TrackRAD2025's organizers welcome you on the challenge repository. The challenge has now been concluded and the challenge report has been published in Medical Image Analysis and can be found here. Other researchers are welcome and invited to compare their algorithms to the challenge results and to use the provided datasets for further research.
The here to go to the challenge website and participate.
The report on the challenge can be found here: https://doi.org/10.1016/j.media.2026.104134
This repository contains:
- A baseline algorithm (
baseline-algorithm/) to the TrackRAD challenge as a starting point. - An example datasets (
dataset/), following the format of the labeled TrackRAD dataset as well as a recommended folder structure for the TrackRAD datasets. - The evaluation code (
evaluation/) used for evaluating submissions, including a compact and fast custom implementation of the metrics. - A testing-script (
test-algorithm.sh) allowing for a local evaluation and testing. - The analysis scripts (
analysis/) and raw challenge results - The contents of the TrackRAD2025 website on grand-challenge.org.
- A number of miscellaneous scripts (
scripts/) including those to convert proprietary cine MRI formats from both vendors into metadata files and images/assets (images/).
The winning submission based on MedSAM2 can be found here.
The use of magnetic resonance imaging (MRI) to visualize and characterize motion is becoming increasingly important in the treatment of cancer patients, especially in radiotherapy. For tumors affected by it, motion management is crucial to ensure a high radiation dose to the tumor while sparing neighboring organs. The recent development of MRI-guided radiotherapy with hybrid MRI-linear accelerator (linac) systems [Keall et al., 2022], called MRI-linacs, offers the possibility to adapt to changes in tumor position during treatment. 2D cine-MRI (a time-resolved sequence of 2D images continuously acquired at the same location) allows real-time tumor motion visualization and closely following the tumor with the radiation beam, but requires automated tumor segmentation/tracking. This needs to be done in real-time, with high accuracy and robustness. Currently, clinically available solutions rely on conventional deformable image registration (DIR) or on template matching. These methods struggle with large non-rigid motion, limiting application to beam gating.
The fast inference of artificial intelligence (AI) methods, obtained by shifting computation time to the training phase, is promising for this task [Lombardo et al., 2024].
TrackRAD2025 aims to advance MRI-guided radiotherapy by providing cine-MRI data from multiple institutions to evaluate real-time tumor tracking methods on a unified platform.
- Challenge website open: 15/12/2024
- Release training cases: 15/03/2025
- Start of challenge - Training phase: 15/03/2025 - 15/08/2025
- Presentation at ESTRO 2025: 05/2025
- Preliminary testing phase (10 submissions per team): 01/06/2025 - 15/08/2025
- Advertisement of the challenge at ESTRO 2025: 04/05/2025
- Testing phase (2 submissions per team): 16/07/2025 - 15/08/2025
- Deadline for algorithm information form and LNCS format algorithm description: 01/09/2025
- Announcements and invitation to present: 09/09/2025
- Workshop at MICCAI 2025: 09/2025
- Post-challenge phase: submission is still possible until grand challenge credits are exhausted
- Presentation of the results at ESTRO 2026: 18/05/2026
- Publication of the challenge report: 23/05/2026
If you use this code in or for a scientific publication, please cite our paper: http://dx.doi.org/10.1088/2057-1976/ae71d8
@article{Blcker2026,
title = {MRIgRT real-time target tracking: TrackRAD2025 challenge report},
volume = {112},
ISSN = {1361-8415},
url = {http://dx.doi.org/10.1016/j.media.2026.104134},
DOI = {10.1016/j.media.2026.104134},
journal = {Medical Image Analysis},
publisher = {Elsevier BV},
author = {Bl\"{o}cker, Tom Julius and G\"{o}rts, Pia A.W. and Wang, Yiling and Lombardo, Elia and Thummerer, Adrian and Fan, Yu and Zhao, Yue and Papadopoulou, Christianna Iris and Hurkmans, Coen and Tijssen, Rob H.N. and Cusumano, Davide and Intven, Martijn PW and Borman, Pim and Riboldi, Marco and Dudáš, Denis and Byrne, Hilary L. and Placidi, Lorenzo and Fusella, Marco and Jameson, Michael and Palacios, Miguel A. and Cobussen, Paul and Finazzi, Tobias and Tetar, Shyama U. and Haasbeek, Cornelis J.A. and Keall, Paul and Maspero, Matteo and Kurz, Christopher and Tarifa, Amparo Soeli Betancourt and He, Kailin and Zhu, Shengqian and Song, Ying and Li, Guangjun and Hu, Junjie and Knispel, Felix and Gatidis, Sergios and Chu, Hung and Guo, Jiapan and Nielsen, Maximilian and Sentker, Thilo and Boussot, Valentin and Hémon, Cédric and Ni, Jing and Georgas, Konstantinos and Vagenas, Theodoros P. and Matsopoulos, George K. and Landry, Guillaume},
year = {2026},
month = July,
pages = {104134}
}
TrackRAD2025 provides a public multi-institutional datasets of both unlabeled and labeled 2D cine-MRI frame sequences for development and testing. Six international centers (3 Dutch, 1 German, 1 Australian, and 1 Chinese) provided data from 0.35 T and 1.5 T MRI-linacs. A further (private) labeled dataset was collected for testing and evaluation.
The dataset is documented in the dataset paper in Medical Physics: https://doi.org/10.1002/mp.17964
@misc{wang2025trackrad2025challengedatasetrealtime,
title={TrackRAD2025 challenge dataset: Real-time tumor tracking for MRI-guided radiotherapy},
author={Yiling Wang and Elia Lombardo and Adrian Thummerer and Tom Blöcker and Yu Fan and Yue Zhao and Christianna Iris Papadopoulou and Coen Hurkmans and Rob H. N. Tijssen and Pia A. W. Görts and Shyama U. Tetar and Davide Cusumano and Martijn P. W. Intven and Pim Borman and Marco Riboldi and Denis Dudáš and Hilary Byrne and Lorenzo Placidi and Marco Fusella and Michael Jameson and Miguel Palacios and Paul Cobussen and Tobias Finazzi and Cornelis J. A. Haasbeek and Paul Keall and Christopher Kurz and Guillaume Landry and Matteo Maspero},
year={2025},
eprint={2503.19119},
archivePrefix={arXiv},
primaryClass={physics.med-ph},
url={https://arxiv.org/abs/2503.19119},
}
The public datasets provided for the TrackRAD2025 challenge can be found on Hugging Face:
@misc {lmu_adaptive_radiation_therapy_lab_2025,
author = { {LMU Adaptive Radiation Therapy Lab} },
title = { TrackRAD2025 (Revision 505ec64) },
year = 2025,
url = { https://huggingface.co/datasets/LMUK-RADONC-PHYS-RES/TrackRAD2025 },
doi = { 10.57967/hf/4539 },
publisher = { Hugging Face }
}
- The labeled and unlabelled training datasets were already available during the challenge.
- The labeled pre-testing and testing datasets were not available during the challenge and only made public after the challenge. Some cases of the testing dataset were not made public due to privacy restrictions.
To download the datasets and place them in the expected locations, you can run the following commands:
from huggingface_hub import snapshot_download
# Labeled dataset (about 3 GiB) for unsupervised learning or evaluation
snapshot_download(repo_id="LMUK-RADONC-PHYS-RES/TrackRAD2025", repo_type="dataset", local_dir="./dataset/", allow_patterns="trackrad2025_labeled_training_data/*")
# the cases are stored in ./dataset/trackrad2025_labeled_training_data/
#Unlabeled dataset (attention, very large, about 200 GiB) for unsupervised learning:
snapshot_download(repo_id="LMUK-RADONC-PHYS-RES/TrackRAD2025", repo_type="dataset", local_dir="./dataset/", allow_patterns="trackrad2025_unlabeled_training_data/*")
# the cases are stored in ./dataset/trackrad2025_unlabeled_training_data/
# Labeled testing dataset (about 3 GiB):
snapshot_download(repo_id="LMUK-RADONC-PHYS-RES/TrackRAD2025", repo_type="dataset", local_dir="./dataset/", allow_patterns="trackrad2025_labeled_testing_data/*")
# the cases are stored in ./dataset/trackrad2025_labeled_testing_data/
# Labeled pre-testing/preliminary testing phase dataset (< 1 GiB):
snapshot_download(repo_id="LMUK-RADONC-PHYS-RES/TrackRAD2025", repo_type="dataset", local_dir="./dataset/", allow_patterns="trackrad2025_labeled_pre-testing_data/*")
# the cases are stored in ./dataset/trackrad2025_labeled_pre-testing_data/We recommend placing the dataset in the dataset/ subfolder of this repository, following the following structure:
dataset/
├── example # a single example case for technical testing
│ └── Z_001
│ ├── b-field-strength.json
│ ├── frame-rate.json
│ ├── images
│ │ └── Z_001_frames.mha
│ ├── scanned-region.json
│ └── targets
│ ├── Z_001_first_label.mha
│ └── Z_001_labels.mha
├── labeled # labeled data for supervised training
├── unlabeled # unlabeled data for unsupervised training
├── preliminary # cases used for the first phase of the challenge
└── testing # cases used for the final phase of the challengeFollow these steps (also available here and here) to set up your development environment and prepare your submission.
Begin by cloning the official TrackRAD2025 repository:
git clone git@github.com:LMUK-RADONC-PHYS-RES/trackrad2025.gitThis repository includes:
- Challenge information
- Code used for evaluation and local testing
- A baseline algorithm to help you get started
Download the dataset and place it in the dataset/ folder. Use the labeled dataset for supervised training and model evaluation. For unsupervised training you may want to use the unlabeled dataset, for supervised training and evaluation of your model the labeled dataset is sufficient.
Tip: A synthetic sequence (formatted like the labeled data) is also provided in the repository to help you get started quickly, even while still downloading the dataset.
Before modifying anything, verify that the baseline and evaluation are working correctly:
sh test-algorithm.sh
This script will compile, execute, and test the baseline algorithm with the synthetic example scan.
This script assumes a unix environment with docker installed correctly, for example on macOS or Linux. If you are using Windows, please look into and make use of the Windows Subsystem for Linux (WSL). Or port the script to PowerShell and create a PR.
Familiarize yourself with the parameters at the top of the script (especially ALGORITHM_DIR and DATASET_DIR). You will need to update these paths when evaluating your own algorithm.
Open baseline-algorithm/model.py in your code editor. You will see the following function:
def run_algorithm(frames: np.ndarray,
target: np.ndarray,
frame_rate: float,
magnetic_field_strength: float,
scanned_region: str) -> np.ndarray:
"""
Implement your algorithm here.
Args:
- frames (numpy.ndarray): A 3D numpy array of shape (W, H, T) containing the MRI linac series.
- target (numpy.ndarray): A 2D numpy array of shape (W, H, 1) containing the MRI linac target.
- frame_rate (float): The frame rate of the MRI linac series.
- magnetic_field_strength (float): The magnetic field strength of the MRI linac series.
- scanned_region (str): The scanned region of the MRI linac series.
"""
# frames.shape == (W, H, T)
# target.shape == (W, H, 1)
# For the example we want to repeat the initial segmentation for every frame
repeated_target = np.repeat(target, frames.shape[2], axis=-1)
# repeated_target.shape == (W, H, T)
return repeated_targetCopy the baseline algorithm folder to a new folder to work in.
cp -R baseline-algorithm your-algorithmNavigate into your new folder and initialise it as a separate git repository:
cd your-algorithm
git initModify the test-algorithm.sh script in the main repository to point to your new folder (update the ALGORITHM_DIR path).
ALGORITHM_DIR="./your-algorithm"Tip: By this time your download of the labeled dataset could be done and you could also update the dataset path, to use the labeled dataset. This is recommended if you want to test more than just the technical functionality of your algorithm.
# This might also be a good time to wherever you stored the labeled cases.
DATASET_DIR="./dataset/labeled"
# When using the command provided above use
# DATASET_DIR="./dataset/trackrad2025_labeled_training_dataEdit the model.py file in your working copy to implement your algorithm. Make sure to keep the function signature of the run_algorithm function intact. This way you ensure that your submission is functional.
def run_algorithm(frames: np.ndarray, # frames.shape == (W, H, T)
target: np.ndarray, # target.shape == (W, H, 1)
frame_rate: float,
magnetic_field_strength: float,
scanned_region: str) -> np.ndarray:
# Example: Repeat the initial segmentation
repeated_target = np.repeat(target, frames.shape[2], axis=-1)
return repeated_target # repeated_target.shape == (W, H, T)Note: When adding files or folders to it make sure to add them to the Dockerfile. This is required for the submission on grand-challenge, where you provide the platform access to github repository you created earlier. Alternatively, but not recommended, you can upload a container image, e.g. created using the save.sh script provided with the baseline algorithm. However, we still require read access to your model source code for prize eligibility.
Have fun developing and contributing the the future of cine-MRI target tracking!
P.S.: When you are ready to test your submission on grand-challenge or submit your final version, go to the Submission Instructions.
Submitted algorithms will be evaluated based on their ability to reproduce ground truth segmentation labels on the test set, using the following metrics:
- Dice similarity coefficient
- 95th percentile of the surface distance distribution
- Average surface distance
- Euclidian 2D center-of-mass distance
- Dosimetric accuracy under simulated MLC tracking
- Speed of inference/runtime
A detailed description of these metrics and the ranking process can be found here or here.
The metrics are implemented here or in a single file here.
While the metrics will remain as described on the challenge website, the organizers retain the option to change the actual evaluation code to fix bugs or exploits.
This license only applies to the Software contained in this repository. All other materials made available in the context of the TrackRAD2025 competition may be licensed under different terms.
Copyright 2025 The TrackRAD2025 organizers
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.