Pediatric wrist fractures too subtle for one view: MV-CortexGuard pairs frontal and lateral evidence and lifts small-fracture AP by 6.6 points

The multi-view network turns subtle pediatric wrist fracture detection from saliency judgment on a single image into a check of cortical continuity along the bone axis, paired with candidate lesions matched between the frontal and lateral projections. Most earlier fracture detectors treated each radiograph as an independent input and improved overall performance by changing the detector, adding generic attention or adjusting multi-scale feature fusion, without systematically modeling the structure of subtle fractures or exploiting the complementary frontal and lateral views of the same examination. On the internal GRAZPEDWRI-DX test set, MV-CortexGuard reached 63.4% AP@50–95, 3.3 percentage points above the best comparator YOLOv10m, with a patient-level bootstrap 95% confidence interval of 2.1 to 4.6 points; on the size-defined small-fracture subset its average precision was 52.7%, 6.6 points above YOLOv10m.

The bone-axis-aligned cortical discontinuity representation first estimates wrist bone orientation from high-resolution features, rotates features into a near-vertical anatomical frame, then computes cortical differences along the bone axis, its normal and the local curvature direction. This representation, BACD, fixes the bone axis from the leading eigenvector of the weighted centroid and covariance matrix of a bone-response map, needs no manual bone-axis annotation, and uses a differentiable spatial transform to align features before mapping detections back to original image coordinates. Ablation shows that adding BACD to a baseline of a shared encoder and set-prediction head raised AP@50–95 from 58.4% to 60.7% and small-fracture average precision from 44.6% to 49.0%; removing directional differences from the full model lowered these by 1.9 and 3.8 points.

Anatomy-constrained multi-view query fusion exchanges frontal and lateral information at the candidate-query level and uses a learnable null query to represent the absence of reliable corresponding evidence in the other view. This module, ACMQF, combines candidate appearance similarity, normalized longitudinal position along the bone axis, osseous-region distribution and evidence reliability into a matching score, avoiding spurious correspondences from forced dense-feature pairing. Compared with standard cross-attention, ACMQF raised small-fracture average precision by 3.7 points and examination-level AUROC from 0.961 to 0.969; for fractures visible in only one view, the full model with the null query reached 88.9% sensitivity, 5.6 points above the model without it.

Multi-source uncertainty and risk-controlled rejection combine epistemic and aleatoric uncertainty, frontal-lateral disagreement, localization instability and image-quality reliability into an examination-level risk score, and examinations above threshold are referred for expert review instead of receiving an automated decision. This module, MURR, estimates predictive disagreement with four detection heads that share the backbone but keep independent low-rank adapter parameters, and sets probability and risk-acceptance thresholds on the held-out validation set. Adding MURR raised AP@50–95 and small-fracture average precision by only 0.4 points each, but cut expected calibration error from 0.047 to 0.024 and the area under the risk–coverage curve from 0.0196 to 0.0107; at the validation-selected operating point targeting a 5% selective error rate, the model accepted 94.3% of internal test examinations.

The results rest on the public GRAZPEDWRI-DX dataset with a patient-level 6:2:2 split, so training, validation and internal test all come from one data source; model selection, probability calibration and rejection thresholds were set on the held-out validation set, and the test set was used only for final evaluation. The small-fracture subset is defined by the 25th percentile of normalized fracture-box area in the training set, with the 0.0016 threshold applied unchanged to validation and test. The authors propose multicenter external validation and reader studies of risk-controlled computer-aided fracture diagnosis as the next step.

The authors note that development and evaluation used a single public dataset, so generalizability across institutions, equipment, acquisition protocols and fracture prevalence is not established; the size-defined small-fracture subset is only a proxy for perceptual subtlety and may miss fractures that are hard because of low contrast, overlap or developmental anatomy, and its threshold may be dataset dependent. The validation set served for model selection, calibration and threshold selection, with no separate external calibration cohort; missing projections and image degradations were simulated, and oblique or other non-standard projections remain untested; subgroup analyses covered only age and sex, and the confidence interval was wider in children under 8. The study had no prospective reader study, workflow integration or patient-outcome evaluation, and the rejection operating points assume deferred examinations receive appropriate expert review.