ribline-3d-ct-centerline: Pretrained Semantic Context Forests, 3D U-Net and Statistical Rib Cage Atlas

Official checkpoints for ribline (v0.3.0), a hybrid machine-learning / deformable-topology pipeline for automatic 3D rib centerline extraction and anatomical numbering (R01..R12, L01..L12) in thoracic CT, described in Hybrid Context Forests and Bilateral Viterbi Matching for 3D Rib Centerline Tracing in CT (PDF) and part of the ribline Hugging Face Collection:

Hugging Face Repository	Model Variant	12-Scan LCD (mm)	12-Scan Acc (%)	5-Fold CV LCD (mm)	5-Fold CV Acc (%)
wq2012/ribline-3d-ct-centerline	Hybrid SCF + 3D U-Net + Stage 3 (Full Suite)	2.05 +/- 0.48	97.2 %	3.64 +/- 3.90	91.8 %
wq2012/ribline-anatomical-scf	Anatomically Anchored SCF + Stage 3 (CPU-only)	2.88 +/- 1.32	95.1 %	5.28 +/- 4.62	86.9 %
wq2012/ribline-3d-unet	Lightweight 3D U-Net + Stage 3	3.40 +/- 1.99	92.0 %	7.44 +/- 8.93	78.2 %
wq2012/ribline-scf-pass2-autocontext	2-Pass Auto-Context SCF Baseline + Stage 3	9.36 +/- 10.57	71.0 %	--	--

All models were trained on the 47 annotated training scans (1,116 ribs) of the Siemens Corporate Research (SCR) 2 mm thoracic CT benchmark (split.json) and evaluated on the 12 evaluation scans (286 ribs after the documented ground-truth fix). Please read the Protocol caveat and Limitations below before quoting any number.

What changed in this revision. The previously shipped rib_unet_3d.pth was a mid-run snapshot (md5 de4b158b...) that did not produce the reported probability maps; it has been replaced by the epoch-20 checkpoint (md5 9e456ba9ff81c402a58436c666f152ef) that reproduces them exactly. The evaluation metrics were corrected (symmetric 10 mm criterion, 50 mm penalty for missed ribs, nan for undefined groups, precision added) and every number below was recomputed with the released code. Architecture and training descriptions were rewritten to match the code (plain U-Net, Dice + BCE loss, raw-intensity z-score input, no augmentation).

---

Included files

File	Component	Description (as implemented)
rib_forest_pass1.joblib + feature_config_pass1.json	Pass-1 forest	120 trees, 228 features: raw intensity, 27 multi-scale Gaussian derivative features (sigma 1 / 2 / 4 mm: smoothed, gradient magnitude, 6 second derivatives, Laplacian), 200 RSID pairs with offsets in [-20, 20] voxels (+/-40 mm).
rib_forest_pass2.joblib + feature_config_img.json + feature_config_ctx.json	Pass-2 auto-context forest	100 trees, 306 features = 228 Pass-1 features + 28 multi-scale features of the Pass-1 probability map + 50 random-shift probability differences in [-40, 40] voxels (+/-80 mm). Trained on Pass-1 maps of an earlier Pass-1 forest; ablation baseline only.
rib_forest_anatomical.joblib + feature_config_anatomical.json	Anatomically anchored SCF	150 trees, max depth 24, 286 features: HU-standardized 28 differential features, a diagonal-curvature tubularity proxy per scale (no Hessian eigen-decomposition / Frangi filter), canonical coordinates `
rib_unet_3d.pth	3D U-Net	Plain 3-level U-Net (16 -> 32 -> 64, Conv-GroupNorm-ReLU x2 per level, no residual blocks), 339,889 parameters. Input: per-volume z-score of raw stored intensities. 64^3 patches, 0.5 soft-Dice + 0.5 BCE, Adam 1e-3, batch 2, 8 patches / scan / epoch, 20 epochs, no augmentation (train_config_unet.json).
rib_atlas.joblib	24-rib statistical atlas	Mean canonical curves (80 points per rib), 5-D descriptor mean / SD, inter-rib spacing statistics, 15,000-point slant-correction cloud. python -m ribline.scripts.build_atlas on the 47 training scans reproduces it exactly.
benchmark_val12_fixed_gt.json, benchmark_val12_original_gt.json, benchmark_gt_original_excl_rib007_R07.json, benchmark_noSko.json, benchmark_train47.json, benchmark_cv5.json	Evaluation reports	Per-rib and per-subject metrics of all ablation rows on the 12 eval scans (including GT-sensitivity variants), the 47 training scans, and the 59-scan 5-fold CV with bootstrap CIs and paired tests.
polar_stats_all59.json, gt_fix_log.json, train_config_unet.json	Statistics / provenance	Frame statistics on all 59 scans; log of the single ground-truth fix; U-Net training recipe.

---

Evaluation protocol

ribline.model.metrics.evaluate_subject_ribs (defaults): labeled Chamfer distance (LCD) averaged over all GT ribs with missing labels counting 50 mm; labeled clDice at 3 / 4 mm; a GT rib is correctly labeled iff the same-label prediction has mean prediction-to-GT and GT-to-prediction distance <= 10 mm and the GT rib is the closest one to it; recall uses the same symmetric criterion over any predicted curve; precision counts predicted curves lying within 10 mm of some GT rib. Subject-level means are given with SD and 95 % bootstrap CI (10,000 resamples), pooled rates with Wilson intervals.

Protocol caveat. The constants of the Stage-3 labeling / refinement code were tuned while inspecting metrics on these 12 scans; the three forests, the U-Net and the atlas were fitted on the 47 training scans only. The numbers below are therefore in-sample for the Stage-3 constants. The split is not random (alphabetical tail of the file list).

Results on the 12 evaluation scans (286 GT ribs, RIB_007 annotation fixed)

Source: benchmark_val12_fixed_gt.json, produced by python -m ribline.scripts.run_ablation with the released code.

Method	LCD (mm)	95 % CI	UCD (mm)	lab. clDice 3 / 4 mm	Acc all	Acc 1st	Acc 2-11	Acc 12th	Recall	Precision (FP)	NLD
RF Pass-1 + Z-sort	14.10 +/- 11.10	[7.96, 20.46]	3.31	0.548 / 0.606	64.3 %	79.2 %	62.9 %	63.6 %	91.3 %	95.5 % (13)	0.264
SCF Pass-2 + Z-sort	21.30 +/- 13.33	[13.79, 28.67]	4.16	0.416 / 0.452	43.0 %	45.8 %	44.2 %	27.3 %	76.2 %	100.0 % (0)	0.254
SCF Pass-2 + Stage 3	9.36 +/- 10.57	[4.24, 16.19]	3.73	0.683 / 0.720	71.0 %	54.2 %	73.8 %	59.1 %	78.0 %	99.6 % (1)	0.226
Anatomical SCF + Stage 3	2.88 +/- 1.32	[2.23, 3.73]	2.41	0.845 / 0.915	95.1 %	91.7 %	95.0 %	100.0 %	95.1 %	99.7 % (1)	0.088
3D U-Net + Z-sort	7.57 +/- 7.43	[3.97, 12.17]	2.75	0.710 / 0.813	83.6 %	70.8 %	86.7 %	63.6 %	92.0 %	100.0 % (0)	0.147
3D U-Net + Stage 3	3.40 +/- 1.99	[2.42, 4.64]	2.30	0.835 / 0.902	92.0 %	66.7 %	95.4 %	81.8 %	93.0 %	100.0 % (0)	0.121
Hybrid + Viterbi (no RibSnake)	2.85 +/- 1.70	[2.07, 3.95]	2.11	0.872 / 0.933	95.8 %	83.3 %	97.5 %	90.9 %	95.8 %	99.6 % (1)	0.147
Hybrid + full Stage 3	2.05 +/- 0.48	[1.80, 2.33]	1.91	0.886 / 0.945	97.2 % (278/286)	91.7 %	97.5 %	100.0 %	97.2 %	100.0 % (0)	0.068

On the unmodified ground truth (285 ribs): 2.08 +/- 0.46 mm, 96.8 % (276/285). All 8 ribs failing the symmetric criterion are partial traces (prediction lies on the rib, but covers only part of it), not label swaps.

Frozen pipeline on the 47 annotated training scans (1,116 ribs; detectors in-sample, Stage-3 constants never inspected on these scans; benchmark_train47.json): LCD 2.63 +/- 1.91 mm (CI [2.14, 3.23], median 1.89), labeled clDice 0.885 / 0.932, labeling accuracy 95.4 % (1065/1116), recall 95.6 %, precision 98.2 % (20 FP curves). 5-fold cross-validation across all 59 scans (1,402 ribs; per-fold re-fitted atlas, Anatomical SCF and 3D U-Net, frozen Stage-3 constants; benchmark_cv5.json): LCD 3.64 +/- 3.90 mm (CI [2.79, 4.74], median 2.39 mm), labeled clDice 0.830 / 0.889, symmetric labeling accuracy 91.8 % (1287/1402, Wilson 95 % CI 90.2 - 93.1 %), recall 94.6 %, precision 97.7 % (33 FP curves), NLD 0.097 (Wilcoxon p = 2.0e-9 for RibSnake on LCD and p = 3.2e-9 on labeled clDice@3mm).

Limitations

• 60-scan single-source dataset at 2 mm; no acquisition metadata; untested outside SCR.
• 12-scan non-random evaluation set; Wilson CI for accuracy 94.6 - 98.6 %; subject-level failure rate only bounded below ~25 %.
• Stage-3 constants fitted with access to the evaluation scans (see caveat).
• Known annotation issues (merged R07 in RIB_007, inconsistent 12th ribs); possible patient overlap Bro_004029 (train) / Sko_004029 (eval) cannot be excluded.
• Published numbers from other papers (Wu 2012, Lenga 2018, RibSeg v2) use other cohorts and criteria; no comparative claim is made.

---

Quick start

from ribline import load_pretrained_models

# Automatically downloads and instantiates the 24-rib atlas, Anatomical SCF, and 3D U-Net
models = load_pretrained_models("wq2012/ribline-3d-ct-centerline", device="cpu")
atlas = models["atlas"]
rf_anat = models["forest"]
unet = models["unet"]

Or load individual local checkpoint files directly:

import torch
from ribline import RibCageStatisticalAtlas, RibForest
from ribline.model.unet.model import RibUNet3D

atlas = RibCageStatisticalAtlas.load("rib_atlas.joblib")
rf_anat = RibForest.load("rib_forest_anatomical.joblib")
unet = RibUNet3D(in_channels=1, out_channels=1, base_filters=16)
unet.load_state_dict(torch.load("rib_unet_3d.pth", map_location="cpu"))
unet.eval()

See the GitHub README for the full inference / Stage-3 / evaluation commands.

---

Citation

• Paper: https://arxiv.org/abs/2610.16047 (PDF)

@article{wang2026hybrid,
  title={Hybrid Context Forests and Bilateral Viterbi Matching for {3D} Rib Centerline Tracing in {CT}},
  author={Wang, Quan},
  journal={arXiv preprint arXiv:2610.16047},
  year={2026},
  url={https://arxiv.org/abs/2610.16047}
}

@inproceedings{wu2012learning,
  title={A learning based deformable template matching method for automatic rib centerline extraction and labeling in {CT} images},
  author={Wu, Dijia and Liu, David and Pusk{\'a}s, Zsolt and Lu, Caiyun and Wimmer, Andreas and Tietjen, Christian and Soza, Grzegorz and Zhou, S Kevin},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages={980--987},
  year={2012},
  organization={IEEE}
}

@inproceedings{wang2013semantic,
  title={Semantic context forests for learning-based knee cartilage segmentation in {3D} {MR} images},
  author={Wang, Quan and Wu, Dijia and Lu, Le and Liu, Meizhu and Boyer, Kim L and Zhou, Shaohua Kevin},
  booktitle={International MICCAI Workshop on Medical Computer Vision},
  pages={105--115},
  year={2013},
  organization={Springer}
}