spinopelvic-seg-checkpoints

spinopelvic-seg-checkpoints

The five-fold nnU-Net v2 ensemble that pseudolabelled the pelvis on the CTSpinoPelvic1K records that lacked one. Spine + pelvis CT segmentation with an LSTV-aware merged-label scheme: L5 and L6 collapse into one last_lumbar class so a lumbosacral transitional vertebra cannot force a wrong count.

Using the weights

The layout is exactly what nnUNetv2_predict expects under nnUNet_results:

nnUNet_results/
  Dataset803_SpineSurgCTFullMerged/
    nnUNetTrainerWandB_500ep_LSTVOversample__nnUNetResEncUNetPlans_100G__3d_fullres/
      plans.json  dataset.json  dataset_fingerprint.json
      fold_0/checkpoint_final.pth ... fold_4/checkpoint_final.pth

pip install nnunetv2
huggingface-cli download OpenSpineConsortium/spinopelvic-seg-checkpoints \
    --local-dir $nnUNet_results
git clone https://github.com/OpenSpineConsortium/spinopelvic-seg   # the trainer class
export PYTHONPATH=$PWD/spinopelvic-seg/tools:$PYTHONPATH
nnUNetv2_predict -i CT_DIR -o OUT_DIR -d 803 -c 3d_fullres \
    -tr nnUNetTrainerWandB_500ep_LSTVOversample -p nnUNetResEncUNetPlans_100G -f 0 1 2 3 4

The trainer class must be importable. nnU-Net rebuilds the network from the trainer named in the checkpoint; nnUNetTrainerWandB_500ep_LSTVOversample is defined in tools/nnunet_wandb_variant.py of the companion repository, not in nnU-Net itself. Without it on the path, nnUNetv2_predict stops with "trainer not found". Inference itself needs nothing from Weights & Biases.

Out-of-fold use: to complete a CTSpinoPelvic1K record with a model that never trained on it, predict with the single fold whose validation split holds that record (-f <fold>), as the release did.

Provenance

Mirrored 2026-09-04 from anonymous-mlhc/spinopelvic-seg-checkpoints (identical files, 11.37 GB). Checkpoints were loaded with torch.load fold by fold after the mirror; all five report epoch 500 and 528.1 M parameters.

• Dataset: https://huggingface.co/datasets/OpenSpineConsortium/CTSpinoPelvic1K (archive of record: https://doi.org/10.5281/zenodo.22139642)
• Code: https://github.com/OpenSpineConsortium/spinopelvic-seg
• Build pipeline and paper: https://github.com/OpenSpineConsortium/CTSpinoPelvic1K

Released under Apache-2.0 for research use only; not a medical device.