deepmriprep 0.3.0

Neural network-based preprocessing of MRI data: Prepping brain images in seconds 🔥

Disclaimer: deepmriprep is not related to fMRIPrep or sMRIPrep and is not part of the NiPreps framework

deepmriprep runs all preprocessing steps needed for Voxel-based Morphometry (VBM) of T1w MR images:

• brain extraction (via deepbet)
• affine registration (via torchreg)
• tissue segmentation
• nonlinear registration
• smoothing

Using neural networks, it only takes ~10 seconds on a GPU (~100 seconds without a GPU) per image 🔥

Additionally, deepmriprep can also run atlas registration needed for Region-based Morphometry (RBM).

Installation 🛠️

For GPU-acceleration (NVIDIA needed), PyTorch should be installed first via the proper install command 🔥

deepmriprep can be easily installed via

pip install deepmriprep

The deepmriprep-gui can look grainy in conda environments. To fix that run

conda install -c conda-forge tk=*=xft_*

Usage 💡

After installation, there are three ways to use deepmriprep

1. deepmriprep-gui runs the Graphical User Interface (GUI)

2. deepmriprep-cli runs the Command Line Interface (CLI)

deepmriprep-cli -bids path/to/bids

3. Run deepmriprep directly in Python

from deepmriprep import run_preprocess

# Apply to BIDS dir. (saves to 'path/to/bids/derivatives')
run_preprocess(bids_dir='path/to/bids')
# Apply to list of files and save to output directory
input_paths = ['path/to/input/sub-1_t1w.nii.gz', 'path/to/input/sub-2_t1w.nii.gz']
run_preprocess(input_paths, output_dir='path/to/output')
# Apply to list of files and save to custom output filepaths
output_paths = [{'tiv': 'outpath1/tivsub-1_t1w.csv', 'p0': 'outpath2/p0sub-1_t1w.nii'},
                {'tiv': 'outpath3/tivsub-2_t1w.csv', 'p0': 'outpath4/p0sub-2_t1w.nii'}]
run_preprocess(input_paths, output_paths=output_paths)

Besides the three shown options to specify input and output paths, run_preprocess has the arguments

• outputs: Output modalities which can be either 'all', 'vbm', 'rbm' or a custom list of output strings
• dir_format: Output directory structure which can be either 'sub', 'mod', 'cat' or 'flat'
• no_gpu: Avoids GPU utilization if set to True (set to False per default for GPU-acceleration 🔥)

outputs set to

• 'all' all output modalities are saved
• 'vbm' the outputs tiv, mwp1, mwp2, s6mwp1 and s6mwp2 (+affine_loss&warp_mse, for QC) are saved
• 'rbm' all available atlases (including regions tissue volumes +affine_loss&warp_mse, for QC) are saved

If output_dir is set, dir_format set to

• 'sub' results in e.g. 'outpath/sub-1/tivsub-1.csv' and 'outpath/sub-1/p0sub-1.nii.gz'
• 'mod' results in e.g. 'outpath/tiv/tivsub-1.csv' and 'outpath/p0/p0sub-1.nii.gz'
• 'cat' results in e.g. 'outpath/sub-1/label/tivsub-1.csv' and 'outpath/sub-1/mri/p0sub-1.nii.gz'
• 'flat' results in e.g. 'outpath/tivsub-1.csv' and 'outpath/p0sub_1.nii.gz'

Quality Control 🔍

Install NiftiView to create PNGs or GIFs of deepmriprep_outputs.csv via

python qc_report.py -csv /path/to/deepmriprep_outputs.csv --gif -c mwp1 mwp2 -n 4

for visual inspection (the above commands creates GIFs of the 'mwp1' and 'mwp2' files of the 4 inputs with the highest warp_mse_value).

The PNGs/GIFs show the percentiles of affine_loss_value and warp_mse_value (compared to >10,000 MRIs from OpenNeuro). High percentile values should be visually inspected to avoid faulty data in further analysis!

Here are two example PNGs indicating "normal" (left) and faulty data (right, missing frontal parts)

Additionally, install the NiftiView-App for quick visual inspection!

Tutorial 🧑‍🏫

In short, deepmriprep (consisting of only ~500 lines of code) internally calls the .run method of the Preprocess class, which sequentially calls the methods .run_bet to .run_atlas_register (see deepmriprep/preprocess.py).

Click here to view the tutorial and get a more detailed look behind the scenes 👀

Citation ©️

If you use deepmriprep in your research, please cite:

@inproceedings{deepmriprep,
Author = {Lukas Fisch, Nils R. Winter, Janik Goltermann, Carlotta Barkhau, Daniel Emden, Jan Ernsting, Maximilian Konowski, Ramona Leenings, Tiana Borgers, Kira Flinkenflügel, Dominik Grotegerd, Anna Kraus, Elisabeth J. Leehr, Susanne Meinert, Frederike Stein, Lea Teutenberg, Florian Thomas-Odenthal, Paula Usemann, Marco Hermesdorf, Hamidreza Jamalabadi, Andreas Jansen, Igor Nenadic, Benjamin Straube, Tilo Kircher, Klaus Berger, Benjamin Risse, Udo Dannlowski, Tim Hahn},
Title = {deepmriprep: Voxel-based Morphometry (VBM) Preprocessing via Deep Neural Networks},
Year = {2024},
eprint = {arXiv:2408.10656}
}

Outputs 📋

When "Output Modalities" is set to "custom" in the deepmriprep-gui, all output strings are shown:

Most output strings follow the CAT12 naming convention of output files.

Here are descriptions for all output strings:

Input

• 't1': T1-weighted MR image

Brain Extraction

• 'mask': Brain mask
• 'brain': 't1' after brain mask is applied
• 'tiv_bet': Total intracranial volume in cm³ based on brain extraction

Affine Registration

• 'affine': Affine matrix moving 'brain' into template space (compatible with F.affine_grid)
• 'translate', 'rotation', 'zoom', 'shear': Tranformation parameters the affine is composed of
• 'mask_large': Affine applied to 'mask' with 0.5mm resolution
• 'brain_large': Affine applied to 'brain' with 0.5mm resolution
• 'affine_loss': Loss value of the affine registration

Tissue Segmentation

• 'p0_large': Tissue segmentation map of 'brain_large'
• 'p0': Tissue segmentation map in 't1' image space (moved 'p0_large')

Tissue Probabilities

• 'nogm': Small area around the brain stem that is masked in subsequent GM output
• 'gmv': Gray matter (GM) volume in cm³
• 'wmv': White matter (WM) ""
• 'csfv': Cerebrospinal fluid (CSF) ""
• 'tiv': Total intracranial volume in cm³ based on tissue segmentation (gmv + wmv + csfv)
• 'rel_gmv': Proportion of GM volume relative to the total intracranial volume
• 'rel_wmv': Proportion of WM ""
• 'rel_csfv': Proportion of CSF ""
• 'p1': GM tissue probability in 't1' image space
• 'p2': WM ""
• 'p3': CSF ""
• 'p1_large': GM tissue probability based on 'p0_large'
• 'p2_large': WM ""
• 'p3_large': CSF ""
• 'p1_affine': GM tissue probability based on 'p0_large' in template resolution
• 'p2_affine': WM ""
• 'p3_affine': CSF ""

Nonlinear Registration

• 'wj_affine': Jacobian determinant of the affine matrix
• 'warp_xy': Forward warping field (compatible with F.grid_sample)
• 'warp_yx': Backward warping field (compatible with F.grid_sample)
• 'warp_mse': Mean squared error between warped p and warp template
• 'wj_': Jacobian determinant of forward warping field
• 'v_xy': Forward velocity field (compatible with F.grid_sample)
• 'v_yx': Backward velocity field (compatible with F.grid_sample)
• 'wp1': Forward warped 'p1'
• 'mwp1': Modulated (multiplied with 'wj_'), forward warped 'p1'
• 's6mwp1': Smoothed with 6mm kernel, modulated (multiplied with 'wj_'), forward warped 'p1'
• ...analogous for 'wp2', 'mwp2', 's6mwp2', 's8mwp1', ...

Atlases

• 'aal3': Registered AAL3 atlas in 't1' image space
• 'anatomy3': Registered Anatomy3 ""
• 'cobra': Registered Cobra atlas ""
• 'hammers': Registered Hammers atlas ""
• 'ibsr': Registered IBSR atlas ""
• 'julichbrain': Registered Julichbrain atlas ""
• 'lpba40': Registered LPBA40 atlas ""
• 'mori': Registered Mori atlas ""
• 'neuromorphometrics': Registered Neuromorphometrics ""
• 'suit': Registered SUIT ""
• 'thalamic_nuclei': Registered Thalamic Nuclei ""
• 'thalamus': Registered Thalamus ""
• 'Schaefer2018_100Parcels_17Networks_order': Registered Schaefer 100 ""
• 'Schaefer2018_200Parcels_17Networks_order': Registered Schaefer 200 ""
• 'Schaefer2018_400Parcels_17Networks_order': Registered Schaefer 400 ""
• 'Schaefer2018_600Parcels_17Networks_order': Registered Schaefer 600 ""

For each atlas, there also can be outputted two more modalities:

• 'aal3_affine': Registered AAL3 atlas in template space
• 'aal3_volumes': GM, WM and CSF volume in mm³ per region of the AAL3 atlas
• ...analogous 'anatomy3_affine', 'anatomy3_volumes', 'cobra_affine', 'cobra_volumes' ...