Answers are grounded in this paper and cite their sources.
Retrospective study demonstrates high hip fracture detection accuracy in pelvic radiographs, suggesting conditional support for radiology residents.
This study aims to externally validate a deep learning model for hip fracture detection on pelvic radiographs and evaluate its association with radiology resident performance. This retrospective study included 600 frontal pelvic radiographs from a tertiary academic centre with ground truth established by musculoskeletal radiologist consensus. A previously developed DenseNet-based model underwent standalone validation. Ten first-year radiology residents interpreted 100 radiographs each (50 fracture-positive, 50 fracture-negative, with cases drawn from the validated dataset and reused across packets), without and with AI assistance after washout. Paired McNemar tests compared unaided versus AI-assisted performance. Generalized linear mixed-effects models with crossed random intercepts assessed the conditional association between AI assistance and positive fracture calls, accounting for repeated case exposure and inter-reader variability. The external validation dataset contained 124 fractures (20.7% prevalence). Standalone AI achieved AUROC 0.959 (95% CI: 0.933–0.979), sensitivity 92.1% (86.9–96.3%), and specificity 85.2% (82.1–88.2%). Aggregate reader sensitivity was 93.1% without AI versus 93.9% with AI (difference + 0.8%, p = 0.45). Specificity was 96.5% without AI versus 97.6% with AI (difference + 1.1%, p = 0.18). Mixed-effects modelling revealed significant AI-fracture interaction (OR 47.8, 95% CI: 4.9–469, p < 0.001), suggesting AI assistance was associated with higher odds of a positive call when a fracture was present, without a corresponding increase in false positives. The wide confidence interval indicates substantial imprecision. The model showed a modest decline in overall discrimination on external validation but remained comparable to published benchmarks. Mixed-effects modeling identified a conditional association between AI assistance and correct fracture calls when fractures were present. However, aggregate paired analyses were non-significant, and these findings require confirmation in larger, prospectively designed studies.
No takes yet. Share an insight, caveat, or question.
Tan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: