SynthRAD2025 - Segmentation model weights

SynthRAD2025 Grand Challenge: generating synthetic CTs for radiotherapy

Medical imaging is crucial in modern radiotherapy, aiding diagnosis, treatment planning, and monitoring. The development of synthetic imaging techniques, particularly synthetic computed tomography (sCT), continues to attract interest in radiotherapy. The SynthRAD2025 dataset and the accompanying SynthRAD2025 Grand Challenge aim to stimulate advancements in synthetic CT generation algorithms by providing a platform for comprehensive evaluation and benchmarking of synthetic CT generation algorithms based on cone-beam CTs (CBCT) and magnetic resonance images (MRI).

Generated sCTs are evaluated based on image similarity, geometric fidelity, and dose accuracy

Goal of this repository:

Geometric fidelity is measured by comparing organ and tissue segmentations derived from sCT to the same segmentations derived from the sCT. These segmentations are automatically generated using TotalSegmentator[1], an nnUNet[2]-based total body segmentation model. This repository serves as an archive that hosts the exact model weights and code that was used during the challenge to generate the gold-standard segmentations, and the on-the-fly generated segmentations from the synthetic CT. 

How to evaluate sCTs used in the SynthRAD2025 challenge?

1. Download the evaluation code from https://github.com/SynthRAD2025/evaluation
2. Download both files in this repository and extract them in the evaluation folder. (note that the Totalsegmentator_weights_synthrad2025.tar.gz archive contains only a hidden folder called .totalsegmentator
3. Prepare the data
   1. In ground_truth, the file structure is separate directories for the ct, mask, and segmentation. Then ct/<patient_id>.mha is the ground-truth CT, mask/<patient_id>.mha is the corresponding mask that is 1 where there is anatomy, 0 otherwise, and segmentation/<patient_id>.mha contains the segmentation as generated by TotalSegmentator. This segmentation contains the class label for every voxel.
   2. In the input, there should be matching files <patient_id>.mha. These are the Synthetic CTs generated from the CBCT or MR images by your algorithm. These sCTs must be generated by you beforehand, evaluation of algorithms is NOT part of this evaluation container.
   3. With these files in place, run input/generate_pred_json.py. This command generated the JSON file to simulate a submission to Grand-Challenge. When CREATE_FOLDERS_AND_FILES=True, the script will also generate the file structure.
4. Build the docker image using docker build -t synthrad2025_evaluation_docker .
5. Run the docker image using 

   docker run -e NPROCS=1 -e DEBUG=0 -v /full/path/to/input:/input -v /full/path/to/groundtruth:/opt/ml/input/data/ground_truth -v /full/path/to/output:/output synthrad2025_eval_docker
   
   
   

License

The original TotalSegmentator and nnUNet code were distributed under the Apache 2.0 License. The (modified) source code files and model weights are redistributed here using the same license. 

[1] Wasserthal, J., Breit, H.-C., Meyer, M.T., Pradella, M., Hinck, D., Sauter, A.W., Heye, T., Boll, D., Cyriac, J., Yang, S., Bach, M., Segeroth, M., 2023. TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images. Radiology: Artificial Intelligence. https://doi.org/10.1148/ryai.230024

[2] Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J., & Maier-Hein, K. H. (2021). nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2), 203-211.