Systole-Conditioned Generative Cardiac Motion Using Deep Learning

📖 Published chapter: Systole-Conditioned Generative Cardiac Motion Using Deep Learning
Artificial Intelligence in Medicine and Surgery, Volume 5, IntechOpen, 2026.
Project Page · DOI

We present a method to generate synthetic 3D flow fields for myocardium deformation, conditioning on real cardiac CT frames using a Conditional Variational Autoencoder (CVAE). These synthetic data serve as ground truth annotations for training myocardium motion models. For a full explanation and results, visit the project page.

Installation

To install the necessary dependencies for the project, follow these steps:

1. Clone this repository:

   git clone https://github.com/shaharzuler/GenerativeCardiacMotion.git

2. Navigate to the project directory:

   cd GenerativeCardiacMotion

3. Install the required Python packages:

   pip install -r requirements.txt

Dataset Preparation

To prepare the dataset, use create_cvae_dataset.py. The data_source directory must contain:

• template_img_<number> files as 3D numpy arrays (shape: x, y, z) representing systole frames.
• flow_<number> files as numpy arrays (shape: 3, x, y, z) representing the flow fields.

Adjust the data_source and target_dir paths in the script and run it to generate the dataset, extracting features from the pyramid feature extractor.

Training

To train the model, use train.py:

• Adjust top_out_path and dataset_path to the correct directories.
• You can enable multiprocessing for running 4 parallel trainings, looping over the entire 44 samples, or specify a single sample using l1o_idx.
• The script outputs tensorboard logs, model weights, generated flow, and a 3D conditioned image to the output directory.
• Modify hyperparameters in config.json as needed.

Results and logs will be saved in the specified top_out_path.

Inference

To perform inference using the trained model, run infer.py:

• Adjust top_out_path, final_model (weights path), and dataset_path.
• Set num_generations to specify how many samples to generate.
• Select the type of latent space sampling to perform from the methods detailed below:

Latent Space Sampling Methods:

1. Random Sampling: Samples latent vectors from a standard normal distribution, creating diverse outputs.
2. Grid Search: Explores the latent space systematically by generating flow fields along a predefined grid of latent variables.
3. 2D Plane Sampling: Random sampling is restricted to a selected 2D plane within the latent space, allowing exploration of variations constrained to a specific slice of the latent distribution. The script will generate synthetic flow fields based on the latent space exploration, outputting the results in the specified path.

Publication

This work was published as a peer-reviewed book chapter:

Systole-Conditioned Generative Cardiac Motion Using Deep Learning
Shahar Zuler, Gal Lifshitz, Hadar Averbuch-Elor, and Dan Raviv
In Artificial Intelligence in Medicine and Surgery – An Exploration of Current Trends, Potential Opportunities, and Evolving Threats, Volume 5, IntechOpen, 2026.

DOI: 10.5772/intechopen.1015635

Citation

If you use this work or code, please cite:

@incollection{Zuler2026SystoleConditioned,
  author    = {Zuler, Shahar and Lifshitz, Gal and Averbuch-Elor, Hadar and Raviv, Dan},
  title     = {Systole-Conditioned Generative Cardiac Motion Using Deep Learning},
  booktitle = {Artificial Intelligence in Medicine and Surgery: An Exploration of Current Trends, Potential Opportunities, and Evolving Threats, Volume 5},
  editor    = {Stawicki, Stanislaw P. and Wojda, Thomas R.},
  publisher = {IntechOpen},
  address   = {London},
  year      = {2026},
  doi       = {10.5772/intechopen.1015635},
  url       = {https://doi.org/10.5772/intechopen.1015635}
}