ribline: 3D Rib Centerline Tracing and Anatomical Labeling in Thoracic CT

ribline is an open-source Python library for automatic 3D rib centerline extraction, anatomical rib numbering (R01..R12, L01..L12) and sub-voxel refinement in thoracic CT (Wang, arXiv:2610.16047). It combines a lung-apex / vertebral-midline thoracic coordinate frame, a hybrid voxel detector (anatomically anchored Semantic Context Forest + lightweight 3D U-Net), and a global labeling stage (bilateral Viterbi dynamic programming + RibSnake refinement). It re-implements and extends ideas from learning-based deformable template matching (Wu et al., CVPR 2012) and Semantic Context Forests (Wang et al., MICCAI MCV 2013). [arXiv] [PDF] [Hugging Face collection]

Version 0.3.1 (revision). This release fixes the metric definitions (symmetric 10 mm labeling criterion, 50 mm penalty for missed ribs, precision / false positives, nan instead of a default 1.0 for undefined groups), ships the U-Net checkpoint that actually produced the reported probability maps, makes the training scripts config-faithful, adds reproducible atlas / ablation / statistics / 5-fold cross-validation scripts and Hugging Face Hub loaders, fixes one merged-rib annotation in the public ground truth, and documents the evaluation-protocol caveats below. All numbers in this README were produced by the scripts named next to them.

Component	ribline
Voxel detector	Anatomically anchored SCF (150 trees, 286 features) max-fused with a plain 3-level 3D U-Net (340k parameters)
Thoracic frame	Automatic lung apex Z_apex, vertebral midline X_spine(Z), kyphosis-aware azimuthal angle (ordering parameter)
Labeling	Fragment consolidation + patient-scaled bilateral Viterbi DP with contralateral coupling
Refinement	Apical clavicle trimming, terminal-rib template recovery, normal-plane centering (RibSnake)
Labeled Chamfer distance (12 eval scans, 286 ribs)	2.05 +/- 0.48 mm, 95 % bootstrap CI [1.80, 2.33], median 1.89 mm
Labeled clDice 3 mm / 4 mm	0.886 / 0.945
Labeling accuracy (symmetric 10 mm)	97.2 % (278 / 286 ribs; Wilson 95 % CI 94.6 - 98.6 %)
Recall / precision	97.2 % / 100.0 % (0 false-positive curves out of 286)

---

Table of Contents

1. Installation
2. Dataset, Split and Ground-Truth Fix
3. Pretrained Models
4. Usage
5. Algorithm (as implemented)
6. Evaluation Protocol
7. Results
8. Limitations and Known Issues
9. Reproducing Every Number
10. Citation

---

Installation

pip install ribline

or for development with PyTorch support:

git clone https://github.com/wq2012/ribline.git
cd ribline
pip install -e ".[torch]"
python run_tests.py

---

Dataset, Split and Ground-Truth Fix

The package supports the Siemens Corporate Research (SCR) 3D thoracic CT rib centerline benchmark (60 scans, 2.0 mm isotropic, Analyze-style .hdr/.img with 3-line ASCII headers, manual B-spline interpolated centerlines *_interp.xls) as well as NIfTI / MetaImage volumes via SimpleITK.

• Split (split.json): 48 training scans (47 with annotations, 1,116 ribs; Gist_p2_19_Anonymous_0015_interp.xls is empty and skipped) and 12 evaluation scans. The split is not random: the 12 evaluation scans are the case-sensitive alphabetical tail of the file list. See Limitations.
• Coordinates: centerline points are in physical millimeters (X, Y, Z) = (i_x s_x, i_y s_y, i_z s_z) with origin (0, 0, 0); X right-to-left, Y anterior-to-posterior, Z inferior-to-superior.
• Intensities: about half of the scans are stored as signed HU and half as unsigned shifted values. ribline.model.anatomy.standardize_hu subtracts 1024 when the volume minimum is >= -100 HU. This standardization is used by the forests and by Stage 3; the U-Net is fed a per-volume z-score of the raw stored intensities (see below).
• Ground-truth fix: in RIB_007_20091210_bone_1.5_0.6_12_38_interp.xls the label R07 contains two ribs (640 points with a 220 mm jump; the second curve is the eighth right rib). python -m ribline.scripts.fix_gt_annotations --gt-dir data/Centerlines_GT --out-dir data/Centerlines_GT_fixed splits it into R07 / R08, touches no other file, and writes fix_log.json. The fixed set has 286 evaluation ribs; results on the unmodified ground truth are reported as well.

---

Pretrained Models

All pretrained checkpoints were trained on the 47 annotated training scans only and are published on Hugging Face in the ribline 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 %	--	--

Checkpoint Files

File	Component	What it actually is
rib_forest_pass1.joblib + feature_config_pass1.json	Pass-1 forest	120 trees, unlimited depth, 228 features: raw intensity, per scale (1, 2, 4 mm) smoothed intensity, gradient magnitude, 6 second derivatives, Laplacian (28), plus 200 RSID pairs with offsets in [-20, 20] voxels (+/-40 mm). Trained with random rigid augmentation.
rib_forest_pass2.joblib + feature_config_img.json + feature_config_ctx.json	Pass-2 auto-context forest	100 trees, 306 features = the 228 Pass-1 features + 28 multi-scale features of the Pass-1 probability map + 50 random-shift probability differences with offsets in [-40, 40] voxels (+/-80 mm).
rib_forest_anatomical.joblib + feature_config_anatomical.json	Anatomically anchored SCF	150 trees, depth 24, 286 features: HU-standardized input, the 28 differential features, a per-scale tubularity proxy from the diagonal second derivatives (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, 64^3 patches, loss 0.5 soft-Dice + 0.5 BCE, Adam 1e-3, batch 2, 8 patches per scan per epoch, 20 epochs, no augmentation. This file is the epoch-20 checkpoint (md5 9e456ba9ff81c402a58436c666f152ef) that produced every probability map reported below (re-inference reproduces them with max abs. difference 0). Earlier releases shipped an unrelated mid-run snapshot by mistake. Training recipe: pretrained_models/train_config_unet.json.
rib_atlas.joblib	24-rib statistical atlas	Mean canonical curves (80 points), 5-D descriptor statistics, inter-rib spacing statistics and a 15,000-point slant-correction cloud. Re-fitting with python -m ribline.scripts.build_atlas on the 47 training scans reproduces this file exactly.
benchmark_val12_fixed_gt.json, benchmark_val12_original_gt.json, benchmark_train47.json, benchmark_cv5.json	Evaluation reports	Per-rib / per-subject metrics of all ablation rows on the 12 eval scans (fixed and unmodified GT), the 47 training scans, and the 59-scan 5-fold CV (corrected metrics, bootstrap CIs, paired tests).
polar_stats_all59.json, gt_fix_log.json	Statistics / provenance	Monotonicity and frame statistics on all 59 scans; log of the ground-truth fix.

---

Usage

Hugging Face Hub Quick Start

from ribline import load_pretrained_models

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

Python API

import numpy as np
import SimpleITK as sitk
from ribline import (
    load_image, parse_rib_xls, extract_thoracic_frame, RibCageStatisticalAtlas,
    CenterlineExtractor, BilateralViterbiLabeler, RibSnake, evaluate_subject_ribs,
)

ct_img = load_image("data/vol2mm/pet_ct_bc__eh_Anonymous_0004.hdr")
ct_arr = sitk.GetArrayFromImage(ct_img).astype(np.float32)
spacing, origin = ct_img.GetSpacing(), ct_img.GetOrigin()

unet_prob = sitk.GetArrayFromImage(load_image("predictions_unet/pet_ct_bc__eh_Anonymous_0004_prob.nii.gz"))
scf_prob = sitk.GetArrayFromImage(load_image("predictions_scf/pet_ct_bc__eh_Anonymous_0004_prob.nii.gz"))
hybrid_prob = np.maximum(unet_prob, 0.75 * scf_prob)

frame = extract_thoracic_frame(ct_arr, spacing=spacing, origin=origin)
atlas = RibCageStatisticalAtlas.load("ribline/configs/rib_atlas.joblib")
curves = CenterlineExtractor(threshold=0.45).extract(
    hybrid_prob, spacing=spacing, origin=origin, min_length=22.0, hu_arr=ct_arr, frame=frame, atlas=atlas)
labeled = BilateralViterbiLabeler(atlas=atlas).match(curves, frame=frame)
refined = RibSnake(spacing=spacing).refine(labeled, ct_arr, origin=origin, prob_np=hybrid_prob, frame=frame, atlas=atlas)

gt = parse_rib_xls("data/Centerlines_GT_fixed/pet_ct_bc__eh_Anonymous_0004_interp.xls")
m = evaluate_subject_ribs(gt, refined)   # symmetric 10 mm criterion, 50 mm missing penalty
print(m["labeled_chamfer_mm"], m["label_acc_all"], m["precision"], m["num_false_positive_curves"])

Command line

# 1. Fix the merged-rib annotation in RIB_007 and build the 24-rib statistical atlas
python -m ribline.scripts.fix_gt_annotations --gt-dir data/Centerlines_GT --out-dir data/Centerlines_GT_fixed
python -m ribline.scripts.build_atlas --image-dir data/vol2mm --gt-dir data/Centerlines_GT_fixed \
    --split split.json --out pretrained_models/rib_atlas.joblib

# 2. Train the Anatomically Anchored SCF and the 3D U-Net
python -m ribline.scripts.train_anatomical_scf --image-dir data/vol2mm --gt-dir data/Centerlines_GT_fixed \
    --split split.json --out-dir pretrained_models
python -m ribline.scripts.train_unet --image-dir data/vol2mm --gt-dir data/Centerlines_GT_fixed \
    --split split.json --epochs 20 --out pretrained_models/rib_unet_3d.pth
# (or from a JSON config: python -m ribline.pipeline.train_unet --config experiments/sota/full_config.json)

# 3. Detector probability maps (any split: val | train | all)
python -m ribline.scripts.infer_anatomical_scf --model pretrained_models/rib_forest_anatomical.joblib \
    --image-dir data/vol2mm --split split.json --out-dir predictions/scf_anatomical
python -m ribline.scripts.infer_unet --model pretrained_models/rib_unet_3d.pth \
    --image-dir data/vol2mm --split split.json --out-dir predictions/unet

# 4. Stage 3 (hybrid fusion, bilateral Viterbi DP, RibSnake)
python -m ribline.scripts.run_stage3 --unet-probs predictions/unet --scf-probs predictions/scf_anatomical \
    --atlas pretrained_models/rib_atlas.joblib --image-dir data/vol2mm --gt-dir data/Centerlines_GT_fixed \
    --split split.json --out pretrained_models/benchmark_val12_fixed_gt.json
# (or via console entry point: ribline-sota --prob-dir predictions/unet --rf-prob-dir predictions/scf_anatomical ...)

# 5. Scoring with bootstrap CIs and paired Wilcoxon tests, plus 5-fold cross-validation
python -m ribline.scripts.benchmark_eval --gt-dir data/Centerlines_GT_fixed \
    --method ours=predictions/unet_stage3 --out benchmark.json
python -m ribline.scripts.run_cv5 --image-dir data/vol2mm --gt-dir data/Centerlines_GT_fixed \
    --folds experiments/cv5/folds.json --out pretrained_models/benchmark_cv5.json

---

Algorithm (as implemented)

1. Thoracic frame (ribline/model/anatomy.py): HU standardization; on a 2x down-sampled grid, torso (> -350 HU), internal air (< -380 HU, opened, <= 4 largest components) = lungs; Z_apex = max of per-side 99.5th-percentile lung Z; spine from the bone projection in the posterior 40 % of the AP depth, per-slice median, Gaussian-smoothed. Kyphosis-aware angle theta = atan2(-(Y - y_c(Z)), |X - x_spine| - 8 mm) with y_c(Z) = Y_ant + 0.25 D + 0.16 clip(Z_apex - Z, 0, 220). Measured on 1,402 GT ribs of the 59 scans (python -m ribline.scripts.polar_angle_stats): theta_end > theta_start on 100 %, strictly monotonic on 78.6 % (vs 57.0 % for a torso-center angle), local reversal > 0.1 rad on 0.6 % (vs 7.6 %). Posterior end of rib 1 lies +11.9 +/- 9.3 mm above Z_apex; midline error 2.5 +/- 2.4 mm.
2. Detectors (ribline/model/features.py, ribline/model/unet/): see the table above. The forests are evaluated on all non-air voxels (gaussian(HU, 1.2) > -20), a speed optimization only; forest training negatives are sampled uniformly over the whole volume.
3. Stage 3 (reconstruction.py, topology.py, refinement.py): anatomical mask (bone band, vertebral / sternal corridors, shoulder exclusion, intercostal valley suppression), threshold 0.45, 3D skeleton, graph decomposition, hairpin / height-jump splitting, azimuthal chaining; per-hemithorax candidate consolidation; patient scale s_Z; relative Mahalanobis + template unaries; bilateral Viterbi DP with clutter costs and skip penalties, two passes with contralateral coupling; RibSnake (apical trimming at 0.55 W, terminal-rib recovery, 5 iterations of normal-plane centering clamped to 1.6 mm, cubic smoothing spline). Output curves have 15 to ~200 points (median ~150).

---

Evaluation Protocol

Implemented in ribline.model.metrics.evaluate_subject_ribs (defaults = protocol below):

Metric	Definition
Labeled Chamfer distance (LCD)	mean over all GT ribs of 0.5 (d_p2g + d_g2p) between a GT rib and the prediction with the same label; a GT rib without prediction counts 50 mm
Labeled clDice (3 / 4 mm)	per-rib centerline Dice, missing ribs = 0
Labeling accuracy	GT rib correct iff same-label prediction exists, d_p2g <= 10 mm and d_g2p <= 10 mm, and the GT rib is the closest one to that prediction; pooled over ribs (all / first / intermediate 2-11 / twelfth); undefined groups give nan
Recall	GT rib detected iff some predicted curve satisfies the symmetric criterion
Precision / FP	predicted curve is a true positive iff d_p2g <= 10 mm to some GT rib (curve_precision_symmetric requires both d_p2g <= 10 mm and d_g2p <= 10 mm)
NLD	normalized length difference, missing = 1
Uncertainty	subject-level mean +/- SD with 95 % bootstrap CI (10,000 resamples); Wilson 95 % CI for pooled rates; two-sided Wilcoxon signed-rank tests between methods

The previous release used a one-sided 8 mm criterion, dropped missed ribs from the LCD mean, defaulted 12th-rib accuracy to 1.0 when no 12th ribs existed and reported unlabeled clDice; the legacy quantities are still returned under *_onesided / *_matched_only keys for comparison.

Protocol caveat (please read). The ~60 constants of Stage 3 were tuned by repeatedly running the pipeline and inspecting the metrics on the 12 evaluation scans (the development history is in the project notes). The three learned detectors and the atlas were fitted on the 47 training scans only. Numbers on the 12 scans are therefore in-sample for the Stage-3 constants. The cross-validation below quantifies the effect.

---

Results

Ablation on the 12 evaluation scans (286 GT ribs, fixed GT)

Source: experiments/rebuttal/ablation_val12/benchmark.json (= pretrained_models/benchmark_val12_fixed_gt.json), produced by python -m ribline.scripts.run_ablation --spec experiments/rebuttal/ablation_val12_spec.json from the stored probability maps with the released Stage-3 code. "Z-sort" = same extraction and consolidation, candidates of each side labeled 1..12 by decreasing height, no RibSnake.

Method	LCD (mm) mean +/- SD	95 % CI	UCD	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 (ours)	2.05 +/- 0.48	[1.80, 2.33]	1.91	0.886 / 0.945	97.2 %	91.7 %	97.5 %	100.0 %	97.2 %	100.0 % (0)	0.068

Paired Wilcoxon tests, ours vs. "Hybrid + Viterbi": LCD p = 0.009, NLD p = 0.0005, labeled clDice@3 mm p = 0.20, accuracy p = 0.50 (not significant on 12 subjects). Ours vs. "3D U-Net + Stage 3": LCD p = 0.001, accuracy p = 0.008.

Per-subject results of the full pipeline

Evaluation scan	GT ribs	Pred. curves	LCD (mm)	UCD (mm)	lab. clDice 3 / 4 mm	Acc all	Recall	FP	NLD	ribs failing the symmetric criterion
R3_anon_0002713245_0002	24	24	2.12	2.13	0.864 / 0.960	100.0 %	100.0 %	0	0.057	--
RIB_007_20091210_bone_1.5_0.6_12_38	24	24	1.53	1.45	0.934 / 0.960	100.0 %	100.0 %	0	0.057	--
Sko_004029_0040290098_0002	24	24	2.50	1.92	0.889 / 0.942	95.8 %	95.8 %	0	0.079	L07 (traced 93 of 291 points)
Tsc_004002_0040027642_0002	22	22	2.70	2.64	0.718 / 0.864	100.0 %	100.0 %	0	0.079	--
Zer_004061_0040610659_0001	24	24	2.98	2.74	0.899 / 0.924	79.2 %	79.2 %	0	0.114	L04-L08 (anterior third missing)
pat1_Anonymous_0002	24	24	2.48	2.28	0.817 / 0.935	91.7 %	91.7 %	0	0.087	R01, L01 (partial)
pat3_Anonymous_0002	24	24	1.87	1.68	0.891 / 0.948	100.0 %	100.0 %	0	0.064	--
pat8_Anonymous_0001	24	24	1.72	1.65	0.902 / 0.953	100.0 %	100.0 %	0	0.038	--
pet_ct_bc__eh_Anonymous_0004	24	24	1.44	1.37	0.946 / 0.973	100.0 %	100.0 %	0	0.047	--
pet_ct_bc__mk_Anonymous_0004	24	24	1.68	1.54	0.946 / 0.959	100.0 %	100.0 %	0	0.075	--
pet_ct_bc__sd_Anonymous_0005	24	24	1.70	1.64	0.927 / 0.961	100.0 %	100.0 %	0	0.040	--
t2_08_20100323_bone_1.5_0.6_8_136	24	24	1.92	1.93	0.895 / 0.956	100.0 %	100.0 %	0	0.078	--
Mean (12 scans, 286 ribs)	286	286	2.05 +/- 0.48	1.91 +/- 0.43	0.886 / 0.945	97.2 %	97.2 %	0	0.068	8 partially traced ribs

All 8 failures are partial traces (d_p2g <= 4.2 mm, d_g2p 10.6 - 34.8 mm), not label swaps; under the legacy one-sided criterion they count as correct. 25 / 286 ribs have HD95 > 10 mm (8 above 20 mm); 52 have clDice@3 mm < 0.8 (5 below 0.5).

Sensitivity: unmodified GT (285 ribs) 2.08 +/- 0.46 mm, 96.8 % (276/285); unmodified GT with the merged R07 excluded (284 ribs) 2.05 +/- 0.48 mm, 97.2 %; without Sko_004029 (possible duplicate of a training patient) 2.01 +/- 0.48 mm [1.75, 2.31], 97.3 % (255/262). Sources: benchmark_gt_original.json, benchmark_gt_original_excl_rib007_R07.json, benchmark_noSko.json in experiments/rebuttal/ablation_val12/.

Cross-validation with frozen Stage-3 constants and evaluation on the training scans

(a) Frozen pipeline on the 47 annotated training scans (experiments/rebuttal/run_train47.py -> experiments/rebuttal/ablation_train47/benchmark.json; detectors in-sample, Stage-3 constants never inspected on these scans, 1,116 GT ribs):

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
Anatomical SCF + Stage 3	3.62 +/- 3.97	[2.67, 4.91]	2.97	0.830 / 0.895	92.4 %	89.2 %	92.4 %	95.7 %	94.8 %	97.4 % (29)	0.094
3D U-Net + Stage 3	5.84 +/- 7.20	[3.95, 8.03]	2.98	0.785 / 0.826	83.9 %	69.9 %	86.0 %	76.3 %	91.6 %	97.8 % (25)	0.142
Hybrid + Viterbi (no RibSnake)	2.98 +/- 2.11	[2.42, 3.63]	2.77	0.875 / 0.930	95.6 %	88.2 %	96.1 %	97.8 %	96.1 %	98.5 % (17)	0.155
Hybrid + full Stage 3	2.63 +/- 1.91	[2.14, 3.23]	2.62	0.885 / 0.932	95.4 % (1065/1116)	86.0 %	96.1 %	97.8 %	95.6 %	98.2 % (20)	0.083

Median LCD 1.89 mm (same as on the 12 scans), 28 / 47 scans labeled perfectly, 4 scans with LCD > 5 mm (worst HP_485533 10.3 mm, 03_Anonymous_0002 9.9 mm, label shifts on one side). The gap to the 12-scan figures (2.05 mm / 97.2 %) is the price of tuning Stage 3 on those scans plus the larger anatomical variety of the training cohort; treat the 12-scan numbers as an optimistic upper bound.

(b) 5-fold cross-validation over all 59 annotated scans (experiments/rebuttal/cv_fold.py {0..4} -> experiments/rebuttal/benchmark_cv5.json, shipped as pretrained_models/benchmark_cv5.json; per-fold re-fitted atlas, Anatomical SCF and 3D U-Net, frozen Stage-3 constants, 1,402 GT ribs):

Method	LCD (mm)	95 % CI	Median LCD	UCD	lab. clDice 3 / 4 mm	Acc all	Acc 1st	Acc 2-11	Acc 12th	Recall	Precision (FP)	NLD
Anatomical SCF + Stage 3	5.28 +/- 4.62	[4.16, 6.56]	3.08	3.44	0.769 / 0.840	86.9 % (1218/1402)	72.6 %	87.7 %	93.0 %	90.8 %	95.3 % (67)	0.126
3D U-Net + Stage 3	7.44 +/- 8.93	[5.31, 9.85]	2.99	3.30	0.745 / 0.785	78.2 % (1097/1402)	61.5 %	80.9 %	67.8 %	87.0 %	98.1 % (26)	0.170
Hybrid + Viterbi (no RibSnake)	4.17 +/- 4.01	[3.29, 5.28]	2.78	3.07	0.807 / 0.879	91.4 % (1282/1402)	81.2 %	92.2 %	93.9 %	94.3 %	97.9 % (29)	0.195
Hybrid + full Stage 3	3.64 +/- 3.90	[2.79, 4.74]	2.39	2.85	0.830 / 0.889	91.8 % (1287/1402)	82.9 %	92.2 %	96.5 %	94.6 %	97.7 % (33)	0.097

Wilson 95 % CI for Acc all of Hybrid + full Stage 3: 90.2 - 93.1 %. Per-fold mean LCD across folds 0..4: 3.12, 4.80, 4.91, 2.84, 2.43 mm. 30 / 59 scans achieve 100 % symmetric labeling accuracy; 8 scans have LCD > 5 mm (worst: BC3_Anonymous_0001 26.58 mm, BC1_Anonymous_0001 15.76 mm). Paired Wilcoxon signed-rank tests across all 59 subjects: Hybrid + full Stage 3 vs. Hybrid + Viterbi (no RibSnake): LCD p = 2.0e-9, labeled clDice@3mm p = 3.2e-9, NLD p = 4.4e-11; vs. Anatomical SCF + Stage 3: LCD p = 2.1e-7, Acc all p = 9.0e-5; vs. 3D U-Net + Stage 3: LCD p = 1.9e-5, Acc all p = 8.5e-6.

Published numbers (not directly comparable)

Wu et al. 2012 report 89.5 - 95.9 % labeling accuracy on an SCR cohort, Lenga et al. 2018 94.1 %, and RibSeg v2 (2024) an LCD of 3.15 mm / 96.5 % on 0.7 mm RibFrac CT. Different scans, resolutions, splits and matching criteria; none of these methods was re-run here, and no claim of outperforming them is made.

---

Limitations and Known Issues

• Small, single-source dataset: 60 scans (59 annotated) at 2 mm from one archive; no acquisition metadata; generalization to other scanners / resolutions / pathologies is untested.
• Non-random, tiny held-out set: 12 scans (286 ribs), alphabetical tail of the file list. 8 errors in 286 ribs give a Wilson CI of 94.6 - 98.6 % for labeling accuracy; 0 subject-level failures in 12 scans only bound the per-scan failure rate below ~25 %.
• Stage-3 constants were tuned with the evaluation scans visible. There is no untouched test set for this dataset. Treat the released constants as fitted to SCR; see the cross-validation.
• Annotation errors / inconsistencies: one merged-rib annotation (fixed, logged); 12th ribs annotated inconsistently (absent in Tsc_004002); some residual error is annotation noise.
• Possible patient overlap between Bro_004029_0040290098_0001 (train) and Sko_004029_0040290098_0002 (eval): different matrix / intensity encoding / FOV (voxel correlation 0.25) but rib-length fingerprints among the 1 % most similar of all scan pairs; cannot be resolved from anonymized data (experiments/rebuttal/provenance.json, bro_sko_lengths.json).
• Dominant failure mode: ribs traced only over their posterior part across a costal-cartilage gap (counted as errors by the symmetric criterion).
• Pass-2 forest provenance: the released Pass-2 forest was trained on Pass-1 maps from an earlier Pass-1 forest (experiments/full_v4), not from the released Pass-1 forest; it is kept as an ablation baseline only.
• ribline/scripts/correct_gt.py is a development-time experiment that produced a copy of the GT identical to the original (verified for all 60 files); it is not part of the protocol.

---

Reproducing Every Number

Quantity	Script / file
Ablation table, per-subject table, CIs, paired tests	python -m ribline.scripts.run_ablation --spec experiments/rebuttal/ablation_val12_spec.json -> benchmark.json
Alternative GT protocols	experiments/rebuttal/score_val12.py
Monotonicity, rib-1 offset, midline error	python -m ribline.scripts.polar_angle_stats ... -> polar_stats_all59.json
Ground-truth fix	python -m ribline.scripts.fix_gt_annotations ... -> fix_log.json
Checkpoint identity, Centerlines_GT_corrected == GT, atlas refit, Bro/Sko	experiments/rebuttal/provenance.py -> provenance.json
Atlas	python -m ribline.scripts.build_atlas ...
Frozen Stage 3 on the 47 training scans	experiments/rebuttal/run_train47.py
5-fold cross-validation	experiments/rebuttal/cv_fold.py {0..4}
Unit tests (25, incl. corrected metrics, multi-format I/O, and CLI entry points)	python run_tests.py

---

Citation

If you find ribline or its pretrained models useful in your research, please cite:

• 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}
}