def preprocess_unlabeled(files_dir, preprocessed_dir):
utils.makedir(preprocessed_dir)
# bb_size = [80, 200, 200] # will be extended by [10,20,20] per side for augmented
# get_left_and_right_BB()
cases = sorted(os.listdir(files_dir))
cases = [x for x in cases if 'case' in x]
for i in range(0, len(cases)):
# if cases[i] == 'case_00223':
out_dir = os.path.join(preprocessed_dir, cases[i])
utils.makedir(out_dir)
print(cases[i])
img = sitk.ReadImage(
os.path.join(files_dir, cases[i], 'imaging.nii.gz'))
img = utils.resampleImage(img, SPACING, sitk.sitkLinear, 0,
sitk.sitkInt16)
# apply region of interest that is guided by bones
bb, smaller_bb = get_threshold_bounding_box(img, 500)
if img.GetSize()[0] > BB_SIZE[0]:
# if img.GetSize()[0] > BB_SIZE[1]+50: # if x-axis is very long, it might be a whole body scan
cropped_img = sitk.RegionOfInterest(
img, [smaller_bb[3], smaller_bb[4], smaller_bb[5]],
[smaller_bb[0], smaller_bb[1], smaller_bb[2]])
sitk.WriteImage(cropped_img, 'cropped_' + str(i) + '.nrrd')
lung_img = sitk.BinaryThreshold(cropped_img, -900, -700, 1,
0) # lung -900, -700
sitk.WriteImage(lung_img, 'lung_img_thresh' + str(i) + '.nrrd')
lung_img = utils.getLargestConnectedComponents(lung_img)
sitk.WriteImage(lung_img, 'lung_img_cc' + str(i) + '.nrrd')
# get boundingbox of lung:
bb_lung = getBB(lung_img, 1)
start_y_w_bb_lung = cropped_img.TransformIndexToPhysicalPoint(
[bb_lung[0], bb_lung[1], bb_lung[2]])
start_y_w_img = img.TransformPhysicalPointToIndex(
start_y_w_bb_lung)
# print('bb_lung: ', bb_lung)
# values for final ROI
start_X = max(
0, bb_lung[0] + bb_lung[3]
) # - 0.1*BB_SIZE[0]) # start 20 voxel above bottom of lung
size_X = BB_SIZE[0]
if start_X + BB_SIZE[0] > img.GetSize()[0]:
start_X = img.GetSize()[0] - size_X - 1
start_y = max(0, (start_y_w_img[1] + bb_lung[4] -
BB_SIZE[1])) ## according to bones bounding box
# start_y = max(0, bb[1] + bb[4] - BB_SIZE[1])
start_z_r = int((bb[2] + bb[5]) / 2)
start_z_l = max(0, int((bb[2] + bb[5]) / 2) - BB_SIZE[2])
else:
### pad volume in x direction
img = utils.pad_volume(img,
target_size_x=BB_SIZE[0],
padValue=utils.getMinimum(img))
# values for final ROI
start_X = 0 # start 20 voxel above bottom of lung
start_y = max(0, (bb[1] + bb[4] + 20 -
BB_SIZE[1])) ## according to bones bounding box
start_z_r = int((bb[2] + bb[5]) / 2)
start_z_l = max(0, int((bb[2] + bb[5]) / 2) - BB_SIZE[2])
left_img = get_final_roi(img, start_X, start_y, start_z_l, BB_SIZE[0],
BB_SIZE[1], BB_SIZE[2])
right_img = get_final_roi(img, start_X, start_y, start_z_r, BB_SIZE[0],
BB_SIZE[1], BB_SIZE[2])
sitk.WriteImage(left_img, os.path.join(out_dir, 'img_left.nrrd'))
sitk.WriteImage(right_img, os.path.join(out_dir, 'img_right.nrrd'))
def preprocess_centered_BB():
utils.makedir(preprocessed_dir)
# get_left_and_right_BB()
cases = sorted(os.listdir(files_dir))
cases = [x for x in cases if 'case' in x]
# bbs_left = np.load('bbs_left.npy')
# bbs_right = np.load('bbs_right.npy')
for i in range(0, len(cases)):
# if cases[i] == 'case_00002':
out_dir = os.path.join(preprocessed_dir, cases[i])
utils.makedir(out_dir)
print(cases[i])
img = sitk.ReadImage(
os.path.join(files_dir, cases[i], 'imaging.nii.gz'))
img = utils.resampleImage(img, SPACING, sitk.sitkLinear, 0,
sitk.sitkInt16)
segm = sitk.ReadImage(
os.path.join(files_dir, cases[i], 'segmentation.nii.gz'))
# threhsold segmentation such that no tumor tissue is available anymore
segm = sitk.BinaryThreshold(segm, 1, 2, 1, 0)
## split image in left and right half
[size_x, size_y, size_z] = img.GetSize()
img_left = sitk.RegionOfInterest(
img, [size_x, size_y, int(size_z / 2)], [0, 0, 0])
img_right = sitk.RegionOfInterest(
img, [size_x, size_y, int(size_z / 2)],
[0, 0, int(size_z / 2) - 1])
img_left, img_right = normalizeIntensities(img_left, img_right)
segm_left = utils.resampleToReference(segm,
img_left,
sitk.sitkNearestNeighbor,
defaultValue=0,
out_dType=sitk.sitkUInt8)
segm_right = utils.resampleToReference(segm,
img_right,
sitk.sitkNearestNeighbor,
defaultValue=0,
out_dType=sitk.sitkUInt8)
sitk.WriteImage(segm_left, 'segm_left.nrrd')
sitk.WriteImage(segm_right, 'segm_right.nrrd')
size_l, centroid_l = get_segmentation_statistics(
segm_left, False, False)
size_r, centroid_r = get_segmentation_statistics(
segm_right, False, False)
start_l_x = max(0, centroid_l[0] - int(BB_SIZE[0] / 2))
start_l_y = max(0, centroid_l[1] - int(BB_SIZE[1] / 2))
start_l_z = max(0, centroid_l[2] - int(BB_SIZE[2] / 2))
start_r_x = max(0, centroid_r[0] - int(BB_SIZE[0] / 2))
start_r_y = max(0, centroid_r[1] - int(BB_SIZE[1] / 2))
start_r_z = max(0, centroid_r[2] - int(BB_SIZE[2] / 2))
segm_l_final = get_final_roi(segm_left, start_l_x, start_l_y,
start_l_z, BB_SIZE[0], BB_SIZE[1],
BB_SIZE[2])
segm_r_final = get_final_roi(segm_right, start_r_x, start_r_y,
start_r_z, BB_SIZE[0], BB_SIZE[1],
BB_SIZE[2])
img_l_final = get_final_roi(img_left, start_l_x, start_l_y, start_l_z,
BB_SIZE[0], BB_SIZE[1], BB_SIZE[2])
img_r_final = get_final_roi(img_right, start_r_x, start_r_y, start_r_z,
BB_SIZE[0], BB_SIZE[1], BB_SIZE[2])
print(size_l, size_r)
# write images
sitk.WriteImage(img_l_final, os.path.join(out_dir, 'img_left.nrrd'))
sitk.WriteImage(img_r_final, os.path.join(out_dir, 'img_right.nrrd'))
sitk.WriteImage(segm_l_final, os.path.join(out_dir, 'segm_left.nrrd'))
sitk.WriteImage(segm_r_final, os.path.join(out_dir, 'segm_right.nrrd'))
print(labelled_files_lst[0:10])
np.random.seed(5)
# np.random.seed(1234) for fold 1 0.25
np.random.shuffle(labelled_files_lst)
labelled_num_considrd = labelled_files_lst[:int(labelled_train_num * p)]
remaining_labelled = set(labelled_files_lst).difference(
set(labelled_num_considrd))
counter = 0
data_path = '/cache/suhita/data/hippocampus/fold_' + str(
fold_num) + '_P' + str(p)
utils.makedir(data_path)
utils.makedir(os.path.join(data_path, 'imgs'))
utils.makedir(os.path.join(data_path, 'GT'))
for i in labelled_num_considrd:
# name = labelled_files_lst[i]
print(i, counter)
np.save(
os.path.join(data_path, 'imgs',
str(counter) + '.npy'),
np.expand_dims(
np.load(
os.path.join(labelled_imgs_path,
i.replace('.nii.gz', '.npy'))), -1))
np.save(
os.path.join(data_path, 'GT',
str(counter) + '.npy'),
def preprocess():
utils.makedir(preprocessed_dir)
bb_size = [80, 200,
200] # will be extended by [10,20,20] per side for augmented
# get_left_and_right_BB()
cases = sorted(os.listdir(files_dir))
cases = [x for x in cases if 'case' in x]
bbs_left = np.load('bbs_left.npy')
bbs_right = np.load('bbs_right.npy')
for i in range(0, len(cases)):
# if cases[i] == 'case_00002':
out_dir = os.path.join(preprocessed_dir, cases[i])
utils.makedir(out_dir)
print(cases[i])
img = sitk.ReadImage(
os.path.join(files_dir, cases[i], 'imaging.nii.gz'))
img = utils.resampleImage(img, [3.0, 0.75, 0.75], sitk.sitkLinear, 0,
sitk.sitkInt16)
segm = sitk.ReadImage(
os.path.join(files_dir, cases[i], 'segmentation.nii.gz'))
# apply region of interest that is guided by bones
bb, smaller_bb = get_threshold_bounding_box(img, 500)
if bb[3] > 100:
cropped_img = sitk.RegionOfInterest(
img, [smaller_bb[3], smaller_bb[4], smaller_bb[5]],
[smaller_bb[0], smaller_bb[1], smaller_bb[2]])
# sitk.WriteImage(cropped_img, 'cropped_' + str(i) + '.nrrd')
lung_img = sitk.BinaryThreshold(cropped_img, -900, -700, 1,
0) # lung -900, -700
# sitk.WriteImage(lung_img, 'lung_img_thresh' + str(i) + '.nrrd')
lung_img = utils.getLargestConnectedComponents(lung_img)
# sitk.WriteImage(lung_img, 'lung_img_cc' + str(i) + '.nrrd')
# get boundingbox of lung:
bb_lung = getBB(lung_img, 1)
# print('bb_lung: ', bb_lung)
# values for final ROI
start_X = max(0, bb_lung[0] + bb_lung[3] -
20) # start 20 voxel above bottom of lung
size_X = 100
if start_X + 100 > img.GetSize()[0]:
start_X = img.GetSize()[0] - size_X - 1
img = sitk.RegionOfInterest(
img, [size_X, img.GetSize()[1],
img.GetSize()[2]], [start_X, 0, 0])
# img = sitk.RegionOfInterest(img, [size_X, bb[4], bb[5]],
# [start_X, bb[1], bb[2]])
sitk.WriteImage(img, 'final_cropped_' + str(i) + '.nrrd')
else:
### pad volume in x direction
img = utils.pad_volume(img,
target_size_x=100,
padValue=utils.getMinimum(img))
## split image in left and right half
[size_x, size_y, size_z] = img.GetSize()
img_left = sitk.RegionOfInterest(
img, [size_x, size_y, int(size_z / 2)], [0, 0, 0])
img_right = sitk.RegionOfInterest(
img, [size_x, size_y, int(size_z / 2)],
[0, 0, int(size_z / 2) - 1])
sitk.WriteImage(img_left, 'img_left_temp.nrrd')
sitk.WriteImage(img_right, 'img_right_temp.nrrd')
# get ROI for left and right kidney
roi_left_start = [
max(0, bbs_left[i, 0] - 10),
max(0, bbs_left[i, 1] - 10),
max(0, bbs_left[i, 2] - 10)
]
roi_right_start = [
max(0, bbs_right[i, 0] - 10),
max(0, bbs_right[i, 1] - 10),
max(0, bbs_right[i, 2] - 10)
]
# apply ROI to image
img_left_final = get_final_roi(img_left, roi_left_start[0],
roi_left_start[1], roi_left_start[2],
bb_size[0] + 20, bb_size[1] + 20,
bb_size[2] + 20)
img_right_final = get_final_roi(img_right, roi_right_start[0],
roi_right_start[1], roi_right_start[2],
bb_size[0] + 20, bb_size[1] + 20,
bb_size[2] + 20)
segm_left_final = utils.resampleToReference(segm,
img_left_final,
sitk.sitkNearestNeighbor,
defaultValue=0,
out_dType=sitk.sitkUInt8)
segm_right_final = utils.resampleToReference(segm,
img_right_final,
sitk.sitkNearestNeighbor,
defaultValue=0,
out_dType=sitk.sitkUInt8)
# write images
sitk.WriteImage(img_left_final, os.path.join(out_dir, 'img_left.nrrd'))
sitk.WriteImage(img_right_final, os.path.join(out_dir,
'img_right.nrrd'))
sitk.WriteImage(segm_left_final, os.path.join(out_dir,
'segm_left.nrrd'))
sitk.WriteImage(segm_right_final,
os.path.join(out_dir, 'segm_right.nrrd'))
str(fold_num) + '.npy')
print(train_fold[0:10])
nr_samples = train_fold.shape[0]
# np.random.seed(5)
np.random.seed(5)
np.random.shuffle(train_fold)
print(train_fold[0:10])
labelled_num_considrd = train_fold[:int(nr_samples * perc)]
remaining_labelled = set(train_fold).difference(set(labelled_num_considrd))
counter = 0
data_path = '/cache/suhita/data/kits/fold_' + str(fold_num) + '_P' + str(perc)
utils.makedir(data_path)
for i in labelled_num_considrd:
# name = labelled_files_lst[i]
print(i, counter)
fold_name = 'case_' + str(counter)
utils.makedir(os.path.join(data_path, fold_name))
copyfile(os.path.join(labelled_path, i, 'img_left.npy'),
os.path.join(data_path, fold_name, 'img_left.npy'))
copyfile(os.path.join(labelled_path, i, 'img_right.npy'),
os.path.join(data_path, fold_name, 'img_right.npy'))
copyfile(os.path.join(labelled_path, i, 'segm_left.npy'),
os.path.join(data_path, fold_name, 'segm_left.npy'))
copyfile(os.path.join(labelled_path, i, 'segm_right.npy'),
os.path.join(data_path, fold_name, 'segm_right.npy'))
counter = counter + 1