kabilasoboka/RibAssist-3D

RibAssist 3D checkpoints (v1.0.0)

Biplanar rib-fracture detection, anatomical addressing, and selective 3D localization from CT-derived projections, with assignment-based cross-view correspondence and explicit abstention.

Resource	Link
This model repo	kabilasoboka/RibAssist-3D
Source code	kabJhai/RibAssist-3D
Paper	arXiv:2608.06914
Demo video	YouTube
Terms	MODEL_TERMS.md

v1.0.0 checkpoints are bundled in this folder (SHA256 in config.yaml). Optional pair_scorer/ weights are omitted unless you run Section 12 in REPRODUCE.md.

Project overview

RibAssist 3D is a research proof-of-concept assistive review system. It detects suspected rib fractures in AP and lateral CT-derived projections, assigns predicted side and rib level, and selectively localizes sufficiently confident cross-view pairs as 3D points under a frozen false-output budget. When correspondence is uncertain, the system abstains from 3D localization while preserving detections and rib addressing for human review.

These checkpoints reproduce the sealed headline result reported in the paper (2.50% end-to-end commitment yield at 0.436 false 3D points per case on the 55-case sealed cohort).

Model architecture

Three trained networks (see GitHub README for layer tables):

1. AP / lateral detector U-Nets (base channels 32, ~1.95M params per view): single-channel 256×256 in → fracture heatmap out; two independent copies for AP and lateral.
2. Addressing network (~49k params): dual-stream CNN on raw AP/lat projections → side + rib level.
3. Learned pair-scorer (~13k params, optional): local-appearance correspondence for 2×2 attribution; not required for the main sealed/demo path.

Cross-view fusion is geometric (SI-gated candidate graph + one-to-one assignment with abstention), not a learned fusion network.

Included checkpoints

Required v1.0.0 weights (SHA256 in config.yaml):

RibAssist-3D/
├── README.md
├── LICENSE
├── MODEL_TERMS.md
├── CITATION.cff
├── config.yaml
├── training_config.yaml
├── requirements.txt
├── assets/
│   └── fig_architecture.svg
├── ap_detector/
│   └── ribassist_ap_detector.pt
├── lateral_detector/
│   ├── ribassist_lateral_detector_frozen.pt
│   └── ribassist_lateral_detector_l2.pt
├── addressing/
│   └── ribassist_addressing.pt
└── pair_scorer/
    └── ribassist_pair_scorer.pt          # optional; not in v1.0.0 bundle

File	Role
ribassist_ap_detector.pt	AP U-Net (shared weights for champion and L2 stacks)
ribassist_lateral_detector_frozen.pt	Champion / frozen-head lateral U-Net
ribassist_lateral_detector_l2.pt	L2 retrained lateral U-Net (sealed headline stack)
ribassist_addressing.pt	Rib-level addressing network
ribassist_pair_scorer.pt	D2b learned pair-scorer (optional)

Training data

These weights were trained on RibFrac CT-derived projections and fracture labels, and use RibSeg-derived annotations and anatomy in the training and evaluation pipeline.

• RibFrac (MICCAI 2020): CC BY-NC 4.0. CT volumes and labels are not included in this repository.
• RibSeg v2 (M3DV/RibSeg): annotations derived from RibFrac source CTs. Segmentation/centerline files are not included here.

Users must obtain RibFrac and RibSeg separately and comply with upstream dataset terms. See GitHub DATA_SETUP.md.

Intended use

• Non-commercial research and educational use only (CC BY-NC 4.0).
• Reproduce sealed evaluation, figures, and the clinician-review demo workflow with locally obtained data.
• Study biplanar fracture detection, assignment with abstention, and assistive review UX.

Limitations

• CT-derived orthographic projections, not independently acquired clinical radiographs.
• Research proof of concept only; not validated for clinical diagnosis or patient care.
• No clinician reader study; no evidence of improved sensitivity, reading time, or outcomes.
• No negative-scan safety evaluation on the sealed cohort (55 fracture-positive cases only).
• Low end-to-end yield (2.5% commitment at the operational budget); not a comprehensive reconstructor.
• Small sealed cohort (55 cases, 601 fractures).
• Rib addressing accuracy was not independently evaluated in the sealed study.
• Checkpoints inherit RibFrac CC BY-NC constraints; commercial use is not authorized without separate permission.

License

CC BY-NC 4.0 — non-commercial research and educational use only. See LICENSE and MODEL_TERMS.md.

RibAssist source code on GitHub is Apache-2.0; these weights are not.

Loading checkpoints

pip install -r requirements.txt

from ribassist import load_ap_detector, load_lateral_detector_l2, load_addressing

ap_net = load_ap_detector("ap_detector/ribassist_ap_detector.pt")
lat_net = load_lateral_detector_l2("lateral_detector/ribassist_lateral_detector_l2.pt")
addr_net = load_addressing("addressing/ribassist_addressing.pt")

Full inference, sealed evaluation, and demo: clone kabJhai/RibAssist-3D and follow REPRODUCE.md.

See examples/inference_example.py.

Citation

You must cite RibAssist 3D, RibFrac, and RibSeg v2 when using these checkpoints.

RibAssist 3D

@article{soboka2026ribassist,
  title   = {RibAssist 3D: Biplanar Rib-Fracture Detection, Addressing, and Selective 3D Localization from CT-Derived Projections},
  author  = {Soboka, Kabila Haile},
  year    = {2026},
  eprint  = {2608.06914},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url     = {https://arxiv.org/abs/2608.06914}
}

RibFrac

@inproceedings{jin2020ribfrac,
  title     = {Accurate Detection and Segmentation of Rib Fractures on Chest {CT}},
  author    = {Jin, Peng and others},
  booktitle = {MICCAI Challenge on Rib Fracture Detection and Classification (RibFrac)},
  year      = {2020},
  url       = {https://ribfrac.grand-challenge.org/}
}

(Replace with the exact BibTeX from the Zenodo record you downloaded.)

RibSeg v2

@article{ribseg2023,
  title  = {RibSeg v2: A Large-Scale Benchmark for Rib Segmentation},
  author = {M3DV RibSeg Team},
  year   = {2023},
  url    = {https://github.com/M3DV/RibSeg}
}

(Use the citation requested in the RibSeg repository.)

Also cite CITATION.cff.

Versioning

Version	Description
v1.0.0	Checkpoints accompanying the current paper (frozen-head + L2 sealed stack)
v1.1.0	(planned) Improved lateral detector
v2.0.0	(planned) Improved correspondence

Tag releases on Hugging Face when updating weights.

Cross-links

• GitHub: kabJhai/RibAssist-3D (code, splits, sealed JSONs, reproduction)
• Paper: arXiv:2608.06914
• Demo: YouTube

Disclaimer

Not for diagnosis, treatment, or patient care. See DISCLAIMER.md.