Automating Model Generation For Image-Based Cardiac Flow Simulation

This repository contains the source code for our paper:

Kong, F., and Shadden, S. C. (August 7, 2020). "Automating Model Generation for Imagebased Cardiac Flow Simulation." ASME. J Biomech Eng. doi: https://doi.org/10.1115/1.4048032

The code repository consists of two parts

• Deep-learning based automatic segmentation of 3D CT/MR image data using an ensemble of 2D UNets
• Down-stream automatic model generation from segmentations for LV CFD simulations

Dependencies

• Segmentation
  • Tensorflow (V 1.14)
  • Python
  • SimpleITK¹
  • h5py (V 2.10.0)
• Model Generation
  • Python
  • VTK
  • SimVascular (Meshing)
  • SimpleElastix (Registration)¹

¹ Install SimpleElastix commit 8244e0001f4137514b0f545f1e846910b3dd7769. This will automatically install a proper version of SimpleITK. Generally speaking, SimpleITK version higher than 2.0 will not work due to syntax changes.

Segmentation Usage

The segmentation models can generate segmentations for LV blood pool, LV myocardium, LA, RA, RV blood pool, aorta and pulmonary artery.

Input Requirements

The preferred input format of the image volumes is .nii.gz or nii. VTK image volumes (.vti) are also accepted; however they should be reoriented to have an orientation matrix of identity. This is because the segmnetation method requires identity-oriented image volumes while the version of VTK within SimVascular does not include orientation matrix with VTI images. It is recommended to number the files starting from 0. The directory containing the input image data should be organized as follows:

image_dir
    |__ ct_test
        |__ patient_id (optional)
          	|__ image_volume0.nii.gz
          	|__ image_volume1.nii.gz
          	|__ image_volume2.nii.gz
          	|__ ...

Trained Models

We used the image and ground truth data provided by MMWHS to train our models. Our segmentation models were trainined simultaneously on CT and MR data and trained weights are here.

Prediction

To generate segmentations for 3D CT or MR image volumes:

python Segmentation/prediction.py \
    --pid patient_id \ # Patient ID.
    --image image_dir \ # the images should be saved in proper format in a folder named [modality]_test/patient_id within image_dir. 
    --output output_dir \
    --model weight_dir \
    --view 0 1 2 \ # Use models trained on axial (0), coronal (1) and/or sagittal (2) view[s].
    --modality ct \ # Image modality, ct or mr.
    --mode test

A shell script (run_seg.sh) is provided for ease of use. Patient ID is optional, and should supply None in the shell script if not used.

LV Modeling Usage

The model construction pipeline takes in the generated segmentation and output reconstructed LV surface meshes for CFD simulations. The pipeline consists of the following four steps: 1) Construct LV surface meshes from segmentation results; 2) Register the surface meshes to get consistent mesh topology; 3) Obtain volumetric mesh using SimVascular; 4) Interpolate the registered surface meshes to obtain sufficient temporal resolution.

1. Construct LV Surface Meshes with Tagged Boundary Faces

• Update run_svsurfaces.sh with correct file and folder names.
• Run the shell script to generate a LV surface mesh for each segmentation file in a folder.

  sv_python_dir=/usr/local/bin
  model_script=Modeling/main.py
  dir=./examples/ct_test_seg
  for file in ${dir}/*.nii.gz; do echo ${file} &&  ${sv_python_dir}/simvascular --python -- ${model_script} --input_dir ${dir} --output_dir ${output_dir} --seg_name ${file##*/} --edge_size 2.5; done
  

• Use --disable_SV to turn off SimVascular (no remeshing would be performed).

  for file in ${dir}/*.nii.gz; do echo ${file} &&  ${sv_python_dir}/simvascular --python -- ${model_script} --input_dir ${dir} --output_dir ${output_dir} --seg_name ${file##*/} --edge_size 2.5 --disable_SV; done
  

2. Construct Point Corresponded LV Meshes from 4D Images

Building point-corresponded LV meshes require segmentations from all time frames. One surface mesh will be created at one time frame and propagated to the others by registering the corresponding segmentations.

• Update run_surfregist.sh with correct file and folder names. Specify the time phase id to construct LV surface mesh.
• Run elastix_main.py through the shell script.

3. Volumetric Meshing using SimVascular

• Update run_volmesh.sh with correct file/folder names and mesh edge size.

• Run volume_mesh_main.py through the shell script.

4. Generate Mesh Motion File for svFSI

• Run run_simulation.sh to generate mesh motion file. Input: Surface mesh from segmented geometry with the same connectivity. Output: Displacement files for all the surfaces in this format.

Acknowledgement

This work was supported by the NSF, Award #1663747.