Spinopelvic Seg Checkpoints

Model summary

The last_lumbar class merges what would otherwise be separate L5 and L6 labels. This eliminates the L5↔L6 channel-swap failure mode that affects fixed-class segmenters on sacralization-count cases (where the lumbar spine has 4 mobile segments instead of 5).

Files

Dataset803_SpineSurgCTFullMerged/
└── nnUNetTrainerWandB_500ep_LSTVOversample__nnUNetResEncUNetPlans_100G__3d_fullres/
    ├── plans.json
    ├── dataset.json
    ├── dataset_fingerprint.json
    ├── fold_0/
    │   ├── checkpoint_best.pth
    │   ├── checkpoint_final.pth
    │   ├── debug.json
    │   ├── progress.png
    │   └── training_log_*.txt
    ├── fold_1/   (same structure)
    ├── fold_2/   (same structure)
    ├── fold_3/   (same structure)
    └── fold_4/   (same structure)

checkpoint_best.pth is the recommended inference checkpoint for each fold (selected by validation EMA Dice). checkpoint_final.pth is the last training epoch.

Quickstart

1. Install dependencies

pip install nnunetv2 huggingface_hub
git clone https://github.com/anonymous-mlhc/spinopelvic-seg.git
cd spinopelvic-seg
pip install -r requirements.txt

2. Download checkpoints

hf download anonymous-mlhc/spinopelvic-seg-checkpoints \
    --repo-type=model \
    --local-dir nnunet/results/Dataset803_SpineSurgCTFullMerged

3. Set environment

export nnUNet_raw=$PWD/nnunet/raw
export nnUNet_preprocessed=$PWD/nnunet/preprocessed
export nnUNet_results=$PWD/nnunet/results
export PYTHONPATH=$PWD/tools:$PYTHONPATH   # makes the custom trainer importable

4. Run inference

Input files must follow nnU-Net's channel-suffix convention (CASE_0000.nii.gz):

nnUNetv2_predict \
    -i  /path/to/input_cts \
    -o  /path/to/predictions \
    -d  803 \
    -c  3d_fullres \
    -p  nnUNetResEncUNetPlans_100G \
    -tr nnUNetTrainerWandB_500ep_LSTVOversample \
    -f  0 1 2 3 4 \
    -chk checkpoint_best.pth

Drop -f 0 1 2 3 4 to -f 0 for single-fold inference (~1 point Dice hit vs the 5-fold ensemble, ~5× faster).

Label scheme

Intended use

• Research-grade segmentation of spine and pelvis on CT for ML benchmarking, anatomy analytics, and downstream LSTV-related studies.
• LSTV characterization — the merged-label scheme is designed for studies that need to characterize the lumbosacral junction without forcing a per-case decision on vertebral count.

Out-of-scope use

• Clinical diagnostic use. Not FDA-cleared, not CE-marked, no clinical validation.
• Patient-specific surgical planning without expert review.
• Pediatric CT. Training data is adult.
• Severe pathology (large tumors, hardware artifacts, multi-level fusion) is underrepresented in the training set.

Training

Trained on the CTSpinoPelvic1K dataset (companion HuggingFace dataset repo). 5-fold cross-validation at the patient-token level with stratification on match-type × LSTV subtype to prevent patient leakage and ensure each fold sees the LSTV-subtype distribution.

The custom trainer (nnUNetTrainerWandB_500ep_LSTVOversample in the companion code repo) adds: queue-based LSTV-case oversampling, CE reweighting on the merged-lumbar and sacrum classes, dedicated LSTV validation passes, and W&B logging with NaN-safe Dice aggregation and offline fallback.

See the companion repo for the full training pipeline (make preprocess, make train-array), the trainer source (tools/nnunet_wandb_variant.py), and ablation configurations.

Evaluation

tools/eval_full.py in the companion repo computes per-case Dice, junction-DSC over a 40 mm L5/S1 window, voxel confusion blocks for the L4 ↔ last_lumbar ↔ sacrum boundary classes, and last_lumbar specificity on sacralization-count cases.

License

Apache-2.0. See the companion repo.