[PET] Improve how to 4D images with 1 volume are handled

[alexandreroutier]

Hi!

Following Discussion #870, it turns out that 4D images with 1 volume are not taken into account in PET pipelines:

• pet-surface*:
  

  clinica/clinica/pipelines/pet_surface/pet_surface_utils.py

  Lines 949 to 952 in 4e5f5d9

  	if len(img.shape) == 4:
  	error_msg = (
  	f"Clinica does not handle 4D volumes for {subject_id} | {session_id}"
  	)

• pet-volume*:
  

  clinica/clinica/pipelines/pet_volume/pet_volume_utils.py

  Lines 13 to 16 in 4e5f5d9

  	if len(img.shape) == 4:
  	error_msg = (
  	f"Clinica does not handle 4D volumes for {image_id.replace('_', ' | ')}"
  	)

A solution would be:

    if len(img.shape) == 4:
        if img.shape[3] > 1:
            error_msg = (
                f"Clinica does not handle 4D volumes for {image_id.replace('_', ' | ')}"
            )

(I have a doubt but a oneliner if len(img.shape) == 4 and img.shape[3] > 1: could also work)

Best,
Alex

[NicolasGensollen]

Thanks for looking into this @alexandreroutier !
I'm wondering whether we should also explicitly squeeze the image in this case (might be done automatically by tools downstream, but doesn't hurt I think...). Maybe something like this:

def _check_img_3d(img):
    """Load a 3D nifti image, raise if not 3D...""" 
    img = nib.load(img)
    dim = len(img.shape) 
    if dim != 3:
        if dim == 4 and img.shape[3] == 1:
            data = img.get_fdata()
            klass = img.__class__
            header = copy.deepcopy(img.header)
            return klass(data[:, :, :, 0], img.affine, header=header)
        error_msg = ( 
             f"Clinica does not handle {dim}D volumes for {image_id.replace('_', ' | ')}" 
         ) 
        cprint(error_msg, lvl="error")
        raise NotImplementedError(error_msg)
    return img

[ghisvail]

I'm wondering whether we should also explicitly squeeze the image in this case

In that case, I'd suggest to rename this function to something more explicit. check implies a boolean or exception raising function. Here we are actually loading and returnung the data.

(might be done automatically by tools downstream, but doesn't hurt I think...).

From the old nipype code base perpsective, I think you are right.

For the pydra pipelines, since they rely on checksumming the input files to handle caching, the 4D to 3D squeezing should probably happen in a separate node which will keep track of the old and new (squeezed) volume files.

On a side note, since the PET volumes likely came from preprocessed ADNI files, maybe it's also the job of our converters to ensure volumes are squeezed and (re)written to 3D. There was also a similar point raised in #622 about normalizing the data type used for writing.

[NicolasGensollen]

In that case, I'd suggest to rename this function to something more explicit. check implies a boolean or exception raising function. Here we are actually loading and returnung the data.

Yes, I agree the name is misleading, I'll change that in #872

[alexandreroutier]

I didn't think about it yesterday but you are raising the fact that we don't known how third party apps are handling this particular case 🤔 Generating an 'adequate' 3D image would be the simplest solution.

(For info, this NotImplementedError was added in order to solve #70)