INR for Cardiovascular Anatomies and Hemodynamic Fields

Code and data to reproduce the results of the paper: Jubilee Lee, Daniele E. Schiavazzi, On the accuracy of implicit neural representations for cardiovascular anatomies and hemodynamic fields.

Installation

Clone the repository from GitHub:

git clone git@github.com:desResLab/nrf.git

The code can be installed using different version of multiresolution hash encoding (MHE). To install a pytroch-compatible version of MHE that can run on CPUs, type:

pip install -e .

To use MHE from the NVIDIA PhysicsNeMo libray run:

pip install -e . [physicsnemo]

Finally, to use MHE from the tiny-CUDA-nn library, run:

pip install -e . [tcnn]

Features

The following type of implicit neural representations are currently supported:

• Multiplicative filter networks - Link.
• Multiresolution hash encoding (physicsnemo, tcnn and pure pytorch implementations) - Link.
• MLP with various types of frequency-based positional encoding - Link.
• SIREN networks - Link.
• Modulated SIREN networks - Link.

Dataset

To reproduce the results in the paper, a Datasets can be downloaded from box. Create the ./data subfolder in the main project directory and download all the subfolders from box in this folder. The following datasets are currently available:

• Two dimensional (XT or XY) datasets:
  • ./data/01_xt_pipe/xt_pipe.npz - Two-dimensional dataset (one-dimensional space and time) of centerline velocities and pressure in a cylindrical pipe.
  • ./data/02_xy_zebra/xt_zebra.npz - Two-dimensional zebra image dataset.
• Three-dimensional (XYZ) datasets:
  • ./data/03_xyz_pipe/xyz_pipe.npz - Steady state three-dimensional hemodynamics in a cylindrical pipe.
  • ./data/04_xyz_aorta/xyz_aorta.npz - Steady state three-dimensional hemodynamics of a patient-specific aortic anatomy.
• Four-dimensional (XYZT) datasets:
  • ./data/05_xyzt_pipe/xyzt_pipe.npz - Pulsatile four-dimensional hemodynamics in a pipe.
  • ./data/05_xyzt_pipe/xyzt_pipe_grid.npz - Regular validation grid for pulsatile four-dimensional hemodynamics in a pipe.
  • ./data/06_xyzt_aorta/xyzt_aorta.npz - Pulsatile four-dimensional hemodynamics in a patient-specific aortic anatomy.
• Signed distance function (SDF) datasets:
  • ./data/10_xyz_sdf/sphere_XXX_YYY.sdf.npz - Ideal spherical geometry.
  • ./data/10_xyz_sdf/aorta_XXX_YYY.sdf.npz - Patient-specific aortic anatomy.
  • ./data/10_xyz_sdf/pulm_XXX_YYY.sdf.npz - Patient-specific pulmonary anatomy. Here XXX refers to the type of dataset (see paper) and can be either M_S_S, S_M_S, S_S_M. YYY refers to the size of the dataset and can be either large, or small.
• Aortic Zoo dataset:
  • ./data/11_xyz_sdf_zoo/ - Contains various *.sdf.npz files with multiple patient-specific aortic anatomies from the vascular model repository.

Available functions

We provide python functions and scripts for the following tasks (use --help sfor a list of avaialble arguments):

• python3 -m nrf.train - Train a neural field from a dataset.

Examle

python3 -m nrf.train --iterations 100 \
                     --exptype gabor \
                     --batch 2048 \
                     --dsfile './data/04_xyz_aorta/xyz_aorta.npz' \
                     --dstype xyz \
                     --param v_x v_y v_z p \
                     --output "checkpoint.pt"

• python3 -m nrf.validation - Evaluate the trained NRF on the training data.

Examle

python3 -m nrf.validation --chkfile "checkpoint.pt" \
                          --validfile "./data/06_xyzt_aorta/xyzt_aorta.npz" \
                          --eval_mode output \
                          --output "output_file"

• python3 -m nrf.validation_to_mesh - Convert validation data to VTK.

Examle

python3 -m nrf.validation_to_mesh --vtkfile "mesh_file.vtu" \
                                  --predfile "prediction_file.npy" \
                                  --output_prefix "output_file"

• python3 -m nrf.tria_to_sdf - Create a training dataset for deep SDF from a triangulated surface model.

Examle

python3 -m nrf.tria_to_sdf --file "sphere.vtp" \
                           --total 100000 8000 8000 \
                           --output "sphere_M_S_S_small"

• python3 -m nrf.eval_zerols_dist - Evaluate distance between a zero level set contour and an existing triangulated surface model.

Examle

python3 nrf/eval_zerols_dist.py --sdf "vtk_image.vti" \
                                --ref "surface_anatomy.vtp" \
                                --output "output.vtp" \
                                --distfile "distance_file"

• python3 -m nrf.plot_loss - Plot loss profile from model checkpoint.

Examle

python3 -m nrf.plot_loss --input "checkpoint.pt"

Example training and validation scripts

Examples training and validation scripts can be found in the run_scripts folder.

Reference

Our work can be cited using the bibtex item below:

@article{lee2025accuracy,
  title={On the accuracy of implicit neural representations for cardiovascular anatomies and hemodynamic fields},
  author={Lee, Jubilee and Schiavazzi, Daniele E},
  journal={arXiv preprint arXiv:2510.20970},
  year={2025}
}