We provide a set of tools packaged as purple-mri for segmentation, parcellation and registration of ultra high-resolution (< 300 microns) postmortem human brain hemisphere at 7 tesla t2w MRI at native subject-space resolution. This pipeline leverages advances in both deep learning and classical surface-based modeling techniques to produce parcellations in any atlas used in neuroimaging. The developed method allows us to perform vertex-wise analysis in the template space and thereby link morphometry measures with pathology measurements derived from histology.
Check out the project page and our latest papers here and here.
In particular, purple-mri allows you to do the following:
- obtain an initial 10 label segmentation: cortical GM, normal appearing WM, WMH, medial temporal lobe, corpus callosum, ventricles, caudate, putamen, globus pallidus and thalamus
- obtain surface-based native subject-space parcellation based on different brain atlases such as: DKT, Economo, Schaeffer etc
- create population specific volumetric and surface-based templates
- perform exvivo to invivo registration in volumetric intensity space
- perform surface-to-surface registration between exvivo or invivo
- perform intensity-based registration between 9.4 tesla MTL to 7 tesla whole hemsiphere registration
- perform vertex-wise and group-wise generalized linear modeling analysis for exvivo subject population for morphometry and histology
purple-mri follows a series of steps making use of bash scripts and Docker.
Perform bias correction and image normalization/standardization. We use N4BiasFieldCorrection as part of the CLI tool ANTs and c3d. We highly recommend using the option of an input mask in N4BiasFieldCorrection which can be obtained via corase threhsolding.
Here is a sample script.
Currently, we have two Docker images. The first image provides the segmentation and the second employs Nighres/CRUISE for post-hoc topology correction. Please follow the [link](https://github.com/Pulkit-Khandelwal/upenn-picsl-brain-ex-vivo/blob/main/exvivo-segm-demo-docker.md for detailed instructions on how to use Docker to get the segmentations. Some key commands are emphasized here:
Place the pre-processed image(s) (with a suffix _0000.nii.gz to your filenames) in a folder named data_for_inference within your working directory is /your/working/directory.
docker pull pulks/docker_hippogang_exvivo_segm:v${LATEST_TAG}
docker run --gpus all --privileged -v /your/working/directory/:/data/exvivo/ -it pulks/docker_hippogang_exvivo_segm:v${LATEST_TAG} /bin/bash -c "bash /src/commands_nnunet_inference.sh ${OPTION}" >> logs.txt
You will see the output in /your/working/directory/data_for_inference/output_from_nnunet_inference.
Next, correct for topology so that adjoining gyri and sulci are clearly separated. Copy the segmentations from output_from_nnunet_inference to a folder data_for_topology_correction in your wokring directory.
docker pull pulks/docker_nighres:v1.0.0
docker run -v /your/working/directory/:/data/cruise_files/ -it pulks/docker_nighres:v1.0.0 /bin/bash -c "bash /data/prepare_cruise_files.sh"
# Locally run the file to get the final combined label file. See instructions [here](https://github.com/Pulkit-Khandelwal/upenn-picsl-brain-ex-vivo/blob/main/exvivo-segm-demo-docker.md) and links referenced therein.
bash clean_labels_final.sh
Once, you have obtained an initial 10-label topology corrected volumetric segmentation, you can proceed to the surface-based pipeline to obtain parcellations based on your favorite atlas. This step will be on your local machine. No GPUs required. To do this, you should have FreeSurfer installed locally. We have used FreeSurfer version 7.4.0 on linux obtained from here. Moreover, there are some Python dependencies which can be found in the dependencies.txt file and installed using pip.
Run the following file which calls in several bash scripts which prepares the data, computes appropriate transformations and re-orients the images, corrects surface topology, and perform the parcellation into Desikan-Killiany-Tourville (DKT), Schaefer, Glasser and the Von Economo-Koskinos atlases.
For the surface-based modeling step, we assume that all the hemishpehrs are right hemispheres. So, we suggest flipping the left t2w MRI and its corresponding segmentation to left using the following c3d command: c3d image_left.nii.gz -flip y image_right_flipped.nii.gz
Clone the current repository and the run the following script run_surface_pipeline.sh from within the purple_mri folder which takes the following mandatory arguments:
freesurfer_path: path to the FreeSurfer installtion
working_dir: directory which will have the outputs for each subject stored
mri_path: mri images path
segm_path: 10-label segmentation path
external_atlases_path: directory with files for other atlases
num_threads: number of threads
Place your t2w MRI in a folder mri_path and the intial deep learning-based segmentations in segm_path.
Make sure your mri images and segmentation files have the same names ending with .nii.gz.
Place the fsaverage in the working_dir folder.
cd purple_mri
bash run_surface_pipeline.sh freesurfer_path working_dir mri_path segm_path external_atlases_path num_threads
You have the option to run the surface-based pipeline as a pip package available at PyPI and can be installed as:
pip3 install purple-mri
Or, using the latest development code:
git clone https://github.com/Pulkit-Khandelwal/purple-mri.git
cd purple-mri
pip3 install pkg_src
Place the fsaverage in the working_dir folder. The cli options remain the same as above but this time you have to pass in the last argument as the path to autodet.gw.stats.binary.rh.dat.
Once, installed you can run, purple_mri as:
python3 -m purple_mri freesurfer_path working_dir mri_path segm_path external_atlases_path num_threads /path/to/autodet.gw.stats.binary.rh.dat
We build intensity-based volumetric templates using the greedy tool. The required binaries (for Linux) and the scripts are located in the the intensity_template within the scripts directory. Follow the instructions here.
Script exvivo_invivo_greedy_registration.sh to register in vivo (t1w) and ex vivo (t2w) MRI is located in the folder scripts. We use greedy to register the segmentations of in vivo aseg+aparc labels derived from FreeSurfer and 10-label initial deep learning segmentation of postmortem MRI. The warps are then used to regsiter the MRIs.
We perform vertex-wise analysis in fsaverage space to fit a generalized linear model (GLM) between cortical thickness (mm) and with global ratings of amyloid-β, Braak staging, CERAD, and semiquantitative ratings of the medial temporal lobe (MTL) neuronal loss and tau pathology, with age, sex and postmortem interval (PMI) as covariates. You can follow the steps detailed here. The same can be repeated for the Deformation-based morphometry.
- Our method has been developed to work on a single exvivo hemisphere.
- The deep learning-based segmentation was primarily trained on 7T t2w MRI. We have tested the model on t2* flash as well and it works pretty well but, if need be, we recommend re-training the model with some manual labels obtained on t2* flash MRI.
- Khandelwal, P., Duong, M. T., Sadaghiani, S., Lim, S., Denning, A. E., Chung, E., ... & Yushkevich, P. A. (2024). Automated deep learning segmentation of high-resolution 7 tesla postmortem MRI for quantitative analysis of structure-pathology correlations in neurodegenerative diseases. Imaging Neuroscience, 2, 1-30.
- Khandelwal, P., Duong, M. T., Fuentes, C., Denning, A., Trotman, W., Ittyerah, R., ... & Yushkevich, P. A. (2024). Surface-based parcellation and vertex-wise analysis of ultra high-resolution ex vivo 7 tesla MRI in neurodegenerative diseases. arXiv preprint arXiv:2403.19497.