MapMetricWrapper("L1_map",
lambda x, y: torch.abs(x - y),
average_method="mean",
mask_keys=['mask2']),
"L2":
MetricWrapper("L2", MSELoss()),
#"SSIM": SSIM3D(average_method="mean"),
"SSIM_mask":
SSIM3D(average_method="mean", mask_keys=["mask", "mask2"]),
"SSIM_Wrapped":
MetricWrapper("SSIM_wrapped",
lambda x, y: functional_ssim(x, y, return_map=False),
use_mask=True,
mask_key="mask"),
"ssim_base":
MetricWrapper('SSIM_base', ssim3D)
}
motion_trsfm = RandomMotionFromTimeCourse(verbose=True,
compare_to_original=True,
metrics=metrics,
oversampling_pct=0.0)
dataset.set_transform(motion_trsfm)
tf = dataset[0]
computed_metrics = tf["T1"]["metrics"]
print("Computed metrics: {}".format(computed_metrics))
ov(tf['T1']['data'].squeeze().numpy())
dataset_not = ImagesDataset(paths_dict, transform=None)
dataload = torch.utils.data.DataLoader(dataset, num_workers=0, batch_size=1)
dataloadnot = torch.utils.data.DataLoader(dataset_not, num_workers=0, batch_size=1)
ddd = dataset[0] #next(iter(dataset))
ii = np.squeeze( ddd['image']['data'][0], axis=1)
ddno = dataset_not[0] #
dd= next(iter(dataload))
ddno = next(iter(dataloadnot))
ii = np.squeeze( dd['image']['data'][0,0,:],axis=1)
iio = np.squeeze( ddno['image']['data'][0,0,:],axis=1)
ov(ii)
ov(iio)
#save
sample = dataset[0]
output = dict(image=Path('/tmp/test_im.nii.gz'),label1=Path('/tmp/test_p1.nii.gz'),label2=Path('/tmp/test_p2.nii.gz'))
dataset.save_sample(sample, output)
#explore
out_dir = '/data/romain/data_exemple/augment/'
suj = [[
Image('T1','/data/romain/data_exemple/nifti_proc/PRISMA_MBB_DB/2017_03_07_DEV_236_MBB_DB_Pilote02/anat_S02_t1mpr_SAG_NSel_S176/cat12/s_S02_t1mpr_SAG_NSel_S176.nii.gz',INTENSITY),
Image('mask','/data/romain/data_exemple/nifti_proc/PRISMA_MBB_DB/2017_03_07_DEV_236_MBB_DB_Pilote02/anat_S02_t1mpr_SAG_NSel_S176/cat12/mask_brain_erode_dilate.nii.gz',LABEL)
]]
transforms = Compose((RandomBiasField(coefficients_range=(-0.5, 0.5),order=3, verbose=True), ))
def display_original_data(self, idx):
volume = self.get_volume_nibabel(idx)
ov(volume.get_data())
header=None)
fpars = np.asarray(mvt)
transforms = get_motion_transform()
suj = [[
Image(
'T1',
'/home/romain/QCcnn/motion_cati_brain_ms/brain_s_S02_Sag_MPRAGE.nii.gz',
'intensity'),
]]
dataset = ImagesDataset(suj, transform=transforms)
s = dataset[0]
ov(s['T1']['data'][0])
tt = dataset.get_transform()
plt.figure()
plt.plot(tt.fitpars.T)
dataset.save_sample(
s, dict(T1='/home/romain/QCcnn/motion_cati_brain_ms/toto10.nii'))
#look at distribution of metric on simulate motion
dir_cache = '/network/lustre/dtlake01/opendata/data/ds000030/rrr/CNN_cache'
dd = gdir(dir_cache, 'mask_mv')
fr = gfile(dd, 'resul')
name_res = get_parent_path(dd)[1]
res = [pd.read_csv(ff) for ff in fr]
sell_col = res[0].keys()
sell_col = [
tmot.nufft_type = '1D_type2'
tmot.preserve_center_frequency_pct = 0
tmot.nT = 218
tmot.maxGlobalDisp, tmot.maxGlobalRot = (4, 4), (3, 3)
tmot._simulate_random_trajectory()
plt.figure()
plt.plot(tmot.fitpars.T)
fitpar = tmot.fitpars
tmot.fitpars = fitpar
tmot.simulate_displacement = False # True; # tmot.oversampling_pct = 1
smot = tmot(sdata)
fitpar = tmot.fitpars
ov(smot.t1.data[0])
#fitpar = interpolate_fitpars(tmot.fitpars, len_output=image.shape[1])
import finufft
def nufft_type1(freq_domain, fitpar, trans_last=False):
#trans_last=False #weird idea while looking to match tio.Affine ... no more usefull
eps = 1E-7
f = np.zeros(image.shape, dtype=np.complex128, order='F')
lin_spaces = [
np.linspace(-0.5, 0.5 - 1 / x, x) * 2 * math.pi
for x in freq_domain.shape
] # todo it suposes 1 vox = 1mm
#remove 1/x to avoid small scaling
header=None)
fpars = np.asarray(mvt)
transforms = get_motion_transform()
suj = [[
Image(
'T1',
'/home/romain/QCcnn/motion_cati_brain_ms/brain_s_S02_Sag_MPRAGE.nii.gz',
'intensity'),
]]
dataset = ImagesDataset(suj, transform=transforms)
s = dataset[0]
ov(s['T1']['data'][0])
tt = dataset.get_transform()
plt.figure()
plt.plot(tt.fitpars.T)
dataset.save_sample(
s, dict(T1='/home/romain/QCcnn/motion_cati_brain_ms/toto10.nii'))
#look at distribution of metric on simulate motion
dir_cache = '/network/lustre/dtlake01/opendata/data/ds000030/rrr/CNN_cache'
dd = gdir(dir_cache, 'mask_mv')
fr = gfile(dd, 'resul')
name_res = get_parent_path(dd)[1]
res = [pd.read_csv(ff) for ff in fr]
sell_col = res[0].keys()
sell_col = [
trans_torch = np.array(trans)/(image_size/2) mr = create_rotation_matrix_3d(rot) ms = np.diag(scale) center = np.ceil(image_size/2) center = center.T - [email protected] center_mat=np.zeros([4,4]) center_mat[0:3,3] = center[0:3].T maff = np.hstack((ms @ mr,np.expand_dims(trans,0).T)) maff_torch = np.hstack((ms @ mr,trans_torch.T)) maff = np.vstack((maff,[0,0,0,1])) nib_fin = nib.Nifti1Image(ii.numpy()[0], affine) new_aff = affine @ np.linalg.inv(maff+center_mat) #new_aff = maff @ affine # other way round new_aff = affine@maff nib_fin.affine[:] = new_aff[:] fout = nbp.resample_from_to(nib_fin, (nib_fin.shape, affine), cval=-1) #fout = nbp.resample_from_to(nib_fin, (nib_fin.shape, new_aff), cval=-1) ov(fout.get_fdata()) #it gives almost the same, just the scalling is shifted with nibabel (whereas it is centred with torch mafft = maff_torch[np.newaxis,:] mafft = torch.from_numpy(mafft) x = ii.permute(0,3,2,1).unsqueeze(0) grid = F.affine_grid(mafft, x.shape, align_corners=False).float() x = F.grid_sample(x, grid, align_corners=False) xx = x[0,0].numpy().transpose(2,1,0) ov(xx) # make the inverse transform xx=torch.zeros(4,4); xx[3,3]=1 xx[0:3,0:4] = mafft[0] imaf = xx.inverse()
