 |
|---|
| Motion tracking using our pipeline. Left: without shape refinement; Right: with shape refinement |
The code in this repository implements 4Dsegment, a pipeline for carrying out deep learning segmentation in UK Biobank, non-rigid co-registration, mesh generation and motion tracking using raw grey-scale cardiac MRI data in NIfTI format. The implementation was first trained using manual annotations and then deployed on pulmonary hypertension (PH) patients to produce segmentation labels and computational meshes. The whole process is fully automated without any manual input.
- Test code in small sample (28-May-2019)
The files in this repository are organized into 3 directories:
- code : contains base functions for segmentation, co-registration, mesh generation, and motion tracking:
- code entrance - code/DMACS.py
- deep learning segmentation with the pre-trained model - code/deepseg.py
- co-registration to fit a high-resolution model - code/p1&2processing.py
- fitting meshes to high-resolution model - code/meshfitting.py
- useful image processing functions used in the pipeline - code/image_utils.py
- downsample mesh resolution while remain its geometry - code/decimation.py
- model : contains a tensorflow model pre-trained on ~400 manual annotations on PH patients
- data : data download address, which contains three sample datasets (4D NIfTI) on which functions from the
codedirectory can be run. You should download the data and place them into this folder.
To run the code in the code directory, we provide a Docker image with all the necessary dependencies pre-compiled.
A Docker image is available on dockerhub https://hub.docker.com/r/jinmingduan/segmentationcoregistration. This image contains a base Ubuntu linux operating system image set up with all the libraries required to run the code (e.g. Tensorflow, nibabel, opencv, etc.). The image also contains pre-compiled IRTK (https://github.com/BioMedIA/IRTK) and MIRTK (https://github.com/BioMedIA/MIRTK) for image registration, as well as external data on which the code can be run.
Click the download button, unzip to your desktop and name the top-level folder 4Dsegment.
Go to /data and download the sample images (nifti format) from the URL in the text file.
For Windows 10 Pro first install Docker. Windows 10 Home users will require Docker toolbox.
Ensure you have the C drive selected as a shared drive in Docker settings (or in VirtualBox on W10 Home).
To visualise the segmentations download ITKsnap.
In W10 open PowerShell from the Windows search box (Win + X then I), in macOS navigate Finder > Applications > Utilities > Terminal, or in Linux any terminal can be used. Then download the pre-compiled image:
docker pull jinmingduan/segmentationcoregistration:latest
docker images
should show jinmingduan/segmentationcoregistration on the list of images on your local system
Note the path to the folder on your desktop eg /c/Users/home/Desktop/4Dsegment and substitute <folder-path> within this command:
docker run -it --rm -v <folder-path>/data/:/data -v <folder-path>/code/:/code -v <folder-path>/model/:/model jinmingduan/segmentationmeshmotion /bin/bash
launches an interactive linux shell terminal that gives users access to the image's internal file system. The command passes the code, model and data into the docker container such that the code can be run within the container.
Typing next
ls -l
will list all the folders in the working directory of the Docker image. You should see the 3 main folders code, data and model, which contain the same files as the corresponding folders with the same name in this github repository.
Typing next
export LD_LIBRARY_PATH=/lib64
will point you where the compiled libraries are
Typing next
cd /code
will bring you to the directory where the code is saved
Finally doing
python DMACS.py --coreNo 8 --irtk True
will run the code using 8 CPU cores on your local computer (change the number to fit your machine) with irtk registration toolbox.
Once the pipeline is finished, under the root directory of each subject, you have three nifti files, i.e., lvsa_.nii.gz, lvsa_ED_enlarged_SR.nii.gz and lvsa_ES_enlarged_SR.nii.gz, and two segmentations, i.e., PHsegmentation_ED.gipl and PHsegmentation_ES.gipl. lvsa_.nii.gz is the original 4D raw data and PHsegmentation_ED.gipl and PHPHsegmentation_ES.gipl are segmentations of lvsa_ED_enlarged_SR.nii.gz and lvsa_ES_enlarged_SR.nii.gz. Note that these segmentations are smooth, high-resolution bi-ventricular three-dimensional models.
You also have meshes (txt files) for left and right ventricles at ED and ES under the root directory. For example, lv_myoed_curvature.txt records the curvature of each vertex on myocardium of left ventricle at ED. lv_myoed_wallthickness.txt records the wall thickness of each vertex on epicardium of left ventricle at ED. lv_myoed_signeddistances.txt records the sign distance of each vertex on epicardium of left ventricle at ED, by referring to a template. lv_myoed_curvature.txt, lv_myoes_wallthickness.txt and lv_myoes_signeddistances.txt have the same meanings for left ventricle at ES. There are also counterparts for right ventricle at ED and ES.
In addition, the pipeline also produces the folders of dofs, segs, sizes, tmps, vtks and motion under the root directory. Apart from motion folder, the files in other folders are intermediate results, which may not be useful for sequential analysis. In motion folder, you have 20 computational meshes (both vtk and txt files) for a complete cardiac cycle. In each of 20 meshes, only spatial locations of vertices are recorded. Vertex spatial position (x, y and z) on the same row in the txt files corresponds to the same anatomical location across the cardiac cycle.
If you find this software useful for your project or research. Please give some credits to authors who developed it by citing some of the following papers. We really appreciate that.
[1] Duan J, Bello G, Schlemper J, Bai W, Dawes TJ, Biffi C, de Marvao A, Doumou G, O’Regan DP, Rueckert D. Automatic 3D bi-ventricular segmentation of cardiac images by a shape-refined multi-task deep learning approach. IEEE Transactions on Medical Imaging (2019).
[2] Bello GA, Dawes TJW, Duan J, Biffi C, de Marvao A, Howard LSGE, Gibbs JSR, Wilkins MR, Cook SA, Rueckert D, O'Regan DP. Deep learning cardiac motion analysis for human survival prediction. Nature Machine Intelligence 1, 95–104 (2019).