Team TUM BIC-MAC 2026 model
The 11 checkpoints used by Team TUM's final BIC-MAC 2026 submission. The model predicts a CT attenuation map from NAC-PET, an X-ray topogram, and Dixon MRI.
Download
hf download ziyuan-li/bic-mac-tum-e612 --local-dir model
The package contains E06, E08, five E13 fold models, E13-full75, E24, E40, and
E43. runtime-closure.json records the expected file sizes and SHA-256 hashes.
Inference
Code and setup instructions are available in
ziyuan-li/bic-mac-2026-methodology.
python scripts/verify_weights.py --model-dir model
PYTHONPATH=src python -m champion.predict \
--features-dir /path/to/features \
--model-dir model \
--output-ct output/ct.nii.gz \
--report output/runtime-report.json \
--device cuda:0
Results
| Metric | Hidden test result |
|---|---|
| CT mu-map MAE | 0.00591 |
| Whole-body SUV MAE | 0.0284 |
| Organ bias | 1.95% |
| Brain outlier score | 0.0249 |
| Mean rank | 6.25 |
Python 3.12, PyTorch 2.10.0, and an NVIDIA GPU are required. The model was developed for the BIC-MAC inputs and geometry and is not a medical device.