Improving Recognition Performance and Reader Efficiency for Femoral-Neck Fracture Detection on Pelvic or Hip...

Background Radiographs sometimes do not depict femoral-neck fractures, particularly radiograph-negative or indeterminate femoral-neck fractures, leading to delayed treatment and complications. Purpose To develop and externally evaluate a deep learning model, OccuNet, for detecting femoral-neck fractures on pelvic or hip radiographs, compare its performance and accuracy with those of radiologists and emergency medicine physicians, and evaluate its effect on reader performance and accuracy. Materials and Methods This multicenter retrospective study included adults suspected of having hip trauma who underwent pelvic or hip radiography and same-episode CT or MRI at four hospitals (January 2009-August 2025). Patients were split into a training set, an internal test set, and three external test sets. A two-stage model, OccuNet, was trained: Stage 1 used contrastive pretraining on paired original and artifact-augmented versions of the same radiograph to learn artifact-robust features, and stage 2 fine-tuned a fracture detector. Performance and accuracy were compared with those of 10 readers (five musculoskeletal radiologists, five emergency medicine physicians), and artificial intelligence (AI) assistance was tested for sensitivity, specificity, area under the receiver operating characteristic curve (AUC), and reading time. Results This study included 2576 patients (mean age [±SD], 69 years ± 11.2; 1380 women). In the pooled group (n = 1766), the model achieved a sensitivity of 97.5% (913 of 936), specificity of 98.8% (820 of 830), and excellent performance (AUC, 0.99 [95% CI: 0.99, 0.99]). For radiograph-negative or indeterminate fractures (n = 189), sensitivity of the model (94.7% [179 of 189]) was higher than that of radiologists (86.2% [163 of 189]; P < .001) and emergency medicine physicians (68.8% [130 of 189]; P < .001). With AI assistance, radiologist and emergency medicine physician sensitivities increased from 93.7% (877 of 936) to 97.2% (910 of 936; P < .001) and from 84.3% (789 of 936) to 95.6% (895 of 936; P < .001), respectively, and mean reading times shortened by 14.9% and 18.9% (both P < .001), respectively. Conclusion OccuNet demonstrated excellent performance for femoral-neck fracture detection on pelvic or hip radiographs, maintained high sensitivity in radiograph-negative or indeterminate (Garden I-II) fractures, and improved reader sensitivity and efficiency. © RSNA, 2026 Supplemental material is available for this article. See also the editorial by Flores and Cantarelli in this issue.