RibFrac2020 — Rib Fracture Detection & Segmentation on Chest CT

The MICCAI 2020 RibFrac Challenge: detect and segment rib fractures in chest–abdomen CT. Fracture lesions are small, numerous, and scattered along the rib cage — this is a hard, high-instance-count lesion segmentation task, not an organ segmentation one.

What this mirror contains — read first

500 GT-annotated CTs, not the full 660-case release. The challenge ships 420 train + 80 validation with ground truth and 160 test without. The test labels were never released and are still withheld for the live leaderboard, so a test split would be image-only and worthless for a segmentation benchmark. This mirror therefore carries the 500 annotated cases only (RibFrac1–RibFrac500).

Ground truth here is BINARY. The upstream masks are instance-labeled (1..N, one integer per fracture). Instance integers are preserved verbatim in labels/, but the intended MedOtter target is mask > 0 → a single rib_fracture foreground class. See Why binary below — this is a deliberate choice forced by the upstream -1 Ignore class, not a shortcut.

The paper's cohort is larger than the public release. The EBioMedicine (FracNet) paper reports a 900-patient / 7,473-fracture in-house cohort with a 720/60/120 split. The public release is 660 patients / 5,304 fractures. FracNet's own README acknowledges the divergence. Numbers quoted from the paper do not describe this data.

Part1 / Part2 is not a semantic split. Zenodo splits the training images across two records purely to stay under a per-file size limit. This mirror merges them into one train split, which is what the challenge intends.

Dataset Details

Field	Value
Modality	CT (chest–abdomen, non-contrast trauma protocol)
Body part	Ribs / thoracic cage
Task	3D rib-fracture lesion segmentation (binary) + instance detection
Cases in this mirror	500 (train 420, val 80) of a 660-case release
Fracture instances	4,422 annotated (5,304 including the withheld test split)
Cohort	Huadong Hospital, Fudan University; Jan 2017 – Dec 2018; ethics NO.2019K146
Scanners	GE Revolution CT, Siemens Somatom Definition Flash
Format	.nii.gz (NIfTI), RibFrac<N>-image.nii.gz / RibFrac<N>-label.nii.gz
In-plane size	512 × 512 (all 500 cases)
Slices per case	74 – 721 (median 357)
Spacing	0.557–0.977 mm in-plane; slice 0.625 / 1.0 / 1.25 / 1.5 mm
License	CC BY-NC 4.0 (consistent across all sources — no discrepancy)
DOIs	10.5281/zenodo.3893508, .3893498, .3893496

Splits use contiguous, non-overlapping ID ranges — RibFrac1–RibFrac420 (train), RibFrac421–RibFrac500 (val), RibFrac501–RibFrac660 (test, not mirrored). public_id is therefore both the case ID and the split key.

Why binary, and what the -1 Ignore class does

The upstream release pairs each instance mask with a per-fracture CSV (public_id, label_id, label_code). The official label_code dictionary — as published on all three Zenodo records and hard-coded in the organizers' ribfrac/evaluation.py — is:

label_code	Meaning
0	Background (one bookkeeping row per case)
1	Displaced
2	Nondisplaced
3	Buckle
4	Segmental
-1	Ignore — genuinely a fracture, but its type could not be determined

58.2 % of all annotated fractures (2,575 / 4,422) are -1. A 4-class typed target would therefore discard the majority of the annotated lesions, turning them into unlabeled holes in the ground truth — which either poisons the loss or demands an ignore-mask the benchmark harness does not model. Binarizing uses all 4,422 instances and matches how the organizers themselves consume the masks (region detection at IoU ≥ 0.2).

⚠️ The -1 class is invisible in the NIfTI. There are zero negative voxel values anywhere in the 500 label volumes — untyped fractures are stored as ordinary positive integers, indistinguishable from typed ones. You can only identify them by joining to the CSV. A pipeline that reads only the mask and assumes every instance is typed will silently train on 58 % mislabeled data.

⚠️ The 2024 TMI challenge paper's Table I column headers are transposed. It prints BK ND DP SG while the counts underneath follow label_code order 1,2,3,4 — so codes 1 and 3 are swapped relative to the official dictionary. The 618 fractures it calls "buckle" are displaced. The openmedlab dataset card propagated this error. Trust the Zenodo dictionary and evaluation.py; the independent AUC23 re-release agrees with Zenodo.

Per-class counts, computed directly from the shipped CSVs (500 cases, 4,422 instances):

label_code	Class	train	val	total	share
1	Displaced	618	69	687	15.5 %
2	Nondisplaced	567	63	630	14.2 %
3	Buckle	291	30	321	7.3 %
4	Segmental	179	30	209	4.7 %
-1	Ignore	2,332	243	2,575	58.2 %

The official CSVs ship unmodified under info/ for anyone who wants the classification task; nothing in this mirror's segmentation target depends on them.

⚠️ RibFrac491 is RAS — the other 499 are LPS

This is the single easiest thing to get wrong with this dataset.

499 of 500 volumes are stored LPS. Exactly one — RibFrac491, in the validation split — is RAS. It is a clean ±1 flip on x and y, not an oblique acquisition, and the image and its label share the same direction, so per-case Dice stays correct even if you ignore it.

Any loader that calls get_fdata() / GetArrayFromImage() without consulting the affine gets that one case left–right and anterior–posterior mirrored relative to every other case.

Measured mitigation — RibFrac491 is also one of the 20 fracture-free controls. Its mask is entirely empty, so mirroring the label is a no-op and the effect on any overlap metric is nil. This is a coincidence of the release, not a design: the image is still mirrored, so canonicalize anyway. The case is unusable for anything with a left/right anatomical prior as-is, and nothing guarantees the withheld 160-case test split — or a future re-release — is equally lucky. Read the affine; do not lean on this accident.

Canonicalize from the affine — never hand-roll a flip:

import nibabel as nib
img = nib.as_closest_canonical(nib.load("images/RibFrac491-image.nii.gz"))  # -> RAS
# or, in a MONAI pipeline, the standard transform:
#   Orientationd(keys=["image", "label"], axcodes="RAS")

The per-case orientation column in the JSONL records which is which.

⚠️ Other verified gotchas

Measured on all 500 shipped label volumes, not taken from documentation:

• 20 validation cases have entirely empty masks. These are the deliberate fracture-free controls (val = 60 fracture-positive + 20 negative; test = 120 + 40). All 420 training cases have ≥ 1 fracture. Empty ground truth breaks naive Dice — decide explicitly whether a correct empty prediction scores 1 or is excluded. This is also why the FracNet paper's "60 tuning / 120 test" counts differ from the release's 80 / 160: the paper counts only positives.

  The full list, verified voxel-wise (note RibFrac491 among them — see the orientation warning above):

  RibFrac435 439 442 446 448 452 453 454 456 461
  RibFrac462 470 471 472 474 475 485 487 490 491
  

• Lesions are tiny. Measured across all 4,422 instances: median 3,243 voxels, p25 1,384, p90 8,930, min 78, max 125,740 — the median lesion is under 0.01 % of a 512×512×357 volume. Fractures per case range 0–36 (mean 8.84). A debug/smoke test that reads only the first N slices will very often see pure background. Do not read that as a broken loader.

• label_id = 0 is a per-case background bookkeeping row in the CSVs, present for all 500 cases, exactly once each. Filter it or you will count 500 phantom "fractures".

• Mixed dtypes, on both sides. Labels: uint8 (439), int16 (32), int32 (29) — do not assume uint8. Images: int32 (258), int16 (242), so a little over half the volumes are stored wider than the HU range needs. This mirror preserves the original dtypes rather than downcasting.

• Slice spacing is bimodal, ~50/50 between 1.0 mm (235 cases) and 1.25 mm (260 cases), plus 0.625 mm (4) and 1.5 mm (1). There is no single natural resample target.

• Instance integers are contiguous 1..N per case, N up to 36, and correspond 1:1 to the CSV's label_id for that public_id.

Ground truth — one gold tier, partly model-assisted

Five radiologists were involved. Two (3–5 yr and 10–20 yr experience) read each CT within 48 h and wrote reports; two more (5 yr each) drew the voxel masks in 3D Slicer 4.8.1 from those reports; a senior radiologist (20 yr) verified every mask. A second, human-in-the-loop pass then ran a FracNet-style model to propose additional candidates, and the senior radiologist adjudicated each one, adding those confirmed.

That second pass is material, not a footnote: the authors estimate ~20 % of fractures were missed by the initial manual reading. Only one mask per case ships, so there is no multi-rater tier and no rater ambiguity to resolve.

The organizers describe these masks as noisy and do not score Dice. The challenge metric is detection-oriented (a candidate counts as a hit at IoU ≥ 0.2), never a segmentation overlap score. Expect low absolute Dice from any model here; that reflects the task and the annotation protocol, not a defective mirror.

⚠️ Cross-dataset overlap — RibSeg is a 100 % patient overlap

RibSeg v1 and v2 are built on these exact CT volumes.

Dataset	Relationship
RibSeg v1 (Zenodo 5336592, CC BY-NC 4.0)	Rib + centerline labels for 490 of the 660 RibFrac CTs. Ships annotations only — the images must be obtained from RibFrac. Filenames are literally RibFrac31-rib-seg.nii.gz next to RibFrac31-image.nii.gz.
RibSeg v2 (IEEE TMI 2023, arXiv:2210.09309, CC BY-NC-ND 4.0)	All 660 RibFrac CTs, 15,466 ribs, reusing the identical 420/80/160 split.

Same patients, same scans, different label semantics (healthy rib anatomy vs. fracture lesions). Deduplication is trivial and exact: public_id (RibFrac<N>) is shared verbatim across RibFrac, RibSeg v1/v2, and the derived files in Zenodo 14864106. Never benchmark a RibFrac model against a RibSeg split as if they were independent cohorts.

No overlap with CTPelvic1K (Jishuitan, pelvis), LIDC-IDRI (TCIA), MSD, or TotalSegmentator (Basel routine CT) — different institutions, body regions and eras. MedOtter/totalsegmentator-ribs shares the word "rib" and nothing else: it is healthy rib bone segmentation on unrelated Basel patients.

The one derivative that is not a substitute: Zenodo 7969559 ("Rib CT fracture", AUC23, Radboud) ships center-cropped per-fracture patches for classification only, with no segmentation masks and restricted access.

Structure

images/RibFrac<N>-image.nii.gz    # 500 CT volumes   (N = 1..500)
labels/RibFrac<N>-label.nii.gz    # 500 instance masks, same grid as the image
info/ribfrac-train-info-1.csv     # official per-fracture CSVs, unmodified
info/ribfrac-train-info-2.csv
info/ribfrac-val-info.csv
train.jsonl                       # 420 rows of per-case metadata
val.jsonl                         # 80 rows
README.md
LICENSE.txt

Images and labels are flat because public_id is globally unique across splits; the split is recorded per row in the JSONL (and is recoverable from the ID range).

JSONL columns:

Column	Meaning
case_id / public_id	"RibFrac1" … "RibFrac500" — the RibSeg cross-reference key
image, mask	repo-relative paths
split	"train" or "val"
shape_zyx, spacing_xyz, origin_xyz	geometry
orientation	"LPS" or "RAS" — see the orientation warning
image_dtype, label_dtype	original dtypes (labels are mixed)
n_fractures	count of distinct instance integers > 0
instance_voxels	{instance_label: voxel_count}
total_fg_voxels	summed foreground voxels
is_fracture_free	true for the 20 empty-mask validation controls
grid_matches_label	true if image and label share size + affine (all 500)

Source & Citation

Zenodo, open access, no registration and no DUA:

• train Part 1 — https://doi.org/10.5281/zenodo.3893508
• train Part 2 — https://doi.org/10.5281/zenodo.3893498
• validation — https://doi.org/10.5281/zenodo.3893496
• test (images only, not mirrored) — https://doi.org/10.5281/zenodo.3993380

The Grand Challenge page (https://ribfrac.grand-challenge.org/) requires an account, but only for leaderboard submission — not for the data.

@article{jin2020ribfrac,
  author  = {Jin, Liang and Yang, Jiancheng and Kuang, Kaiming and Ni, Bingbing
             and Gao, Yiyi and Sun, Yingli and Gao, Pan and Ma, Weiling and
             Tan, Mingyu and Kang, Hui and Chen, Jiajun and Li, Ming},
  title   = {Deep-learning-assisted detection and segmentation of rib fractures
             from {CT} scans: Development and validation of {FracNet}},
  journal = {EBioMedicine},
  volume  = {62},
  pages   = {103106},
  year    = {2020},
  doi     = {10.1016/j.ebiom.2020.103106}
}

@article{yang2024ribfrac_challenge,
  author  = {Yang, Jiancheng and Shi, Rui and Jin, Liang and Huang, Xiaoyang and
             Kuang, Kaiming and Wei, Donglai and Gu, Shixuan and Liu, Jianying
             and Liu, Pengfei and Chai, Zhizhong and Xiao, Yongjie and Chen, Hao
             and Xu, Liming and Du, Bang and Yan, Xiangyi and Tang, Hao and
             Alessio, Adam and Holste, Gregory and Zhang, Jiapeng and
             Wang, Xiaoming and He, Jianye and Che, Lixuan and Pfister, Hanspeter
             and Li, Ming and Ni, Bingbing},
  title   = {Deep Rib Fracture Instance Segmentation and Classification from
             {CT} on the {RibFrac} Challenge},
  journal = {arXiv preprint arXiv:2402.09372},
  year    = {2024},
  doi     = {10.48550/arXiv.2402.09372}
}