Converting RDS to LDS?

Is there an existing issue for this?

• I have searched the existing issues

Problem summary

I am attempting to follow the workflow you used for https://github.com/fepegar/torchio/issues/647 , but I am still getting issues. I am loading from two different data sources for semantic segmentation one .nii and one .npy. When I run it through the transformation below and call the first time of the training set I get an error. Thank you in advance.

Code for reproduction

def numpy_reader(path):
    return np.load(path), np.eye(4)

subjects=[]
for subject_path in merged_list:
    subject = tio.Subject({"MRI":tio.ScalarImage(subject_path[0]), "Label":tio.LabelMap(subject_path[1],reader=numpy_reader)})
    subjects.append(subject)

process = tio.Compose([
            tio.ToCanonical(),
            tio.CropOrPad((192, 192, 192)), #This crops it to the desired metirc
            tio.RescaleIntensity((-1, 1)) # This normalizes the data
            ])
augmentation = tio.RandomAffine(scales=(0.9, 1.1), degrees=(-10, 10))
val_transform = process
train_transform = tio.Compose([process, augmentation])

train_dataset = tio.SubjectsDataset(subjects, transform=train_transform)
val_dataset = tio.SubjectsDataset(subjects, transform=val_transform)

train_dataset[0] #error

Actual outcome

I get an error that says my two images are in different space which is true from my run-through.

When I check to see how different they look very similar

Error messages

RuntimeError: More than one value for "direction" found in subject images:
{'Label': (1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0),
 'MRI': (-1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0)}
....
RuntimeError: As described above, some images in the subject are not in the same space. You probably can use the transforms ToCanonical and Resample to fix this, as explained at https://github.com/fepegar/torchio/issues/647#issuecomment-913025695

Expected outcome

Subject(Keys: (‘MRI’, ‘Label’); images: 2)

System info

# flake8: noqa

import re
import sys
import platform
import torchio
import torch
import numpy
import SimpleITK as sitk


sitk_version = re.findall('SimpleITK Version: (.*?)\n', str(sitk.Version()))[0]

print('Platform:  ', platform.platform())
print('TorchIO:   ', torchio.__version__)
print('PyTorch:   ', torch.__version__)
print('SimpleITK: ', sitk_version)
print('NumPy:     ', numpy.__version__)
print('Python:    ', sys.version)

[Jeffotter]

I don’t know what changed but my error has moved to: RuntimeError: More than one value for “origin” found in subject images: {‘Label’: (0.0, 0.0, 0.0), ‘MRI’: (-91.0, -126.0, -72.0)} I’ll look into this further

[fepegar]

Thanks for the info. In that case, that’s definitely the issue. I would stick to a format more suitable for medical images such as NIfTI or NRRD.