, cmap='gray')
# functions.show_slice(image_list[10], cmap='gray')
sampled_image = fastmri.ifft2c(
masked_kspace) # inverse fast FT to get the complex image
sampled_image = T.complex_center_crop(sampled_image, (320, 320))
sampled_image_abs = fastmri.complex_abs(sampled_image)
sampled_image_list.append(sampled_image_abs)
sampled_image_list_tensor = torch.stack(
sampled_image_list, dim=0) # torch.Size([35, 640, 368])
total_sampled_image_list.append(sampled_image_list_tensor)
target = T.to_tensor(target)
total_target_list.append(target)
#target_image_tensor = torch.cat(target_image_list, dim=0) # torch.Size([6965, 640, 368])
total_target = torch.cat(total_target_list, dim=0)
total_sampled_image_tensor = torch.cat(
total_sampled_image_list, dim=0) # torch.Size([6965, 640, 368])
total_sampled_image_tensor, mean, std = T.normalize_instance(
total_sampled_image_tensor, eps=1e-11)
total_sampled_image_tensor = total_sampled_image_tensor.clamp(-6, 6)
target_image_tensor = T.normalize(total_target, mean, std, eps=1e-11)
target_image_tensor = target_image_tensor.clamp(-6, 6)
# total_image_tensor = torch.stack(total_image_list, dim=0) # torch.Size([199, 35, 640, 368])
# total_sampled_image_tensor = torch.stack(total_sampled_image_list, dim=0) # torch.Size([199, 35, 640, 368])
#print(target_image_tensor.shape)
#print(total_sampled_image_tensor.shape)
return target_image_tensor, total_sampled_image_tensor
def forward(self, masked_kspace, mask):
sens_maps = self.sens_net(masked_kspace, mask)
kspace_pred = masked_kspace.clone()
for cascade in self.cascades:
kspace_pred = cascade(kspace_pred, masked_kspace, mask, sens_maps)
return fastmri.rss(fastmri.complex_abs(fastmri.ifft2c(kspace_pred)),
dim=1)
def add_ge_oversampling(masked_kspace):
# call all of these "masked_kspace" to preserve memory
readout_len = masked_kspace.shape[-3]
pad_size = (0, 0, 0, 0, readout_len // 2, readout_len // 2)
masked_kspace = fastmri.ifft2c(masked_kspace)
masked_kspace = F.pad(masked_kspace, pad_size)
masked_kspace = fastmri.fft2c(masked_kspace)
return masked_kspace
def test_ifft2(shape):
shape = shape + [2]
x = create_input(shape)
out_torch = fastmri.ifft2c(x).numpy()
out_torch = out_torch[..., 0] + 1j * out_torch[..., 1]
input_numpy = transforms.tensor_to_complex_np(x)
input_numpy = np.fft.ifftshift(input_numpy, (-2, -1))
out_numpy = np.fft.ifft2(input_numpy, norm="ortho")
out_numpy = np.fft.fftshift(out_numpy, (-2, -1))
assert np.allclose(out_torch, out_numpy)
def forward(
self,
current_kspace: torch.Tensor,
ref_kspace: torch.Tensor,
mask: torch.Tensor,
) -> torch.Tensor:
zero = torch.zeros(1, 1, 1, 1).to(current_kspace)
soft_dc = torch.where(mask, current_kspace - ref_kspace,
zero) * self.dc_weight
model_term = fastmri.fft2c(self.model(fastmri.ifft2c(current_kspace)))
return current_kspace - soft_dc - model_term
def forward(
self,
masked_kspace: torch.Tensor,
mask: torch.Tensor,
num_low_frequencies: Optional[int] = None,
) -> torch.Tensor:
sens_maps = self.sens_net(masked_kspace, mask, num_low_frequencies)
kspace_pred = masked_kspace.clone()
for cascade in self.cascades:
kspace_pred = cascade(kspace_pred, masked_kspace, mask, sens_maps)
return fastmri.rss(fastmri.complex_abs(fastmri.ifft2c(kspace_pred)),
dim=1)
def test_step(self, batch, batch_idx):
masked_kspace, _, fname, slice_num, _, attrs, _ = batch = batch
# image, _ = to_cropped_image(masked_kspace, None, attrs)
image = fastmri.ifft2c(masked_kspace)
image = fastmri.complex_abs(image)
output = self.forward(image)
return {
"fname": fname,
"slice": slice_num,
"output": output.cpu().numpy(),
}
def forward(self, masked_kspace: torch.Tensor,
mask: torch.Tensor) -> torch.Tensor:
if self.mask_center:
pad, num_low_freqs = self.get_pad_and_num_low_freqs(
mask, self.num_sense_lines)
masked_kspace = transforms.batched_mask_center(
masked_kspace, pad, pad + num_low_freqs)
# convert to image space
images, batches = self.chans_to_batch_dim(
fastmri.ifft2c(masked_kspace))
# estimate sensitivities
return self.divide_root_sum_of_squares(
self.batch_chans_to_chan_dim(self.norm_unet(images), batches))
def validation_step(self, batch, batch_idx):
masked_kspace, mask, target, fname, slice_num, max_value, _ = batch
kspace_pred = self(masked_kspace, mask)
output = fastmri.complex_abs(fastmri.ifft2c(kspace_pred))
target, output = transforms.center_crop_to_smallest(target, output)
return {
"batch_idx": batch_idx,
"fname": fname,
"slice_num": slice_num,
"max_value": max_value,
"output": output,
"target": target,
"val_loss": self.loss(kspace_pred, masked_kspace),
}
def forward(
self,
current_kspace: torch.Tensor,
ref_kspace: torch.Tensor,
mask: torch.Tensor,
) -> torch.Tensor:
zero = torch.zeros(1, 1, 1, 1).to(current_kspace)
soft_dc = torch.where(mask, current_kspace - ref_kspace,
zero) * self.dc_weight
with torch.enable_grad():
X = current_kspace.clone().detach().requires_grad_(True)
histl = hist_loss(X, ref_kspace.clone().detach())
histl.backward()
soft_histc = X.grad * self.hist_weight
model_term = fastmri.fft2c(self.model(fastmri.ifft2c(current_kspace)))
return current_kspace - soft_dc - soft_histc - model_term
def test_step(self, batch, batch_idx):
masked_kspace, mask, _, fname, slice_num, _, crop_size = batch
crop_size = crop_size[0] # always have a batch size of 1 for varnet
kspace_pred = self(masked_kspace, mask)
output = fastmri.complex_abs(fastmri.ifft2c(kspace_pred))
# check for FLAIR 203
if output.shape[-1] < crop_size[1]:
crop_size = (output.shape[-1], output.shape[-1])
output = transforms.center_crop(output, crop_size)
return {
"fname": fname,
"slice": slice_num,
"output": output.cpu().numpy(),
}
def forward(self, masked_kspace: torch.Tensor,
mask: torch.Tensor) -> torch.Tensor:
# get low frequency line locations and mask them out
cent = mask.shape[-2] // 2
left = torch.nonzero(mask.squeeze()[:cent] == 0)[-1]
right = torch.nonzero(mask.squeeze()[cent:] == 0)[0] + cent
num_low_freqs = right - left
pad = (mask.shape[-2] - num_low_freqs + 1) // 2
x = transforms.mask_center(masked_kspace, pad, pad + num_low_freqs)
# convert to image space
x = fastmri.ifft2c(x)
x, b = self.chans_to_batch_dim(x)
# estimate sensitivities
x = self.norm_unet(x)
x = self.batch_chans_to_chan_dim(x, b)
x = self.divide_root_sum_of_squares(x)
return x
def _base_fastmri_unet_transform(
kspace,
mask,
ground_truth,
attrs,
which_challenge="singlecoil",
):
kspace = fastmri_transforms.to_tensor(kspace)
mask = mask[..., :kspace.shape[-2]] # accounting for variable size masks
masked_kspace = kspace * mask.unsqueeze(-1) + 0.0
# inverse Fourier transform to get zero filled solution
image = fastmri.ifft2c(masked_kspace)
# crop input to correct size
if ground_truth is not None:
crop_size = (ground_truth.shape[-2], ground_truth.shape[-1])
else:
crop_size = (attrs["recon_size"][0], attrs["recon_size"][1])
# check for FLAIR 203
if image.shape[-2] < crop_size[1]:
crop_size = (image.shape[-2], image.shape[-2])
# noinspection PyTypeChecker
image = fastmri_transforms.complex_center_crop(image, crop_size)
# absolute value
image = fastmri.complex_abs(image)
# apply Root-Sum-of-Squares if multicoil data
if which_challenge == "multicoil":
image = fastmri.rss(image)
# normalize input
image, mean, std = fastmri_transforms.normalize_instance(image, eps=1e-11)
image = image.clamp(-6, 6)
return image.unsqueeze(0), mean, std
def training_step(self, batch, batch_idx):
subsampled_kspace, _, _, _, _, _, mask_loss = batch
mask_train = -(mask_loss-1)
subsampled_kspace_train = subsampled_kspace * mask_train + 0.0
subsampled_kspace_loss = subsampled_kspace * mask_loss + 0.0
image_train = fastmri.ifft2c(subsampled_kspace_train)
image_train = fastmri.complex_abs(image_train)
output_image = self(image_train)
output_kspace = torch.fft.fft2(output_image)
output_kspace = torch.stack((output_kspace.real, output_kspace.imag), axis=-1)
output_kspace_loss = output_kspace * mask_loss + 0.0
loss = l1_l2_loss(output_kspace_loss, subsampled_kspace_loss)
self.log("loss", loss.detach())
return loss
def data_transform(kspace, mask_function, target, data_attributes, filename,
slice_num):
"""
Perform preprocessing of the kspace image, in order to get a proper input for the net. Should be invoked from
the SliceData class.
Args:
- kspace: complete sampled kspace image
- mask_func: masking function to apply mask to kspace (TODO not working: we are passing from outside)
- target: the target image to be reconstructed from the kspace
- data_attributes: attributes of the whole HDF5 file
Returns:
- normalized_masked_image: original kspace with mask applied and cropped to 320 x 320
- mask: mask generated by masking function
- normalized_target: normalized target
- max_value: highest entry in target tensor (for SSIM loss)
"""
kspace_t = transforms.to_tensor(kspace)
kspace_t = transforms.normalize_instance(kspace_t)[0]
masked_kspace, mask = transforms.apply_mask(
data=kspace_t, mask_func=mask_func
) # apply mask: returns masked space and generated mask
masked_image = fastmri.ifft2c(
masked_kspace
) # Apply Inverse Fourier Transform to get the complex image
masked_image = transforms.complex_center_crop(
masked_image, (320, 320)) # center crop masked image
masked_image = masked_image.permute(
2, 0, 1) # permuting the masked image fot pytorch n x c x h x w format
masked_image = transforms.normalize_instance(masked_image)[0] # normalize
target = transforms.to_tensor(target)
target = transforms.normalize_instance(target)[0] # normalize
target = torch.unsqueeze(target, 0) # add dimension
return kspace_t, masked_image, target, mask, data_attributes[
'max'], slice_num
def forward(self, masked_kspace, mask):
def get_low_frequency_lines(mask):
l = r = mask.shape[-2] // 2
while mask[..., r, :]:
r += 1
while mask[..., l, :]:
l -= 1
return l + 1, r
l, r = get_low_frequency_lines(mask)
num_low_freqs = r - l
pad = (mask.shape[-2] - num_low_freqs + 1) // 2
x = T.mask_center(masked_kspace, pad, pad + num_low_freqs)
x = fastmri.ifft2c(x)
x, b = self.chans_to_batch_dim(x)
x = self.norm_unet(x)
x = self.batch_chans_to_chan_dim(x, b)
x = self.divide_root_sum_of_squares(x)
return x
def visualize_kspace(kspace, dim=None, crop=False, output_filepath=None):
kspace = fastmri.ifft2c(kspace)
if crop:
crop_size = (kspace.shape[-2], kspace.shape[-2])
kspace = T.complex_center_crop(kspace, crop_size)
kspace = fastmri.complex_abs(kspace)
kspace, _, _ = T.normalize_instance(kspace, eps=1e-11)
kspace = kspace.clamp(-6, 6)
else:
# Compute absolute value to get a real image
kspace = fastmri.complex_abs(kspace)
if dim is not None:
kspace = fastmri.rss(kspace, dim=dim)
img = np.abs(kspace.numpy())
if output_filepath is not None:
if not output_filepath.parent.exists():
output_filepath.parent.mkdir(parents=True)
plt.imshow(img, cmap='gray')
plt.axis("off")
plt.savefig(output_filepath, bbox_inches="tight", pad_inches=0)
else:
plt.imshow(img, cmap='gray')
plt.show()
def forward(self, mr_img, mk_space, mask):
# Loop over num = cascades number
for i in range(self.n_cascade):
# The laplacian decomposition will produce different images composing the gaussian/laplacian pyramids
# The gaussians are the downsampled images from the original one
# The laplacian represents the 'errors' within the images related to the gaussian images
gaussian_3, lap_1, lap_2 = self.lapl_dec(mr_img.cpu())
# Resize along with the input/output channels
mr_img = self.shuffle_down_4(mr_img)
# The convBlock1 is here performed to extract shallow features
mr_img = self.convBlock1(mr_img)
# ---- 3 branches ----
branch_1 = self.shuffle_up_4(mr_img)
branch_1 = self.branch1(branch_1)
branch_2 = self.shuffle_up_2(mr_img)
branch_2 = self.branch2(branch_2)
branch_3 = self.branch3(mr_img)
branch_out1 = torch.stack((self.pixel_conv(branch_1[...,0], lap_1[...,0]),
self.pixel_conv(branch_1[...,1], lap_1[...,1])), dim=-1)
branch_out2 = torch.stack((self.pixel_conv(branch_2[...,0], lap_2[...,0]),
self.pixel_conv(branch_2[...,1], lap_2[...,1])), dim=-1)
branch_out3 = torch.stack((self.pixel_conv(branch_3[...,0], gaussian_3[...,0]),
self.pixel_conv(branch_3[...,1], gaussian_3[...,1])), dim=-1)
# Performing the 2x linear upsample in order to add the different branches correctly
output = self.lapl_rec(branch_out3, branch_out2)
output = self.lapl_rec(output, branch_out1)
# The DataConsistency Layer step is done here
mr_img = fastmri.ifft2c((1.0 - mask) * fastmri.fft2c(output) + mk_space)
return mr_img
def __call__(
self,
kspace: np.ndarray,
mask: np.ndarray,
target: np.ndarray,
attrs: Dict,
fname: str,
slice_num: int,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, str,
int, float]:
"""
Args:
kspace: Input k-space of shape (num_coils, rows, cols) for
multi-coil data or (rows, cols) for single coil data.
mask: Mask from the test dataset.
target: Target image.
attrs: Acquisition related information stored in the HDF5 object.
fname: File name.
slice_num: Serial number of the slice.
Returns:
tuple containing:
image: Zero-filled input image.
target: Target image converted to a torch.Tensor.
mean: Mean value used for normalization.
std: Standard deviation value used for normalization.
fname: File name.
slice_num: Serial number of the slice.
"""
kspace = to_tensor(kspace)
# check for max value
max_value = attrs["max"] if "max" in attrs.keys() else 0.0
# apply mask
if self.mask_func:
seed = None if not self.use_seed else tuple(map(ord, fname))
masked_kspace, mask = apply_mask(kspace, self.mask_func, seed)
else:
masked_kspace = kspace
# inverse Fourier transform to get zero filled solution
image = fastmri.ifft2c(masked_kspace)
# crop input to correct size
if target is not None:
crop_size = (target.shape[-2], target.shape[-1])
else:
crop_size = (attrs["recon_size"][0], attrs["recon_size"][1])
# check for FLAIR 203
if image.shape[-2] < crop_size[1]:
crop_size = (image.shape[-2], image.shape[-2])
image = complex_center_crop(image, crop_size)
# absolute value
image = fastmri.complex_abs(image)
# apply Root-Sum-of-Squares if multicoil data
if self.which_challenge == "multicoil":
image = fastmri.rss(image)
# normalize input
image, mean, std = normalize_instance(image, eps=1e-11)
image = image.clamp(-6, 6)
# normalize target
if target is not None:
target = to_tensor(target)
target = center_crop(target, crop_size)
target = normalize(target, mean, std, eps=1e-11)
target = target.clamp(-6, 6)
else:
target = torch.Tensor([0])
return image, target, mean, std, fname, slice_num, max_value
def sens_reduce(self, x: torch.Tensor,
sens_maps: torch.Tensor) -> torch.Tensor:
return fastmri.complex_mul(fastmri.ifft2c(x),
fastmri.complex_conj(sens_maps)).sum(
dim=1, keepdim=True)
def sens_reduce(x):
x = fastmri.ifft2c(x)
return fastmri.complex_mul(x, fastmri.complex_conj(sens_maps)).sum(
dim=1, keepdim=True)
