TeNCA — Temporal Neural Cellular Automata

Modeling contrast enhancement in breast MRI as a temporal process using Neural Cellular Automata

---

Overview

Dynamic contrast-enhanced (DCE) breast MRI is a sensitive screening modality, but it is slow, expensive, and requires intravenous injection of a gadolinium-based contrast agent. Synthetic contrast enhancement aims to remove the dependency on contrast agent injection by generating post-contrast sequences from a pre-contrast acquisition, cutting scan time and cost while avoiding contrast-agent exposure.

However, contrast enhancement is a physical process that unfolds over time, yet DCE acquisitions are temporally sparse and non-uniformly sampled, i.e., MRIs are only taken at specific timestamps, not as a dense video. Therefore, DCE modeling does not simply involve producing a single plausible post-contrast frame, but producing a temporally consistent trajectory.

TeNCA (Temporal Neural Cellular Automata) reframes this contrast modeling as an iterative cellular-automata update, where the number of NCA steps maps to elapsed physical time.

---

Repository structure

TeNCA/
├── config/
│   └── <model>.yaml         # experiment configuration (model / training / data)
├── data/                    # dataset root, see "Data"
├── src/
│   ├── dataset.py           # PyTorch Dataset
│   ├── metrics.py           # MSE / MAE / SSIM-MAE losses, masked SSIM, masked FID
│   ├── utils.py             # helper functions for logging
│   ├── nca/
│   │   ├── model.py         # NCA model
│   │   └── trainer.py       # NCA training loop
│   └── unet/
│       ├── unet.py          # U-Net baseline
│       └── trainer.py       # U-Net training loop
│
├── utils/                   # data preparation (download, patchify, train-test split, ...)
├── main.py                  # run models
├── requirements.txt         # Python dependencies
├── LICENSE                  # Apache-2.0
└── README.md

Installation

git clone https://github.com/LangDaniel/TeNCA.git
cd TeNCA

# recommended: isolated environment
conda create -n tenca python=3.12 -y
conda activate tenca

pip install -r requirements.txt

Usage

python main.py config/<model>.yml

Parameters can be adjusted in the config/<model>.yml file, with config/nca.yml and config/unet.yml reproducing the paper's setup.

main.py creates a timestamped run directory under output.output_dir, snapshots the config and source files for reproducibility, then builds and trains the selected model.

Citation

If you use this code or build on TeNCA, please cite:

@inproceedings{lang2025temporal,
  title={Temporal Neural Cellular Automata: Application to modeling of contrast enhancement in breast MRI},
  author={Lang, Daniel M and Osuala, Richard and Spieker, Veronika and Lekadir, Karim and Braren, Rickmer and Schnabel, Julia A},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={604--614},
  year={2025},
  organization={Springer}
}