MV-CortexGuard: detecting subtle pediatric wrist fractures through cortical structure representation, anatomy-constrained multi-view consistency, and risk- controlled abstention
Year: 2026
Type: preprint
Abstract: Background Subtle pediatric wrist fractures may appear on radiographs only as focal cortical interruption, mild cortical bulging, or a fine linear lucency. Their detection is complicated by the epiphysis, growth plate, overlapping bones, and variations in projection. Most existing fracture detectors process each image independently, do not jointly model directional cortical continuity and complementary evidence from AP/PA (frontal) and lateral views, and cannot reliably identify high-confidence errors. Methods We developed MV-CortexGuard, a multi-view, risk-aware network for detecting subtle pediatric wrist fractures. Its bone-axis-aligned cortical discontinuity representation module estimates the wrist bone axis, maps features into normalized anatomical coordinates, and extracts cortical differences along the bone axis, its normal, and local curvature directions. Its anatomy-constrained multi-view query fusion module establishes cross-view correspondences using candidate appearance, normalized longitudinal location, osseous-region distribution, and evidence reliability, while a learnable null query preserves view-specific evidence. A multi-source uncertainty and risk-controlled rejection module integrates epistemic and aleatoric uncertainty, cross-view disagreement, localization instability, and image-quality reliability into an examination-level risk score and refers high-risk examinations for expert review. Results On GRAZPEDWRI-DX, MV-CortexGuard achieved an average precision (AP) of 63.4% at intersection-over-union thresholds of 0.50–0.95 and an AP of 52.7% in the size-defined small-fracture subset, exceeding YOLOv10m by 3.3 and 6.6 percentage points, respectively. The examination-level area under the receiver operating characteristic curve (AUROC) was 0.969. The uncertainty module reduced the expected calibration error to 0.024 and the area under the risk–coverage curve to 0.0107. At the validation-selected operating point targeting a selective error rate of 5%, the model accepted 94.3% of internal test examinations for automated classification. Conclusions MV-CortexGuard improved detection accuracy, cross-view diagnostic consistency, and predictive reliability for subtle pediatric wrist fractures. The results support further multicenter external validation and reader studies of risk-controlled computer-aided fracture diagnosis.
Source: Research Square
Institution: Yuhuangding Hospital
Language: English
Topic: Medical Imaging and Analysis
Subfield: Biomedical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225630073