Specialized scans

This tutorial will cover how HippUnfold can be applied to non-standard data including ex-vivo scans, super-high resolution data (eg. <0.3mm isotropic), non-MRI 3D imaging data, or scans where a corresponding whole-brain T1w image is not available.

We will show how the available flags can be adapted for these use-cases with several worked examples.

Case 1: super high resolution

In this example, we have only a limited field of view covering the hippocampus, and the resolution and contrast do not closely match the training data of HippUnfold (0.3-1.0mm isotropic T1w, T2w, or DWI data). This could be ex-vivo MRI data, or it could even be 3D microscopy data as in our recent 3D BigBrain publication. Thus we don’t expect HippUnfold’s inbuilt UNet to be successful in segmenting hippocampal tissue before unfolding, and we do not want to downsample our data to accommodate HippUnfold’s usual UNet and unfolding workflow in space-corobl (which consists of 0.3mm isotropic resampling cropped coronally-oblique to the hippocampus).

This will require manual segmentation of hippocampal grey matter, SRLM, and neighbouring structures, though in the future we hope to include models trained with higher resolution data (and contrasts more common in ex-vivo scanning). This should be done according to the protocol outlined here or, more recently, the video example here. This manual segmentation file should have the _dseg suffix.

Here is an example of what the input directory might look like:

└── sub-001/
    ├── sub-001_hemi-R_desc-hippo_T2w.nii.gz
    └── sub-001_hemi-R_desc-hippo_dseg.nii.gz

This can be unfolded with the command:

hippunfold . PATH_TO_OUTPUT_DIR participant --modality cropseg \
--path_cropseg exvivo/sub-{subject}/sub-{subject}_hemi-{hemi}_desc-hippo_dseg.nii.gz \
--hemi R --skip_inject_template_labels

Explanation: --modality cropseg informs HippUnfold that the input manual segmentation should not be resampled and UNet does not need to be run. Because of a limitation in bids parsing for the hemi entity, we need to use the generic path input, --path_cropseg in this case, making sure we use the {subject} and {hemi} wildcards in the filename. Output files will be named with space-corobl because HippUnfold is coded to effectively treat all files as already being in this space. We need the --hemi R to prevent HippUnfold looking for both hemispheres. Finally, because this segmentation was performed manually on very high resolution data, we can optionally consider skipping the template shape injection step with --skip_inject_template_labels. Template shape injection can fix minor errors in segmentation from UNet or from an imperfect manual rater, at the cost of smoothing out some details of the hippocampus due to the fact that it uses deformable registration with inherent smoothness contraints.

Note that because we are not resampling to the CITI168 template or using UNet, the T2w image in this example is effectively not being used at all. Instead, the provided manual segmentation makes up the basis for unfolding.

Case 2: one ex-vivo hemisphere

In this example, we have a single hemisphere that was scanned ex-vivo at a nearly standard resolution and T2w contrast. Because the resolution and contrast are similar to the HippUnfold training data, we expect UNet will work and so we don’t need to perform manual segemntation. However, due to gross deformations and the missing hemisphere, we don’t expect this sample to register well to the standard CITI168 template. Thus, we will need to manually resample the image to the CITI168 template prior to running HippUnfold, focusing in particular on aligning the hippocampus. Once done, we may have a directory like this:

└── sub-001/
    ├── sub-001_hemi-R_desc-exvivo_T2w.nii.gz
    ├── sub-001_hemi-R_affine-exvivo-to-CITI168_xfm.txt
    └── sub-001_hemi-R_desc-exvivo_space-CITI168_T2w.nii.gz

Note that only the last file is needed for unfolding:

hippunfold . PATH_TO_OUTPUT_DIR participant --output_spaces corobl --hemi R --no_reg_template \
--path_T2w PATH_TO_EXVIVO_DIR/sub-001/sub-001_hemi-R_desc-exvivo_space-CITI168_T2w.nii.gz \
--output_spaces corobl

Here we need to use --path_T2w to specify which input should be used, and --no_reg_template to specify that it is already in space-CITI168. In this case, we also specified --output_spaces corobl. This is not needed, but is useful when we are interested in only the hippocampus as space-corobl is higher resolution and cropped more nicely around the hippocampus then the original scan, making it a good space to perform subsequent analyses. Alternatively, outputs can be transformed back to the original space using the inverted transform sub-001_hemi-R_affine-exvivo-to-CITI168_xfm.txt.

This same usage could also be applied in a standard MRI case where no T1w image is available.