This repository provides the full code and configurations used for our submission to the TrackRAD2025 Challenge, focused on real-time tumor tracking in cine-MRI sequences. Our approach is based on SAM 2.1 (Segment Anything Model), adapted for the medical domain using prompt-based fine-tuning on the small labeled subset provided by the challenge.

We evaluated several foundation model variants and prompting strategies on the annotated validation subset of TrackRAD2025.

Fine-tuning the SAM 2.1 l model improved all metrics, confirming the value of adapting to annotator-specific mask styles.

We compared the performance of SAM 2.1 b+ with and without fine-tuning on unseen preliminary test data from TrackRAD2025:

👉 Fine-tuning on the annotated TrackRAD2025 data significantly improved segmentation and tracking performance.

Our final submission achieved the 6th best overall position on the TrackRAD2025 leaderboard, with top-tier performance in multiple metrics:

We achieved:

• 4th best Dice among all teams
• 3rd best relative dose (D98) accuracy

🔗 View the full TrackRAD2025 leaderboard

Automatically download the fine-tuned SAM model from Hugging Face:

python download.py

This will create the following file:

./SAM2.1_b+_finetune.pt

🔗 Model on Hugging Face

You can prepare the dataset manually or automatically.

Expected folder structure:

./Dataset/
├── A_001/
│   ├── IMAGE.mha
│   └── MASK.mha
├── A_002/
│   ├── IMAGE.mha
│   └── MASK.mha
└── ...

• IMAGE.mha is the input cine-MRI frame
• MASK.mha is the binary segmentation mask of the first annotated frame

Run the following command to automatically convert the TrackRAD2025 labeled dataset into the proper format:

python data.py

This will generate the ./Dataset/ folder with one subfolder per patient (A_001/, etc.), ready for training or inference.

Use the following command to run inference on the prepared dataset:

python model.py

Predictions will be saved as:

./Dataset/A_XXX/SAM21FINETUNE.mha

To evaluate the predictions against the ground truth masks:

konfai EVALUATION -y

• Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A. C., Lo, W.-Y., et al. (2023).
  Segment Anything. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 4015–4026.
  https://arxiv.org/abs/2304.02643

• Ma, J., He, Y., Li, F., Han, L., You, C., & Wang, B. (2024).
  Segment Anything in Medical Images. Nature Communications, 15, 654.
  https://www.nature.com/articles/s41467-024-44979-3

• Boussot, V., & Dillenseger, J.-L. (2025).
  KonfAI: A Modular and Fully Configurable Framework for Deep Learning in Medical Imaging. arXiv preprint arXiv:2508.09823.
  https://arxiv.org/abs/2508.09823