Longitudinal4DMed: Models and Tools for Longitudinal and Spatio Temporal Medical Imaging

This repository is the official implementation of CRONOS (ICLR) and the continuations Temporal Flow Matching (TFM), a spatio-temporal and generative framework for longitudinal medical imaging. The repository also hosts LAUGEN, a method for generating longitudinal sequences from single images (Syndata4CV @ CVPR 2025).

Features

• Flow Matching and other methods for sequence-to-image forecasting.
• Discrete variant (grid-based, e.g. regular follow-up times).
• Continuous time reconstructions.
• Supports 3D+T or 4D sequences (e.g. MRI volumes, CT or US).
• Simple, dependency-light PyTorch code.
• Longitudinal augmentations and data generation.
• 5 different longitudinal and spatio-temporal medical imaging datasets (from preprocessing to dataloader)

Status

Actively maintained. Recently added:

• CRONOSFlex: see src/method/cronos_flex.py; more flexible handling of channels, supports the loading of pretrained architectures more generally
• Latent FM / CRONOS: for if you do not want to use voxel space
• /laugen for longitudinal augmentation and data generation
• eval.py now supports segmentation and segmentation masked metrics.
• minor additions: EMA, tests...

Installation

Clone this repository and install the required packages:

git clone https://github.com/MIC-DKFZ/Longitudinal4DMed.git
cd Longitudinal4DMed
pip install -e .

To launch TensorBoard during or after training:

tensorboard --logdir checkpoints/logs

Training

Each dataset has a ready-made config in configs/:

# ACDC (cardiac MRI)
python src/train.py --config configs/acdc.yaml

# ISLES 2024 (stroke CTP)
python src/train.py --config configs/isles.yaml

# Lumiere (glioma MRI)
python src/train.py --config configs/lumiere.yaml

CLI flags override any value from the config file, e.g. to run a quick debug pass:

python src/train.py --config configs/acdc.yaml --debug

For a quick install check without real data:

# see src/examples/train_dummy.ipynb
python src/train.py --dummy --device cpu --debug

Contact

For further information, or if you want to reach out to us, visit our webpage.

Citation

If you find this work useful for your research, please consider citing:

@misc{disch2025temporalflowmatchinglearning,
      title={Temporal Flow Matching for Learning Spatio-Temporal Trajectories in 4D Longitudinal Medical Imaging}, 
      author={Nico Albert Disch and Yannick Kirchhoff and Robin Peretzke and Maximilian Rokuss and Saikat Roy and Constantin Ulrich and David Zimmerer and Klaus Maier-Hein},
      year={2025},
      eprint={2508.21580},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.21580}, 
}
@misc{disch2025cronoscontinuoustimereconstruction,
      title={CRONOS: Continuous Time Reconstruction for 4D Medical Longitudinal Series}, 
      author={Nico Albert Disch and Saikat Roy and Constantin Ulrich and Yannick Kirchhoff and Maximilian Rokuss and Robin Peretzke and David Zimmerer and Klaus Maier-Hein},
      year={2025},
      eprint={2512.16577},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.16577}, 
}