Automated systems now find rib fractures on CT about as well as senior radiologists. They still cannot reliably say which kind of fracture they found: a buckle (a subtle bulge in the bone's outer shell), a clean crack that has not moved, a displaced break, or a segmental injury with two breaks on the same rib. Surgeons need exactly that distinction. The problem gets worse at a new hospital, because routine CT there is often reconstructed into thick slices (3–5 mm instead of under 1 mm) and with different sharpening filters. This proposal starts from a physical observation. Each fracture subtype is defined by a small set of geometric facts, such as whether the bone's outer shell is continuous, how many breaks there are along the bone, and how far the pieces have moved. Each of those facts has its own smallest size below which a given scan cannot resolve it. Thick slices do not blur every subtype equally. They erase the sub-millimeter facts and leave the millimeter-scale ones intact.
We propose a classifier that measures these geometric facts explicitly, in a 3D coordinate frame that runs along each bone. It is paired with a measurability rule that is fixed before deployment. The rule takes the scan's slice thickness, its reconstruction filter, and the local angle of the bone to the scan axis, and decides which facts can be trusted for this particular fracture. When a fact cannot be trusted, the model does not guess. It reports a coarser label one level up the subtype hierarchy, for example "buckle-or-non-displaced" instead of a coin flip between the two. The trained model stays frozen across hospitals; only scan-header metadata changes its behavior. This fits the locked-model setting that nearly all cleared clinical AI operates under.
Core Claim
Morphological fracture labels are combinations of geometric attributes, and these attributes have very different resolution floors. Cross-protocol failure in fracture subtyping is therefore mostly a predictable loss of specific attributes, not a generic distribution shift. Three predictions follow. First, an end-to-end 3D classifier trained on thin slices should lose accuracy mainly on the subtypes whose attributes lie below the new slice thickness. Thick-slice augmentation cannot fix this, because the information is physically absent. Second, a classifier that measures the attributes and coarsens its output when they become unmeasurable should keep high accuracy on the labels it does emit. Its errors would become honest abstentions rather than confident mistakes. Third, the measurability decision depends on 3D geometry. A rib fracture whose bone axis runs nearly parallel to the scan axis loses through-plane detail far faster than one lying in-plane. Any approach that processes slices or patches without the bone's 3D orientation cannot represent this.
The spine makes the argument concrete. Genant grading assigns grades by vertebral height loss: under 20% is normal, 20–25% is grade 1, 25–40% grade 2, and over 40% grade 3. These heights are measured roughly along the scanner's slice axis. The grade-1 band spans only five percentage points of a vertebral body a few centimeters tall, which is on the order of a millimeter or two. A 3–5 mm slice cannot place a vertebral endplate that precisely. So grade 0 versus grade 1 should be unrecoverable from thick-slice spine CT, while grade 2 versus grade 3 should largely survive. We will test this prediction directly.
Relation to Prior Work
Three papers motivate the gap.
- Deep Rib Fracture Instance Segmentation and Classification from CT on the RibFrac Challenge defines the four-class rib subtyping task and builds FracNet+, which adds global skeleton context through a point-cloud branch. Its authors report that subtyping remains unsolved, especially for buckle and segmental fractures. They do not test protocol shift.
- Video-CT MAE shows that video-pretrained transformers can process CT volumes natively. It addresses only binary fracture detection at a single center.
- Regulating AI Adaptation shows that 98.6% of FDA-cleared devices are locked. It also shows that site-local retraining causes forgetting on the original site, costing up to 0.268 AUC on chest radiographs. This constrains us to a frozen model.
Our measurement pipeline borrows from Interpretable Vertebral Fracture Quantification via Anchor-Free Landmarks Localization. That paper straightens the spine along its centerline and grades fractures from six height landmarks. It generalizes zero-shot to VerSe, but it flattens the volume to a single sagittal image and never studies slice thickness. We keep the straightening idea but use it to build a tube-shaped 3D reformat that preserves each bone's full cross-section. We extend the same idea from vertebrae to ribs.
Two further papers constrain the design. Curriculum learning for improved femur fracture classification reports expert agreement around 71% on AO subtypes. That level of label noise means the coarsened labels must be evaluated against reader disagreement, not treated as pure loss. Enhancing Super-Resolution Networks through Realistic Thick-Slice CT Simulation supplies the method we reuse for simulating realistic thick-slice reconstructions from thin ones.
Research Program
Stage 1: bone-aligned 3D frames. For ribs, we take rib centerlines from an existing segmentation (a STUNet-style network pretrained on TotalSegmentator, as in FracNet+). Around each centerline we resample a straightened tube, roughly 2 cm wide. For vertebrae, we straighten the spine following Chen et al., but keep a 3D slab rather than a single mid-sagittal plane. Every voxel in the tube records how far it lies from native acquired samples along the original slice axis. This makes the information the scan actually contains explicit.
Stage 2: attribute heads. A small 3D network on the tube predicts the attributes that define each class:
- cortical continuity at the fracture site;
- the number of distinct breaks along the bone;
- displacement of the bone pieces relative to each other, in millimeters, measured perpendicular to the centerline;
- for vertebrae, the six Genant height landmarks.
The subtype is then a fixed, auditable rule over these attributes. Segmental means two or more breaks. Displaced means displacement above a threshold. Non-displaced means a break with no displacement. Buckle means a cortical bulge with no break. RibFrac does not label attributes directly. Break count and displacement can be derived from its voxel-level fracture masks, and cortical continuity is supervised weakly through the subtype labels. Genant heights follow the existing landmark supervision.
Stage 3: a frozen measurability rule. On thin-slice source data, we simulate each target protocol: slice thicknesses of 1, 2, 3 and 5 mm, and smooth versus sharp reconstruction filters. For each attribute $k$ we fit its measurement error standard deviation $\sigma_k(t,\theta)$ as a function of slice thickness $t$ and the angle $\theta$ between the local bone axis and the slice axis. At test time, attribute $k$ is trusted for a given fracture only if its estimate $\hat a_k$ lies clearly on one side of its class threshold $\tau_k$:
$$|\hat a_k - \tau_k| > 2\,\sigma_k(t,\theta).$$
The factor 2 gives roughly 95% one-sided confidence under Gaussian error and will be checked empirically. A class decision that relies on an untrusted attribute is merged with its neighbor in the hierarchy. The error table is fit once on source data and then locked. At a new site the model reads only header fields, so no retraining or local labels are required.
Baselines
The naive reference is a 3D patch classifier trained on native-resolution RibFrac data and applied to thick-slice inputs after trilinear resampling, with no special handling. It measures how much subtyping accuracy protocol shift destroys. Anything we build must beat it on the labels it emits.
The strongest existing method is FracNet+ from Yang et al. We will retrain it with its point-cloud context branch and TotalSegmentator pretraining, using the same RibFrac training split (420 train, 80 validation, 160 test). It reports 84.29% test detection sensitivity on RibFrac, against about 76% for a senior radiologist. We found no reported per-class subtyping macro-F1 for RibFrac on Lacuna. We will therefore reproduce FracNet+'s classification score ourselves at native resolution before any shift.
The strongest competitor on our specific question is FracNet+ (or a Video-CT MAE backbone) trained with thick-slice augmentation from the same simulator and the same compute and tuning budget. It sees exactly the degraded data our model sees. If it matches us, the explicit geometry adds nothing.
For the spine, the reference is the Chen et al. landmark pipeline. It reports patient-level zero-shot AUC 0.95 (sensitivity 0.85, specificity 0.90) on VerSe, and we will rerun it under each simulated slice thickness.
Two ablations isolate the ingredients. The first keeps the attribute heads but removes the measurability rule, so the model always outputs a full four-class label. The second keeps the coarsening idea but replaces the geometric rule with a learned confidence threshold on the end-to-end classifier.
Beating the baselines means two things. First, on held-out protocols, the full four-class accuracy restricted to cases where we do not coarsen exceeds the augmented FracNet+ on the same cases. Second, across the whole test set, the curve of error against coverage lies below every baseline's curve, with the comparison made at matched coverage.
Why not just…?
Why not just augment with thick slices and let a 3D network learn robustness? This is the most serious threat. If buckle fractures leave some faint residual signal at 3 mm, augmentation could recover it, and our coarsening would discard information unnecessarily. The augmented-FracNet+ baseline settles this. If its accuracy on buckle versus non-displaced at 3–5 mm stays near its thin-slice level, our physical premise is wrong for that attribute, and we will report that. The prediction we stake the proposal on is that augmentation collapses buckle and grade-1 discrimination toward chance while preserving displaced versus segmental.
Why not just super-resolve thick slices back to thin ones and run the existing model? Super-resolution networks such as the one trained with the simulator of Enhancing Super-Resolution Networks through Realistic Thick-Slice CT Simulation produce plausible thin slices. They cannot create a sub-millimeter cortical bulge that was never sampled; they can only invent one. We will run this pipeline as an additional baseline. We expect it to produce confidently wrong buckle calls, which the error–coverage curve will expose as a higher error at full coverage than our coarsened output.
Why not just abstain using the end-to-end model's own uncertainty, for example Monte Carlo dropout? This is the second real threat, and it is the second ablation. Learned uncertainty is calibrated on the training protocols and tends to be overconfident on unseen ones. It also cannot tell the clinician which distinction became unmeasurable. The decisive comparison is coverage-matched error on unseen thickness and filter combinations, together with whether the abstentions fall on the hierarchy pairs physics predicts.
Why not just fine-tune at each hospital? Wu et al. show that local retraining costs up to 0.268 AUC on the original site, and that cleared devices are almost always locked. Subtype re-annotation is also expensive and noisy, given the roughly 71% inter-expert agreement reported by Jiménez-Sánchez et al. We include per-site fine-tuning only as an upper bound on target-site accuracy, reporting its source-site forgetting alongside.
Proposed Experiments and Decisive Checks
Rib subtyping under simulated shift. We degrade RibFrac test scans to 2, 3 and 5 mm slices and to smooth or sharp filters. The model sees only some of these combinations during training and is tested on the rest. We report per-class F1, coarsened-label accuracy, and error–coverage curves. The first decisive check is whether the accuracy loss concentrates in buckle and non-displaced fractures, and whether it grows with the bone-to-axis angle $\theta$ as the measurability model predicts.
Genant grading. On A Vertebral Segmentation Dataset with Fracture Grading (VerSe: 160 series, 289 graded fractures, at most 1 mm resolution), we degrade along the slice axis and test the prediction that grade 0 versus 1 falls to chance by about 3 mm while grade 2 versus 3 holds. The second decisive check is whether the measured cross-over thickness matches the threshold-plus-two-sigma rule derived from source data alone.
Real multi-center geometry. Simulation could misrepresent real reconstructions. As a check of the fragment-separation attribute on genuine cross-center data, we will evaluate break and fragment counting on the six-center CT set of the PENGWIN pelvic fracture benchmark, whose winning entry reports IoU-F 0.930. We will compare our attribute error by center. Our labeled data contain no real thick-slice scans, so the thick-slice claims rest on simulation fidelity, and we will state this limitation plainly.
Clinical value of coarsened labels. We check whether the coarsened outputs agree with what radiologists actually distinguish. On the RibFrac cases with multiple annotations, we test whether the pairs the model merges are the same pairs on which annotators disagree.
References
- Deep Rib Fracture Instance Segmentation and Classification From CT on the RibFrac Challenge
- Video-CT MAE: Self-supervised Video-CT Domain Adaptation for Vertebral Fracture Diagnosis
- Regulating AI Adaptation: An Analysis of AI Medical Device Updates
- Interpretable vertebral fracture quantification via anchor-free landmarks localization
- Curriculum learning for improved femur fracture classification: Scheduling data with prior knowledge and uncertainty
- Enhancing Super-Resolution Networks through Realistic Thick-Slice CT Simulation
- A Vertebral Segmentation Dataset with Fracture Grading
- Benchmark of Segmentation Techniques for Pelvic Fracture in CT and X-Ray: Summary of the PENGWIN 2024 Challenge