def __getitem__(self, index):
"""Return the specific index pair subvolume (input, ground truth).
Args:
index (int): Subvolume index.
"""
coord = self.indexes[index]
seg_pair, _ = self.handlers[coord['handler_index']]
# Clean transforms params from previous transforms
# i.e. remove params from previous iterations so that the coming transforms are different
# Use copy to have different coordinates for reconstruction for a given handler
metadata_input = imed_loader_utils.clean_metadata(
copy.deepcopy(seg_pair['input_metadata']))
metadata_gt = imed_loader_utils.clean_metadata(
copy.deepcopy(seg_pair['gt_metadata']))
# Run transforms on images
stack_input, metadata_input = self.transform(sample=seg_pair['input'],
metadata=metadata_input,
data_type="im")
# Update metadata_gt with metadata_input
metadata_gt = imed_loader_utils.update_metadata(
metadata_input, metadata_gt)
# Run transforms on images
stack_gt, metadata_gt = self.transform(sample=seg_pair['gt'],
metadata=metadata_gt,
data_type="gt")
# Make sure stack_gt is binarized
if stack_gt is not None and not self.soft_gt:
stack_gt = imed_postpro.threshold_predictions(stack_gt, thr=0.5)
shape_x = coord["x_max"] - coord["x_min"]
shape_y = coord["y_max"] - coord["y_min"]
shape_z = coord["z_max"] - coord["z_min"]
# add coordinates to metadata to reconstruct volume
for metadata in metadata_input:
metadata['coord'] = [
coord["x_min"], coord["x_max"], coord["y_min"], coord["y_max"],
coord["z_min"], coord["z_max"]
]
subvolumes = {
'input':
torch.zeros(stack_input.shape[0], shape_x, shape_y, shape_z),
'gt':
torch.zeros(stack_gt.shape[0], shape_x, shape_y, shape_z)
if stack_gt is not None else None,
'input_metadata':
metadata_input,
'gt_metadata':
metadata_gt
}
for _ in range(len(stack_input)):
subvolumes['input'] = stack_input[:, coord['x_min']:coord['x_max'],
coord['y_min']:coord['y_max'],
coord['z_min']:coord['z_max']]
if stack_gt is not None:
for _ in range(len(stack_gt)):
subvolumes['gt'] = stack_gt[:, coord['x_min']:coord['x_max'],
coord['y_min']:coord['y_max'],
coord['z_min']:coord['z_max']]
return subvolumes
def __getitem__(self, index):
"""Return the specific processed data corresponding to index (input, ground truth, roi and metadata).
Args:
index (int): Slice index.
"""
seg_pair_slice, roi_pair_slice = self.indexes[index]
# Clean transforms params from previous transforms
# i.e. remove params from previous iterations so that the coming transforms are different
metadata_input = imed_loader_utils.clean_metadata(
seg_pair_slice['input_metadata'])
metadata_roi = imed_loader_utils.clean_metadata(
roi_pair_slice['gt_metadata'])
metadata_gt = imed_loader_utils.clean_metadata(
seg_pair_slice['gt_metadata'])
# Run transforms on ROI
# ROI goes first because params of ROICrop are needed for the followings
stack_roi, metadata_roi = self.transform(sample=roi_pair_slice["gt"],
metadata=metadata_roi,
data_type="roi")
# Update metadata_input with metadata_roi
metadata_input = imed_loader_utils.update_metadata(
metadata_roi, metadata_input)
# Run transforms on images
stack_input, metadata_input = self.transform(
sample=seg_pair_slice["input"],
metadata=metadata_input,
data_type="im")
# Update metadata_input with metadata_roi
metadata_gt = imed_loader_utils.update_metadata(
metadata_input, metadata_gt)
if self.task == "segmentation":
# Run transforms on images
stack_gt, metadata_gt = self.transform(sample=seg_pair_slice["gt"],
metadata=metadata_gt,
data_type="gt")
# Make sure stack_gt is binarized
if stack_gt is not None and not self.soft_gt:
stack_gt = imed_postpro.threshold_predictions(stack_gt,
thr=0.5)
else:
# Force no transformation on labels for classification task
# stack_gt is a tensor of size 1x1, values: 0 or 1
# "expand(1)" is necessary to be compatible with segmentation convention: n_labelxhxwxd
stack_gt = torch.from_numpy(seg_pair_slice["gt"][0]).expand(1)
data_dict = {
'input': stack_input,
'gt': stack_gt,
'roi': stack_roi,
'input_metadata': metadata_input,
'gt_metadata': metadata_gt,
'roi_metadata': metadata_roi
}
return data_dict