Atlas-guided attribution ranking reveals anatomy-dependent localization gaps in C2 fracture classifiers
Year: 2026
Type: article
Abstract: Binary C2 fracture classifiers can identify whether a fracture is present but do not show which part of C2 is fractured. We evaluated a framework that combines detailed C2-specific multi-atlas anatomy with classifier attribution maps to produce regional rankings, without requiring additional model training. Three CT examinations from the RSNA 2022 Cervical Spine Challenge were manually annotated into 16 C2 subregions and used as anatomical atlases. Deformable registration transferred these labels to individual patients. The resulting anatomical maps were combined with Grad-CAM maps from two existing classifiers and their ensemble using Peak-Guided Iterative Region Proposal. Primary localization analysis included 97 examinations classified as positive by at least one classifier; secondary detection-to-localization analysis included all 101 confirmed C2 fracture cases. Three-atlas leave-one-out evaluation of label propagation yielded macro-Dice scores of 0.88 for five anatomical groups and 0.77 for 16 subregions, with weaker agreement for small articular facets. Ensemble mean per-case F1@3 was 0.66 and 0.54, respectively. Ensemble detection-to-localization micro-Recall@3 was 0.83, compared with 0.86 among the 97 detector-positive cases. Odontoid and body–lateral mass–superior articular facet complex fractures were retrieved most reliably; transverse-process, posterior-arch, and pars-inferior-facet retrieval was reduced or variable. This framework converts patient-level fracture predictions into structured anatomical rankings and reveals region-specific weaknesses that are not apparent from patient-level detection performance. These findings suggest opportunities for targeted dataset enrichment and future model refinement.
Source: Scientific Reports
Language: English
Topic: Medical Imaging and Analysis
Subfield: Biomedical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7226003701
