RibAssist 3D: Biplanar Rib-Fracture Detection, Addressing, and Selective 3D Localization from CT-Derived Projections
Kabila Haile Soboka
cs.CV
2026-08-07
A staged diagnostic study of biplanar rib-fracture reconstruction from CT-derived projections finds triangulation exact (0.0 mm round trip) and conditional localization accurate (median 1.49 mm), but cross-view correspondence commits on only 15 of 601 sealed fractures at 93% rib-exactness. The bottleneck is detector confidence, not geometry or matching algorithms.
Rib fractures are common on chest CT and slow to localize, yet fracture counts feed directly into clinical risk scoring. Volumetric 3D detectors like FracNet need dense 3D annotation and heavy compute. A cheaper route: render the CT into two orthogonal projections (AP and lateral), detect fractures independently in each view, then pair detections across views and triangulate them back into 3D points. Everything rides on one sub-problem, cross-view correspondence: deciding which AP detection and which lateral detection are the same physical fracture, at a controlled rate of false 3D outputs. This single-author study from UT Austin refuses to report one end-to-end number; instead it runs a staged diagnostic, swapping each pipeline stage from oracle to real model to find exactly where the system fails.
| Stage | Result |
| Geometry (oracle) | 0.0 mm round-trip error; triangulation is sound |
| Oracle correspondence + real detections | median 4.0 mm, 88% within 10 mm, 93.6% rib-exact |
| Frozen detector + every matcher | 0% recall at budget; learned pair-scorer AUROC 0.57 |
| Retrained lateral head (dev) | 0% to 2.44% recall at budget |
| Sealed test (55 cases, 601 fractures) | 15 correct commitments (2.50%), 0.436 false points/case, median 1.49 mm, 93% rib-exact |
The detector-by-correspondence factorial shows the only cell that leaves zero is the retrained detector with deterministic confidence assignment; smarter matching algorithms move nothing. The root cause sits in the lateral heatmap: confidence is compressed just above the 0.05 extraction floor and roughly 188 spurious peaks per case flood the candidate graph. Retraining the lateral head lifts dual-view availability from 0.52 to 0.76, and only then does the system dare to commit.
For anyone building 2D-to-3D reconstruction, the finding is directly portable: the bottleneck is neither geometry nor localization but detector confidence failing to support cross-view correspondence, so invest in the weak view's detector, not in matching algorithms. Methodologically, the combination of staged diagnosis, a sealed cohort, a pre-registered fixed policy, and fail-closed provenance is rare in medical AI and far more credible than tune-until-presentable. Practically, abstention buys precision: 2.5% coverage with 93% rib-exact committed points positions this as an assistive reviewer where 2D detections and rib addressing stay visible even when 3D commitment abstains.
The author's own list is thorough: projections are simulated renderings of CT, not real biplanar radiographs, so scatter, tissue overlap, and calibration error are unmodeled; the sealed cohort is only 55 cases, so the 2.50% estimate carries a wide interval (0.69% to 4.48%); the cohort contains no fracture-negative studies, so false-output behavior on clean scans is unmeasured; the retrained head was selected on development data, so the sealed pass confirms direction but not an unbiased effect size. Two more from this reading: no clinician reader study exists, so workflow value is hypothetical, and as a single-author work it awaits independent replication.