Our new work on longitudinal lesion tracking was early accepted at MICCAI 2026 (top 9%). The method also won 1st place in the MICCAI autoPET IV challenge.

LongiSeg 2.0 introduces Verified Tracking, a clinically safe paradigm for lesion follow-up. Instead of silently accepting fully automatic tracking results, the workflow separates lesion retrieval and delineation:

1. A registration-based method proposes the corresponding lesion location in the follow-up scan.
2. A clinician verifies or corrects the proposed point prompt.
3. LongiSeg segments the lesion using the verified prompt and the baseline lesion appearance as longitudinal context.

The model combines early prompt fusion, latent temporal difference weighting, and large-scale synthetic longitudinal pretraining. This allows LongiSeg 2.0 to use both the current scan and prior lesion appearance when segmenting lesions across timepoints. It sets a new state of the art in lesion tracking and even its automatic tracking outperforms the verified tracking results of competing methods.

We also release PanTrack, the first public longitudinal benchmark for pancreatic cancer lesion tracking. It contains 45 patients and 161 CT scans with temporally matched annotations of pancreatic tumors and selected liver metastases.

Paper:
Dataset:

LongiSeg is an extension of the popular nnU-Net framework, designed specifically for longitudinal medical image segmentation. By incorporating temporal information across multiple timepoints, LongiSeg enhances segmentation accuracy and consistency, making it a robust tool for analyzing medical imaging over time.

LongiSeg includes several methods for temporal feature merging, including the newly introduced Difference Weighting Block.    
For more details on the underlying nnU-Net framework, visit the nnU-Net repository.

Please cite the following works when using LongiSeg in your research:

@article{kirchhoff2026exploiting,
  title={Exploiting Longitudinal Context in Clinician-Verified Interactive Lesion Tracking},
  author={Kirchhoff, Yannick and Rokuss, Maximilian and Mertens, Daniel Philipp and F{\"u}ller, David and Hamm, Benjamin and Schreyer, Andreas and Ritter, Oliver and Maier-Hein, Klaus},
  journal={arXiv preprint arXiv:2605.23118},
  year={2026}
}
@inproceedings{rokuss2024longitudinal,
  title={Longitudinal segmentation of MS lesions via temporal Difference Weighting},
  author={Rokuss, Maximilian R and Kirchhoff, Yannick and Roy, Saikat and Kovacs, Balint and Ulrich, Constantin and Wald, Tassilo and Zenk, Maximilian and Denner, Stefan and Isensee, Fabian and Vollmuth, Philipp and Kleesiek, Jens and Maier-Hein, Klaus},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={64--74},
  year={2024},
  organization={Springer}
}
@article{isensee2021nnu,
  title={nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation},
  author={Isensee, Fabian and Jaeger, Paul F and Kohl, Simon AA and Petersen, Jens and Maier-Hein, Klaus H},
  journal={Nature methods},
  volume={18},
  number={2},
  pages={203--211},
  year={2021},
  publisher={Nature Publishing Group}
}

LongiSeg is easy to use and follows the basic principles of the nnU-Net framework.

In order to get started, we recommend to create a virtual environment with at least Python 3.10, e.g. using conda

conda create -n longiseg python=3.12

LongiSeg is not yet available via pip, therefore you need to clone the repository and install it locally. This will also allow for easy customization of the code.

conda activate longiseg
git clone https://github.com/MIC-DKFZ/LongiSeg.git
cd LongiSeg
pip install -e .

Finally, you need to set the paths for raw data, preprocessed data and results

export LongiSeg_raw="/path_to_data_dir/LongiSeg_raw"
export LongiSeg_preprocessed="/path_to_data_dir/LongiSeg_preprocessed"
export LongiSeg_results="/path_to_experiments_dir/LongiSeg_results"

If these are not set, LongiSeg will fall back to the respective nnU-Net paths, ensuring compatability with nnU-Net setups.

For more details on installation requirements and dataset structure, refer to the nnU-Net installation and the path setup documentation.

Detailed usage instructions for LongiSeg can be found in the documentation.

TL;DR:

1. Prepare your dataset, ensuring it includes a patientsTr.json file.
2. Run experiment planning and preprocessing: LongiSeg_plan_and_preprocess -d DATASET_ID.
3. Train your model: LongiSeg_train DATASET_NAME_OR_ID UNET_CONFIGURATION FOLD
4. Run inference on unseen data: LongiSeg_predict -i INPUT_FOLDER -o OUTPUT_FOLDER -path /path/to/patients.json -d DATASET_ID
5. Run evaluation on the predicted segmentation: LongiSeg_evaluate_folder GT_FOLDER PRED_FOLDER -djfile /path/to/dataset.json -pfile /path/to/plans.json -patfile /path/to/patients.json

LongiSeg is fully compatible with nnU-Net and can be installed alongside it in the same environment. This allows users to seamlessly reuse existing nnU-Net structures, datasets, and preprocessing pipelines without modification.

If you're working on lesion or tumor segmentation and tracking, make sure to also check out our LesionLocator framework, introduced at CVPR 2025:

🎯 LesionLocator is a powerful zero-shot framework for segmentation and longitudinal tracking of lesions in 3D whole-body imaging — no lesion-specific training required. It supports prompt-based segmentation (e.g., 3D point or box prompts) and autoregressive tracking across timepoints.

👉 GitHub: https://github.com/MIC-DKFZ/LesionLocator

LesionLocator and LongiSeg share a focus on longitudinal analysis, but with different strengths:

• LongiSeg is ideal for fully supervised, dataset-specific training with temporal modeling across timepoints.
• LesionLocator excels at zero-shot generalization and supports interactive & promptable workflows out of the box.

Use them together to benchmark traditional vs. zero-shot approaches — or combine insights from both for even better longitudinal segmentation performance.

LongiSeg is developed and maintained by the Division of Medical Image Computing at the German Cancer Research Center (DKFZ), supported by the Helmholtz Association under the joint research school “HIDSS4Health – Helmholtz Information and Data Science School for Health.

nnU-Net is developed and maintained by the Applied Computer Vision Lab (ACVL) of Helmholtz Imaging and the Division of Medical Image Computing at the German Cancer Research Center (DKFZ).