RibAssist 3D: Biplanar Rib-Fracture Detection, Addressing, and Selective 3D Localization from CT-Derived Projections

• Projection-based reconstruction offers a lower-dimensional alternative to dense volumetric detection, but depends on pairing AP and lateral detections of the same fracture.The study uses simulated biplanar projections derived from CT datasets rather than claiming a deployable detector.Source: §1 · p. 1
• The study decomposes the pipeline into geometry, localization, and correspondence to isolate the stage responsible for failures.Source: §1 · p. 1
• The tested correspondence methods are compared through controlled experiments, with lateral-detector quality identified as the operational lever.Source: §1 · p. 1
• A pre-specified confirmation on an untouched sealed cohort reproduces the development estimate out of sample.Source: §1 · p. 2
• Even when 3D correspondence is not committed, per-view detections highlight suspicious regions and support rib-level addressing.Source: §1 · p. 2

Deep rib-fracture detection on CT

• FracNet detects and segments rib fractures directly in CT volumes, whereas this work studies a cheaper 2D projection route with budget-aware evaluation.Source: §2 · p. 2

Rib labeling and anatomical centerlines

• RibSeg v2 supplies rib labels and centerlines that this study uses for addressing and for checking reconstructed-point rib exactness.Source: §2 · p. 2

Biplanar-to-3D reconstruction

• Unlike X2CT-GAN’s dense generative volume reconstruction, this method triangulates a sparse set of fracture points and explicitly decides correspondence.Source: §2 · p. 2

Cross-view correspondence and selective prediction

• Shared superior-inferior-axis agreement simplifies candidate generation but is weaker than full epipolar geometry because many fractures share an axis coordinate.Source: §2 · p. 2
• Assignment with abstention applies selective prediction at correspondence commitments, trading coverage for reliability under a controlled false-output budget.Source: §2 · p. 2

Data and projection

• CT volumes and annotations are rendered into 256 × 256 AP and lateral orthographic projections sharing the superior-inferior axis.Source: §3.1 · p. 2

Data splits

• Patient-disjoint splits prevent scans and derived projections, peaks, or crops from crossing development and test boundaries.Source: §3.1 · p. 2

End-to-end pipeline

• Per-view U-Nets produce heatmaps, NMS extracts peaks, SI agreement gates candidate pairs, and abstaining assignment commits only pairs clearing a threshold for triangulation.Source: §3.2 · p. 2, Figure 1 · p. 3

Networks

• The architecture uses independent AP and lateral U-Nets, geometric SI-axis fusion, an addressing network, and a learned pair-scorer.Source: Figure 2 · p. 3, §3.3 · pp. 3–4

Operational endpoint

• The primary endpoint measures recall within 10 mm while allowing at most one false 3D point per case.Source: §3.4 · p. 4

Staged diagnosis

• A diagnostic elimination swaps oracle and model components across geometry, localization, correspondence, calibration, and retraining stages to charge failures to individual steps.Source: §3.5 · p. 4

Experimental setup

• The deterministic correspondence frontier jumps from accepting nothing to many false points, leaving no controlled-budget operating point.Source: Table 2 · p. 4

Geometry and localization

• Median 4.0 mm localization error with oracle correspondence, 88% within 10 mm, and 93.6% rib-exact show localization is tolerable rather than dominant.Triangulation from correct paired ground-truth centers is exact, with 0.0 mm error and 98.6% rib-exactness.Source: §4.1 · p. 4

Frozen-detector correspondence

• Only 4 of 205 dual-view fractures were correctly and uniquely paired by the frozen SI-only matcher, producing about 0.6% end-to-end within-10 mm performance.The failure occurs in obtaining correct pairs; when correspondence is correct, reconstruction remains accurate.Source: §4.2 · p. 4

Candidate-field calibration

• 68% of spurious lateral peaks occupy [0.05, 0.055), but extraction-policy recalibration leaves controlled-budget recall at zero despite shrinking the candidate field.Candidate-field density alone is therefore insufficient; the tested frontier depends on confidence in true correspondence edges.Source: §4.3 · pp. 4–5, Table 3 · p. 5

Detector intervention

• Lateral retraining raises amplitude ceiling from 0.091 to 0.155 and dual-view availability from 0.52 to 0.76, moving development recall at one false point per case from 0% to 2.44%.In the detector-by-correspondence factorial, only the retrained detector with deterministic assignment moves off zero; learned appearance scoring does not beat detector confidence.Source: §4.4 · p. 5, Table 4 · p. 5, Figure 4 · p. 6

Sealed-test confirmation

• 15 of 601 fractures receive correct sealed-cohort 3D localizations at 0.436 false points per case, a 2.50% end-to-end commitment yield.Committed points have median error 1.49 mm and 93% rib-exactness; the frozen policy commits 0 of 601.Source: §4.5 · p. 6, Table 5 · p. 6, Figure 5 · p. 7

Impact

• The workflow retains per-view highlighting and rib-level addressing when confidence-gated 3D correspondence abstains, making 3D localization an additive output.The system is presented as an assistive, secondary-review workflow rather than an autonomous diagnostic system, and clinical benefit was not measured.Source: §4.6 · p. 7, §4.7 · p. 7

Simulated inputs

• Real biplanar radiography is not evaluated because AP and lateral inputs are deterministic orthographic CT renderings.Scatter, tissue overlap, and acquisition calibration error are not modeled.Source: §5 · p. 7, Figure 6 · p. 8

Small cohort

• At n=55, the sealed-cohort confidence interval is wide, making the estimate directional rather than precise.Source: §5 · p. 8

Safety validation

• Negative-scan safety remains unvalidated because the sealed cohort contains no confirmed fracture-negative studies.False-output behavior on truly negative scans is therefore unmeasured.Source: §5 · p. 8

Experimental interpretation

• Development selection and joint intervention effects limit causal interpretation of the retraining gain.The selected head was chosen on development data, while retraining changed availability, prevalence, confidence, and candidate ceiling together.Source: §5 · p. 8

Metric scope

• Recall at 10 mm with independent matching does not assess fracture typing, displacement, or acute-versus-healed status.Source: §5 · p. 8

Research status

• The study uses public, de-identified RibFrac and RibSeg v2 datasets under their licenses without patient interaction or prospective decisions.Source: §6 · p. 8

Intended use

• The prototype must not be used for diagnosis or patient care.Source: §6 · p. 8

Validation boundary

• The clinician-review interface demonstrates human-in-the-loop review but has not undergone usability or clinical validation.Source: §6 · p. 8

Generalization and equity

• Performance on independently acquired radiographs and demographic subgroup or fairness effects are not claimed or analyzed.The available metadata and evaluation did not support subgroup analysis.Source: §6 · p. 8

Main finding

• Projection geometry is exact and detector localization is rib-accurate, but weak lateral detections limit confident cross-view commitments.Source: §7 · p. 8

Sealed confirmation

• 2.50% correct 3D commitment yield was confirmed on an untouched 55-case cohort at 0.436 false 3D points per case.Source: §7 · p. 8

Intervention

• Lateral retraining moved the development frontier from 0% to 2.44% under the controlled-budget evaluation.Source: §7 · p. 8

Future work

• Future work should prioritize lateral-detector quality while expanding correspondence methods, real biplanar inputs, and 3D confirmation of accepted points.Source: §7 · pp. 8–9

Selective assistance

• The framework preserves detections and rib information while adding accurate 3D localization only when confidence is sufficient.Source: §7 · p. 9

• Source code, split identifiers, frozen policies, evaluation scripts, and reproducibility instructions are available in the RibAssist 3D repository.Source: Code Availability · p. 9
• Third-party datasets and trained checkpoints are not redistributed because of licensing and artifact-size constraints.Source: Code Availability · p. 9