HRpQCT-Knee-Seg

This repo contains an installable package for doing automated HR-pQCT knee segmentation. You can use it to do inference with an existing model, or to train a new model on your own data. The package is built on top of the bonelab and bonelab-pytorch-lightning packages.

---

Publication

Pre-print: https://doi.org/10.1101/2024.05.20.24307643

This repository contains code for automating peri-articular analysis of bone in knee HR-pQCT images. There are utilities for training, and doing inference with, segmentation models using deep learning (UNet variants specifically), for atlas-based registration, and for ROI generation. While the other repo contains all of the associated code for the publication, this repo contains only the code necessary to use the workflow, is installable, has command line hooks, and has some additional utilities for automating the generation of slurm and shell scripts to run the analysis. If you want to use the method yourself, this is the repo you want.

This work is not yet peer-reviewed.

Atlases and trained models

The atlases and trained models are available on zenodo.

---

Installation

1. Go to https://github.com/Bonelab/bonelab-pytorch-lightning and follow the instructions to install the bonelab and bonelab-pytorch-lightning packages
2. Clone this repository and install it. You can do this by running the following commands:

cd ~/Projects # or wherever you keep your repositories
git clone git@github.com:Bonelab/HRpQCT-Knee-Seg.git
cd HRpQCT-Knee-Seg
pip install -e .

NOTE 1: The -e flag is for "editable" mode. This means that if you make changes to the code in this repository, you don't have to reinstall the package. The changes will be reflected in the package immediately.

NOTE 2: The above install command will use the requirements.txt file and will install the default versions of all of the dependencies. There are two special cases with special requirements file:

requirements_arc_c11.1.txt is for using the NVIDIA A100 GPU nodes on the ARC cluster at the University of Calgary, and also on Groot. You can install the package with this requirements file by running:

pip install -r requirements_arc_c11.1.txt -e .

requirements_thegnu_c10.2.txt is for using the NVIDIA Quadro P6000 GPUs on the TheGNU remote server in the lab. You can install the package with this requirements file by running:

pip install -r requirements_thegnu_c10.2.txt -e .

If you have other package version requirements, feel free to make your own requirements file and install the package with that. However please note that it's import that bonelab-pytorch-lightning and this package be installed with a consistent requirements file.

---

Command Line Apps

This is a list of all of the command-line apps that get installed when you install this package. For more specific usage instructions, run a command with the -h flag.

Command	Description
hrkAIMs2NIIs	Convert an AIM and optionally, its associated masks, to NIIs
hrkMask2AIM	Convert a mask to AIM, requires a base AIM that the mask will be lined up on.
hrkMasks2AIMs	Convert multiple masks to AIMs, requires a base AIM that the masks will be lined up on.
hrkParseLogs	Parse/collate the logs from a set of pytorch lightning model training runs.
hrkCombineROIMasks	Combine multiple ROI masks into a single compartmental mask.
hrkGenerateAffineAtlas	Use a set of images and masks to construct an average atlas with affine registration.
hrkInferenceEnsemble	Perform inference on an image uby ensembling multiple segmentation models.
hrkIntersectMasks	Compute the intersection of two binary masks.
hrkPostProcessSegmentation	Use morphological filtering operations to post-process a predicted bone compartment segmentation - designed for knee HR-pQCT images specifically.
hrkMaskImage	Given an image and a mask, will dilate the mask by some amount and then set the image voxels to zero outside of the dilated mask.
hrkPreProcess2DSlices	Preprocess a directory of AIMs to generate 2D slice patch samples to minimize file IO and processing time when training a segmentation model.
hrkPreProcess2dot5DSliceStacks	Preprocess a directory of AIMs to generate 2.5D stacked slice patch samples to minimize file IO and processing time when training a segmentation model.
hrkPreProcess3DPatches	Preprocess a directory of AIMs to generate 3D patch samples to minimize file IO and processing time when training a segmentation model.
hrkPreProcess3DPatchesFromNPZ	Preprocess a directory of AIMs that have been converted to NPZs to generate 3D patch samples to minimize file IO and processing time when training a segmentation model.
hrkTrainSeGAN_CV	Train a SeGAN segmentation model usinf cross-validation.
hrkTrainSegResNetVAE_CV	Train a SegResNetVAE using cross-validation.
hrkTrainUNet_CV	Train a UNet (or variant) using cross-validation.
hrkTrainSeGAN_Final	Train a single SeGAN using 100% of the provided data as training data.
hrkTrainSegResNetVAE_Final	Train a single SegResNetVAE using 100% of the provided data as training data.
hrkTrainUNet_Final	Train a single UNet (or variant) using 100% of the provided data as training data.
hrkTrainSeGAN_TransferCV	Train a SeGAN using cross-validation, starting from a SeGAN that has been trained on another dataset.
hrkTrainSegResNetVAE_TransferCV	Train a SegResNetVAE using cross-validation, starting from a SeGAN that has been trained on another dataset.
hrkTrainUNet_TransferCV	Train a UNet (or variant) using cross-validation, starting from a SeGAN that has been trained on another dataset.
hrkVisualize2DPanning	Reads in an image and masks and generates a 2D video or GIF that pans through the image, for qualitative data checking, motion scoring, etc.
hrkGenerateROIs	Generate peri-articular ROIs from a bone compartment segmentation and a atlas-based compartmental segmentation.
hrkCrossSectional	Create bash/slurm files to perform all steps for a cross-sectional study design.
hrkLongitudinal	Create bash/slurm files to perform all steps for a longitudinal study design.

---

Segmenting Knee Images

Regardless of the study design, the first steps are to move your data from the DATA disk to the PROJECTS disk on the VMS system, and then transfer your data to ARC. This package does not provide scripts for automating any of that.

Cross-sectional Data

Your data directory must be organized in the following way:

<STUDY NAME>
├── aims
├── atlas_registrations
├── model_masks
├── niftis
├── roi_masks
└── visualizations

In a cross-sectional study design, we will process all images independently from each other. So the steps listed below are to be applied to each image.

Steps:

1. Put the AIMs in the aims directory.
2. Convert the AIMs to NIfTI format using the hrkAIMs2NIIs command, and put the outputs in the niftis directory.
3. Perform inference on the images using the hrkInferenceEnsemble command, and put the outputs in the model_masks directory.
4. Post-process the segmentations using the hrkPostProcessSegmentation command, and put the outputs in the model_masks directory.
5. Convert the post-processed masks to AIMs using the hrkMasks2AIMs command, and put the outputs in the roi_masks directory.
6. If the image is from a LEFT knee, then mirror the NII using the blImageMirror command, and put the output in the niftis directory.
7. Register the (if LEFT, mirrored) image to the atlas using the blRegistrationDemons command, and put the output in the atlas_registrations directory.
8. Transform the atlas masks to the image using the blRegistrationApplyTransform command and put the output in the atlas_registrations directory.
9. If the image is from a LEFT knee, then mirror the transformed atlas masks using the blImageMirror command, and put the output in the atlas_registrations directory.
10. Generate the ROI masks using the hrkGenerateROIs command, and put the output in the roi_masks directory.
11. Convert the peri-articular ROI masks to AIMs using the hrkMasks2AIMs command, and put the outputs in the roi_masks directory.
12. Visualize the segmentations using the hrkVisualize2DPanning command, and put the outputs in the visualizations directory.
13. Visually check the visualization outputs to make sure the ROI masks are OK - you may need to go back and recrop your AIMs, or you may need to do some corrections to the bone compartment segmentations and rerun parts of the workflow. You can also use these visualizations for doing motion scoring.
14. Transfer all *.AIM files from roi_masks back to the VMS system, and do your analysis.
15. Clean up after yourself by deleting everything that is not an *.AIM file.

Longitudinal Data

Your data directory must be organized in the following way:

<STUDY NAME>
├── aims
├── atlas_registrations
├── model_masks
├── niftis
├── registrations
├── roi_masks
└── visualizations

In a longitudinal study design, we will process images as time series. The steps below are to be applied to each set of images that are the same subject, the same side, the same bone, etc. at each time point.

Steps:

1. Put the AIMs in the aims directory.
2. Convert the AIMs to NIfTI format using the hrkAIMs2NIIs command, and put the outputs in the niftis directory.
3. Perform inference on the images using the hrkInferenceEnsemble command, and put the outputs in the model_masks directory.
4. Post-process the segmentations using the hrkPostProcessSegmentation command, and put the outputs in the model_masks directory.
5. Convert the post-processed masks to AIMs using the hrkMasks2AIMs command, and put the outputs in the roi_masks directory.
6. If the images are from a LEFT knee, then mirror the NIIs using the blImageMirror command, and put the outputs in the niftis directory.
7. Register the (if LEFT, mirrored) images to the atlas using the blRegistrationDemons command, and put the output in the atlas_registrations directory.
8. Transform the atlas masks to the images using the blRegistrationApplyTransform command and put the outputs in the atlas_registrations directory.
9. If the images are from a LEFT knee, then mirror the transformed atlas masks using the blImageMirror command, and put the outputs in the atlas_registrations directory.
10. Use the respective post-processed bone compartment masks to mask the images using the hrkMaskImages command, and put the outputs in the niftis directory.
11. Perform a longitudinal registration with the masked images using the blRegistrationLongitudinal command, and put the outputs in the registrations directory.
12. Delete the masked images.
13. Transform all follow-up bone compartment segmentations and atlas-based compartmental segmentations to the baseline image using the blRegistrationApplyTransform command, and put the outputs in the registrations directory.
14. Find the intersection of all atlas-based compartmental segmentations in the baseline frame using the hrkIntersectMasks command and put the output in the registrations directory.
15. Generate the peri-articular ROI masks in the baseline reference frame using the intersected atlas masks and the bone compartment segmentations that have been transformed to the baseline reference frame, using the hrkGenerateROIs command, and put the output in the roi_masks directory.
16. Transform the follow-up peri-articular ROI masks from the baseline frame to their respective follow-up frames using the blRegistrationApplyTransform command, and put the outputs in the roi_masks directory.
17. Convert the peri-articular ROI masks to AIMs using the hrkMasks2AIMs command, and put the outputs in the roi_masks directory.
18. Visualize the segmentations using the hrkVisualize2DPanning command, and put the outputs in the visualizations directory.
19. Visually check the visualization outputs to make sure the ROI masks are OK - you may need to go back and recrop your AIMs, or you may need to do some corrections to the bone compartment segmentations and rerun parts of the workflow. You can also use these visualizations for doing motion scoring.
20. Transfer all *.AIM files from roi_masks back to the VMS system, and do your analysis.
21. Clean up after yourself by deleting everything that is not an *.AIM file.

Miscellaneous

The steps don't need to be done in the exact order presented above. For example, you can do the bone compartment segmentation steps at the same time as you register the image to the atlas and get the compartmental masks (and also do the longitudinal registrations in parallel). Also, only model inference will use a GPU, so every other step should be executed on a cluster node or machine where you are not blocking someone else's access to a GPU. Many of the CPU-intensive steps are slow, so it's not recommended to just run the whole workflow all the way through on a GPU node on ARC, or a GPU machine such as TheGNU or Groot.

---

TODO:

1. Add a command line utility for each step in the pipeline (DONE)
2. Add a command line utility that you can run in a different directory that will create all of the shell and slurm files required for either cross-sectional or longitudinal processing of HR-pQCT knee images
3. Improve the documentation in this repo to properly explain how to use the utilities and how everything works (DONE)
4. Add tests and github actions CI/CD
5. Long term goal - perhaps add command line utilities that link up to the preprocessing and training scripts so that someone could retrain the models if they had more data and wanted to - or so you could use this to train models on entirely different datasets.. the approach here isn't necessarily specific to just knees
6. Might be nice to have sphinx auto-documentation, but that's post-thesis stuff