Problem with creating a subject and resampling

[Elmisiry99]

Hello,
I am working on the prostate dataset of the Medical decathlon datasets and I wanted first to store each example in a subject. When I create the subject there are no errors, but if I try to print anything inside this subject I get the following error:

Traceback (most recent call last):
  File "/home/students/elmisiry/Prostate_network/Data_Inspection.py", line 37, in <module>
    sub1 = transform(sub)
  File "/work/scratch/elmisiry/miniconda3/envs/pytorchnew/lib/python3.8/site-packages/torchio/transforms/transform.py", line 124, in __call__
    subject = copy.copy(subject)
  File "/work/scratch/elmisiry/miniconda3/envs/pytorchnew/lib/python3.8/copy.py", line 84, in copy
    return copier(x)
  File "/work/scratch/elmisiry/miniconda3/envs/pytorchnew/lib/python3.8/site-packages/torchio/data/subject.py", line 70, in __copy__
    value = copy.copy(value)
  File "/work/scratch/elmisiry/miniconda3/envs/pytorchnew/lib/python3.8/copy.py", line 84, in copy
    return copier(x)
  File "/work/scratch/elmisiry/miniconda3/envs/pytorchnew/lib/python3.8/site-packages/torchio/data/image.py", line 190, in __copy__
    tensor=self.data,
  File "/work/scratch/elmisiry/miniconda3/envs/pytorchnew/lib/python3.8/site-packages/torchio/data/image.py", line 203, in data
    return self[DATA]
  File "/work/scratch/elmisiry/miniconda3/envs/pytorchnew/lib/python3.8/site-packages/torchio/data/image.py", line 182, in __getitem__
    self.load()
  File "/work/scratch/elmisiry/miniconda3/envs/pytorchnew/lib/python3.8/site-packages/torchio/data/image.py", line 478, in load
    tensor, affine = self.read_and_check(paths[0])
  File "/work/scratch/elmisiry/miniconda3/envs/pytorchnew/lib/python3.8/site-packages/torchio/data/image.py", line 504, in read_and_check
    tensor, affine = self.reader(path)
  File "/work/scratch/elmisiry/miniconda3/envs/pytorchnew/lib/python3.8/site-packages/torchio/data/io.py", line 25, in read_image
    result = _read_sitk(path)
  File "/work/scratch/elmisiry/miniconda3/envs/pytorchnew/lib/python3.8/site-packages/torchio/data/io.py", line 57, in _read_sitk
    data, affine = sitk_to_nib(image, keepdim=True)
  File "/work/scratch/elmisiry/miniconda3/envs/pytorchnew/lib/python3.8/site-packages/torchio/data/io.py", line 321, in sitk_to_nib
    affine = get_ras_affine_from_sitk(image)
  File "/work/scratch/elmisiry/miniconda3/envs/pytorchnew/lib/python3.8/site-packages/torchio/data/io.py", line 339, in get_ras_affine_from_sitk
    rotation_ras = np.dot(FLIPXY_33, rotation_lps)
UnboundLocalError: local variable 'rotation_lps' referenced before assignment

The code looks as follows:

imagesdir = sorted(os.listdir(image_dir))
labelsdir = sorted(os.listdir(label_dir))
for idx, (image_path, label_path) in enumerate(zip(imagesdir, labelsdir)):
        if not label_path.startswith('.'):
            sub = tio.Subject(image=tio.ScalarImage(image_name),
                              map=tio.LabelMap(label_name))
            img = sub['image']
            print(sub['image'])
            transform = tio.Resample((0.625,0.625,3.6))
            sub1 = transform(sub)
            #print("After resampling: ", sub1['image'].shape)

Please note that I found this problem by first trying to output the shape of an image after applying resampling but I got this error so I tried to simplify first checking whether the subjects are defined or not.
Do you have an idea what this might be?
I wanted to ask a second question about 2D slices. Torchio doesn't work on 2D images, right? If I want to work only on 2D slices, I would then have to define the subjects with 3D images and index the slices I want or is there a way to apply transformations only on 2D slices? For example in the example above, I would like to resample only the in-plane dimensions without doing anything to the depth as I will not use it. Do I still have to apply the resamling on the whole volume?
Thanks a lot in advance

[fepegar]

Do you have an idea what this might be?

It looks like an old bug. What TorchIO version are you using? Can you update and try again?

Torchio doesn't work on 2D images, right?

It does. It's just optimized for 3D.

I would like to resample only the in-plane dimensions without doing anything to the depth as I will not use it. Do I still have to apply the resampling on the whole volume?

I think resampling the whole volume might be faster. I guess there are different choices:

1. Pass the original spacing to the last dimension so it's untouched
2. Resample slice by slice
3. Add support to do one of the things above if a special arg is passed for one of the spacings, e.g. None

The simplest solution, as it doesn't need a modification of the library, is number 1. What do you think?

[Elmisiry99]

Whether it will have an effect on the quality of resapmling of the 2D slices. I don’t know if I am expressing the question in the right way or not.
I will try to formulate in another way:
Considering the three choices you stated above, If I use the first choice on the whole 3D volume and index a certain axial slice(let us denote it with slice_1). Then I access the same slice in the original volume that was not resampled, and resample it to the target spacing,just like the second option you wrote above( let us call this slice_2). Then both slice_1 and slice_2 will have the same size, but will they have the same intensity values for each pixel/voxel. I hope I expressed it in an understandable way.

[Elmisiry99]

I had by the way version 0.18.57 but I updated to the 0.18.72 and it worked and the images were resampled successfully. Is is possible to access the new spacing of the subject? I checked the keys of the ScalarImage in the subject and it contains : ['data', 'affine', 'path', 'stem', 'type'], but but no spcaing.

[fepegar]

I think the results should be the same for both approaches. I would try if I were you. Maybe it depends on the type of interpolation, and small numerical differences might show up.

---

spacing is a property of the Image class that is computed dynamically from the associated affine matrix. If you do my_image.spacing or just print the image, you'll see it.

[Elmisiry99]

Thanks a lot for the help :).

[fepegar]

No worries!