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

Technical Summary: RibAssist 3D

Problem Statement
Rib fractures are clinically significant indicators of thoracic trauma severity, yet localizing them on CT volumes is tedious and error-prone, particularly for non-displaced fractures. While automated 3D detection exists, full volumetric networks are computationally heavy and require dense 3D annotation. This paper investigates a projection-based alternative: rendering CT volumes into two orthogonal orthographic views (anteroposterior [AP] and lateral), detecting fractures independently in each, and triangulating matched detections into 3D points. The core challenge is not the geometry of triangulation, but solving the cross-view correspondence problem—deciding which AP detection and which lateral detection correspond to the same physical fracture—at a strictly controlled rate of false 3D outputs. The study aims to determine if such a system can provide reliable, selective 3D localization without emitting spurious points, functioning as an assistive workflow rather than a fully automated reconstructor.

Methodology
The study employs a staged diagnostic design using a sealed cohort to isolate failure modes within the pipeline. The system consists of four main components:

1. Data & Projections: CT volumes from the RibFrac and RibSeg datasets are rendered into 256×256 AP and lateral orthographic projections. The data is split into training, diagnostic (held-out for selection), and a sealed test cohort (55 cases, 601 fractures).
2. Detection: Two independent U-Net detectors (one per view) generate fracture heatmaps. Peaks are extracted via non-maximum suppression (NMS) with view-specific thresholds.
3. Correspondence & Triangulation: Candidates are paired based on agreement in the shared superior-inferior (SI) axis. A one-to-one assignment with abstention is applied: a pair is only committed to 3D triangulation if its matching cost (based on detector confidence or a learned scorer) clears a strict threshold. Committed pairs are back-projected to 3D world coordinates.
4. Addressing: A separate dual-stream CNN predicts the rib level and side for each detection to support anatomical organization.

The evaluation metric is recall at 10 mm (distance to ground truth) under a controlled budget of at most one false 3D point per case. The study systematically swaps components (e.g., oracle vs. model, frozen vs. retrained detectors, deterministic vs. learned scorers) to identify the operational bottleneck.

Key Contributions

1. Pipeline Decomposition: The paper isolates the biplanar reconstruction pipeline into geometry, localization, and correspondence, demonstrating that geometry is exact and localization is accurate, while correspondence is the dominant failure point.
2. Identification of the Bottleneck: Through staged elimination, the study proves that the operational failure is not due to triangulation geometry or the correspondence algorithm itself, but rather low confidence in the lateral detector. The frozen detector emits too many low-confidence lateral peaks, preventing the formation of correct pairs that clear the abstention threshold.
3. Lateral Retraining Intervention: Retraining only the lateral detector head (with hard-negative mining) significantly improves detector confidence and dual-view availability. This intervention is the primary driver of the first nonzero controlled-budget results.
4. Sealed-Cohort Validation: The study provides a single, pre-specified pass on an untouched sealed cohort, confirming the development estimates with fail-closed provenance (SHA-256 verification of inputs and policies).

Results

• Geometry & Localization: Back-projection of correct paired centers yields 0.0 mm round-trip error. With correct correspondence, detector peaks reconstruct to a median 3D error of 4.0 mm (88% within 10 mm). When the policy successfully commits a correct pair, the median error drops to 1.49 mm with 93% rib-exactness.
• The Correspondence Failure: On the frozen detector, deterministic and learned pair-scoring methods achieve 0% recall at the controlled budget (≤1 false/case). The candidate graph is dominated by non-fracture pairs, and the lateral detector's confidence is insufficient to distinguish true matches.
• The Intervention: Retraining the lateral head (Stage L2) raises dual-view availability from 0.52 to 0.76 and lifts the operational frontier.
• Final Performance: On the sealed cohort (55 cases, 601 fractures), the retrained system with deterministic assignment achieves a 2.50% end-to-end commitment yield (15 correct 3D localizations) at 0.436 false 3D points per case. The 95% case-bootstrap confidence interval is [0.69%, 4.48%], excluding zero. In contrast, the original frozen detector's selected policy yields 0% (0/601).
• Learned Scorers: A learned pair-scorer did not outperform the detector's own geometric-mean confidence at the operational budget; in fact, it performed worse on the retrained head.

Significance and Claims
The paper positions RibAssist 3D not as a deployable clinical diagnostic system, but as a diagnostic study and an assistive workflow prototype. Its primary significance lies in:

• Establishing Feasibility: Proving that confidence-gated, selective 3D localization from CT-derived projections is possible, provided the detector confidence is high enough to clear strict safety thresholds.
• Defining the Bottleneck: Explicitly identifying that for this specific setup, cross-view correspondence is the operational bottleneck, driven by lateral-detector quality, not by the matching algorithm or geometric constraints.
• Assistive Value: Demonstrating a layered assistance model where the system highlights suspicious regions and assigns rib levels even when 3D correspondence is abstained. The system prioritizes safety (omission) over coverage, ensuring that any emitted 3D point is geometrically accurate and highly confident, while uncertain findings remain visible as 2D detections for human review.

The author concludes that while the end-to-end yield is low (2.5%), the system successfully demonstrates a framework where reliable 3D guidance is provided without unsupported predictions, highlighting the need for improved detector quality and global correspondence strategies for broader coverage.