# Third-party provenance and licenses

This repository distributes BodyCompositionCT-ResEncM trained weights,
documentation, metadata, and nnU-Net training and evaluation records. Source
CT images, source annotations, upstream checkpoints, and third-party code are
not redistributed.

| Material or software | Use | Upstream terms |
| --- | --- | --- |
| SAROS segmentations and split metadata | Sparse reviewed body-region labels used during target construction | CC BY 4.0; [dataset](https://doi.org/10.25737/SZ96-ZG60), [paper](https://doi.org/10.1038/s41597-024-03337-6) |
| SAROS source CT images | Training images | Collection-specific TCIA terms, including TCIA Restricted, CC BY 3.0, CC BY-NC 3.0, and CC BY 4.0; not redistributed. [SAROS/TCIA page](https://wiki.cancerimagingarchive.net/pages/viewpage.action?pageId=157287899) |
| BOA Task542 body-region weights | Dense body-region predictions used during target construction | MIT; [weights](https://doi.org/10.5281/zenodo.7918824), [paper](https://doi.org/10.1097/RLI.0000000000001040) |
| TotalSegmentator v2 CT dataset | Training CT images and anatomical labels | CC BY 4.0; [dataset](https://doi.org/10.5281/zenodo.10047292) |
| TotalSegmentator `total` and `body` tasks | Anatomical corrections and exclusions | Apache-2.0; [repository](https://github.com/wasserth/TotalSegmentator), [paper](https://doi.org/10.1148/ryai.230024) |
| nnU-Net v2 and residual-encoder configuration | Training framework and model architecture | Apache-2.0; [repository](https://github.com/MIC-DKFZ/nnUNet), [nnU-Net paper](https://doi.org/10.1038/s41592-020-01008-z), [residual-encoder paper](https://doi.org/10.48550/arXiv.2404.09556) |

CC BY 4.0 applies to the BodyCompositionCT trained weights, documentation,
metadata, and included training/evaluation records. Source data, source labels,
upstream checkpoints, and third-party code are not included and remain under
their original terms. The listed upstream terms do not impose a
non-commercial or share-alike condition on this model package.

The target construction used the open TotalSegmentator `total` and `body`
tasks, not the separately licensed `tissue_types` or `vertebrae_body` tasks.