def recover_medimg(out_path):
step = 'medimg'
prev_step = 'resample'
save_nii_dir, save_mat_dir = make_save_dirs(step, out_path)
data = load_data_label(prev_step)
original_ss = Cparams['original_ss']
# de-resample: resampled unified spacing to original spacings (or according to shape ratio)
for case_name in data.keys():
label = data[case_name]['label']
label_shape = label.shape
if print_verbose:
print(case_name)
# recover original label id (3 for trachea)
if Params['data_name'] == 'Segthor':
label[label == 1] = 3 # trachea label id
# recover XYZ order
original_spacing = original_ss[case_name]['spacing'].copy()
if Params['depth_transpose']:
label = np.transpose(label, (1, 2, 0))
original_spacing = [
original_spacing[1], original_spacing[2], original_spacing[0]
]
if print_verbose:
print('after recover medimg', label.shape, original_spacing)
# equip label file with original spacings (spacing setting)
original_spacing.append(1)
image_nib = nib.Nifti1Image(label, np.diag(original_spacing))
nib.save(image_nib,
os.path.join(save_nii_dir, '{}.nii.gz'.format(case_name)))
def align_volume_center(out_path):
step = 'align'
prev_step = 'resample'
save_nii_dir, save_mat_dir = make_save_dirs(step, out_path)
data = load_data(prev_step)
max_shape = Cparams['max_shape']
max_center = np.array(max_shape) // 2
for case_name in data.keys():
image = data[case_name]['image']
image_shape = np.array(image.shape)
image_center = image_shape // 2
start = (max_center - image_center).astype(int)
end = (start + image_shape).astype(int)
print(case_name)
print('before', image.shape, type(image[0, 0, 0]), image.max(),
image.sum())
image_new = np.zeros(max_shape).astype(np.int16)
image_new[start[0]:end[0], start[1]:end[1],
start[2]:end[2]] = copy.deepcopy(image)
image = image_new
data[case_name]['image'] = image
print('after ', image.shape, type(image[0, 0, 0]), image.max(),
image.sum())
save_case(step, case_name, image, data, save_nii_dir, save_mat_dir)
def reverse_resample(out_path):
step = 'resample'
prev_step = 'align'
save_nii_dir, save_mat_dir = make_save_dirs(step, out_path)
data = load_data_label(prev_step)
original_ss = Cparams['original_ss']
# de-resample: resampled unified spacing to original spacings (or according to shape ratio)
for case_name in data.keys():
label = data[case_name]['label']
label_shape = label.shape
original_shape = original_ss[case_name]['shape']
zoom_factor = [original_shape[i] / label_shape[i] for i in range(3)]
if print_verbose:
print(case_name)
print('before ', label.shape, type(label[0, 0, 0]), label.max(),
label.min())
orders = [3, 1, 3, 1]
if zoom_factor != [1, 1, 1]:
label = ndimage.zoom(label,
zoom=[1.0, zoom_factor[1], zoom_factor[2]],
order=orders[1])
label = ndimage.zoom(label,
zoom=[zoom_factor[0], 1.0, 1.0],
order=orders[3])
data[case_name]['label'] = label
assert sorted(label.shape) == sorted(original_shape)
if print_verbose:
print('after de-resample', label.shape, type(label[0, 0, 0]),
label.max(), label.min())
save_case_label(step, case_name, label, data, save_nii_dir,
save_mat_dir)
def load_nii(root, phase, out_path, depth_transpose=False):
step = 'load'
save_nii_dir, save_mat_dir = make_save_dirs(step, out_path)
paths = get_nii_paths(root, phase)
data = {}
for image_p in tqdm(sorted(paths)):
if Params['data_name'] == 'Segthor':
case_id = image_p.split('/')[-1][8:10]
case_name = 'Case%02d' % (int(case_id))
elif Params['data_name'] == 'LA':
case_name = image_p.split('/')[-1][:6]
nii_image = sitk.ReadImage(image_p)
spacing = nii_image.GetSpacing()
image = nib.load(image_p).get_data()
if depth_transpose:
image = np.transpose(image, (2, 0, 1))
spacing = (spacing[2], spacing[0], spacing[1])
data[case_name] = {'image': image, 'spacing': spacing}
print('load ', image.shape, type(image[0, 0, 0]), image.max(),
image.sum() / image.shape[0] / image.shape[1] / image.shape[2])
print(case_name, spacing)
print(image_p)
save_case(step, case_name, image, data, save_nii_dir, save_mat_dir)
def reverse_align(out_path):
step = 'align'
prev_step = 'crop'
save_nii_dir, save_mat_dir = make_save_dirs(step, out_path)
data = load_data_label(prev_step)
# check max_shape consistency
max_shape = Cparams['max_shape']
original_ss = Cparams['original_ss']
target_spacing = Cparams['target_spacing']
for case_name in data.keys():
label = data[case_name]['label']
label_shape = np.array(label.shape)
label_center = label_shape // 2
if print_verbose:
print(case_name)
# calculate resampled shape
original_spacing, original_shape = original_ss[case_name][
'spacing'], original_ss[case_name]['shape']
resampled_shape = [
int(
round(original_shape[ith] * original_spacing[ith] /
target_spacing[ith]))
for ith in range(len(original_shape))
]
original_center = np.array(resampled_shape) // 2
start = (label_center - original_center).astype(int)
end = (start + resampled_shape).astype(int)
if print_verbose:
print('before ', label.shape, type(label[0, 0, 0]), label.max(),
label.min())
# de-align: depadding to separate shape (after resample)
label_new = label[start[0]:end[0], start[1]:end[1],
start[2]:end[2]].copy()
label = label_new
data[case_name]['label'] = label
if print_verbose:
print('after reverse-align', label.shape, type(label[0, 0, 0]),
label.max(), label.min())
save_case_label(step, case_name, label, data, save_nii_dir,
save_mat_dir)
def crop_roi(out_path, changed_axis=2):
step = 'crop'
prev_step = 'align'
save_nii_dir, save_mat_dir = make_save_dirs(step, out_path)
data = load_data(prev_step)
zoom_factor, roi, down_n_all = Cparams['zoom_factor'], Cparams[
'roi'], Cparams['down_n_all']
for case_name in data.keys():
image = data[case_name]['image']
print(case_name)
print('before ', image.shape, type(image[0, 0, 0]), image.max(),
image.sum())
orders = [3, 1, 3, 1]
image = ndimage.zoom(image,
zoom=[1.0, zoom_factor[1], zoom_factor[2]],
order=orders[0])
# downsample along Z axis
if changed_axis != 2:
image = ndimage.zoom(image,
zoom=[zoom_factor[0], 1.0, 1.0],
order=orders[2])
print('after down', image.shape, type(image[0, 0, 0]), image.max(),
image.sum())
if Params['data_name'] == 'LA' and roi[1] - roi[0] > image.shape[0]:
new_shape = (roi[1] - roi[0], roi[3] - roi[2], roi[5] - roi[4])
image_new = np.zeros(new_shape).astype(np.int16)
zlen = roi[1] - roi[0]
z_start = (zlen - image.shape[0]) // 2
image_new[z_start:z_start +
image.shape[0], :, :] = image[:, roi[2]:roi[3],
roi[4]:roi[5]]
image = image_new
else:
image = image[roi[0]:roi[1], roi[2]:roi[3], roi[4]:roi[5]]
data[case_name]['image'] = image
print('after crop', image.shape, type(image[0, 0, 0]), image.max(),
image.sum())
save_case(step, case_name, image, data, save_nii_dir, save_mat_dir)
return zoom_factor, roi, down_n_all
def load_pred(load_dir, out_path, phase):
step = 'load'
save_nii_dir, save_mat_dir = make_save_dirs(step, out_path)
# load labels from given dir (record their ids)
paths = sorted(glob(load_dir + '*.nii.gz'))
data = {}
for image_p in sorted(paths):
case_name = image_p.split('/')[-1][:6]
label = nib.load(image_p).get_data()
data[case_name] = {
'label': label,
}
if print_verbose:
print('load ', label.shape, type(label[0, 0, 0]), label.max(),
label.min())
print(case_name, image_p)
save_case_label(step, case_name, label, data, save_nii_dir,
save_mat_dir)
def normalize_intensity(out_path, data_name='Promise'):
step = 'normalize'
prev_step = 'crop'
save_nii_dir, save_mat_dir = make_save_dirs(step, out_path)
data = load_data(prev_step)
for case_name in data.keys():
image = data[case_name]['image']
print(case_name)
print('before', image.shape, type(image[0, 0, 0]), image.max(),
image.sum(), image.min())
if data_name in ['Promise', 'promise']:
image_value = image[image != 0]
mean = image_value.mean()
std = image_value.std()
image = (image - mean) / std
data[case_name]['image'] = image
elif data_name in ['Segthor', 'segthor']:
clip_window = [-986, 271]
mean, std = 20.78, 180.50
image[image < clip_window[0]] = clip_window[0]
image[image > clip_window[1]] = clip_window[1]
image = (image - mean) / std
data[case_name]['image'] = image
elif data_name in ['LA']:
mean = image.mean()
std = image.std()
image = (image - mean) / std
# data[case_name]['image'] = image # We forget this re-assignment in current exps, yet it doesn't bring hurt anayway.
print('mean std', mean, std, image.shape,
image.shape[0] * image.shape[1] * image.shape[2])
print('after ', image.shape, type(image[0, 0, 0]), image.max(),
image.min())
save_case(step, case_name, image, data, save_nii_dir, save_mat_dir)
def resample(out_path):
step = 'resample'
prev_step = 'load'
save_nii_dir, save_mat_dir = make_save_dirs(step, out_path)
data = load_data(prev_step)
for case_name in data.keys():
image = data[case_name]['image']
spacing = data[case_name]['spacing'][0]
print(case_name)
print('before', image.shape, type(image[0, 0, 0]), image.max(),
image.sum() / image.shape[0] / image.shape[1] / image.shape[2])
if max(spacing) / min(spacing) > 3:
anisotropic = True
else:
anisotropic = False
if re_sample[0] != 0 or re_sample[1] != 0 or re_sample[2] != 0:
zoom_factor = [spacing[i] / re_sample[i] for i in range(3)]
if zoom_factor != [1, 1, 1]:
if anisotropic:
orders = [3, 1, 0, 0]
else:
orders = [3, 1, 3, 1]
assert np.argmax(spacing) == 0 # z axis
image = ndimage.zoom(
image,
zoom=[1.0, zoom_factor[1], zoom_factor[2]],
order=orders[0])
image = ndimage.zoom(image,
zoom=[zoom_factor[0], 1.0, 1.0],
order=orders[2])
data[case_name]['image'] = image
print('after ', image.shape, type(image[0, 0, 0]), image.max(),
image.sum() / image.shape[0] / image.shape[1] / image.shape[2])
print(case_name, spacing)
save_case(step, case_name, image, data, save_nii_dir, save_mat_dir)
def reverse_crop(out_path):
step = 'crop'
prev_step = 'load'
save_nii_dir, save_mat_dir = make_save_dirs(step, out_path)
data = load_data_label(prev_step)
zoom_factor, roi, down_n_all = Cparams['zoom_factor'], Cparams[
'roi'], Cparams['down_n_all']
# calculate downsampled max_shape
max_shape = Cparams['max_shape']
downsampled_max_shape = [
int(round(ia * ib)) for ia, ib in zip(max_shape, zoom_factor)
]
for case_name in data.keys():
label = data[case_name]['label']
if print_verbose:
print(case_name)
print('before ', label.shape, type(label[0, 0, 0]), label.max(),
label.min())
# de-crop: original unified shape
label_new = np.zeros(downsampled_max_shape).astype(np.uint8)
if Params['data_name'] == 'LA':
zlen = roi[1] - roi[0]
roi_start = (zlen - label_new.shape[0]) // 2
label_new[:, roi[2]:roi[3],
roi[4]:roi[5]] = label[roi_start:roi_start +
label_new.shape[0], :, :]
else:
label_new[roi[0]:roi[1], roi[2]:roi[3],
roi[4]:roi[5]] = label.copy()
label = label_new
if print_verbose:
print('after de-crop', label.shape, type(label[0, 0, 0]),
label.max(), label.min())
# de-downsample: original max shape
label_new = np.zeros(max_shape).astype(np.uint8)
orders = [3, 1, 3, 1]
label = ndimage.zoom(
label,
zoom=[1.0, 1 / zoom_factor[1], 1 / zoom_factor[2]],
order=orders[1])
if zoom_factor[0] != 1:
label = ndimage.zoom(label,
zoom=[1 / zoom_factor[0], 1.0, 1.0],
order=orders[3])
if label_new.shape != label.shape:
if label.shape[1] > label_new.shape[1] and label.shape[
2] > label_new.shape[2]: # for trachea
# label_new = label[:, :label_new.shape[1], :label_new.shape[2]] #00
# label_new = label[:, :label_new.shape[1], 1:] # 01
# label_new = label[:, 1:, :label_new.shape[1]] # 10
_t1, _t2 = label.shape[1] - label_new.shape[1], label.shape[
2] - label_new.shape[2]
label_new = label[:, _t1:,
_t2:] # for trachea, we find this way better # 11
elif label.shape[1] < label_new.shape[1] and label.shape[
2] < label_new.shape[2]: # for prostate
# label_new[:, :label.shape[1], :label.shape[2]] = label # 00
# label_new[:, :label.shape[1], 1:] = label # 01
# label_new[:, 1:, :label.shape[1]] = label # 10
_t1, _t2 = label_new.shape[1] - label.shape[
1], label_new.shape[2] - label.shape[2]
label_new[:, 1:, 1:] = label # 11
else:
ipdb.set_trace()
print(label_new.shape, label.shape)
raise NotImplementedError
label = label_new
if print_verbose:
print('after de-downsample', label.shape, type(label[0, 0, 0]),
label.max(), label.min())
data[case_name]['label'] = label
save_case_label(step, case_name, label, data, save_nii_dir,
save_mat_dir)