Abstract Background/Aims Osteoporotic vertebral fractures occur in up to 25% of at-risk individuals and are associated with a 1.2-fold increase in age-adjusted mortality. Diagnosis rates remain low, with many cases undetected during routine care. This study aimed to develop and evaluate AI-based models to identify vertebral fractures on routinely acquired lumbar spine DEXA scans. Methods We analysed lumbar spine DEXA images with corresponding demographic and clinical data from a regional scanner in northwest England. Each scan included vertebrae L1-L4 and adjacent soft tissues. Patient data included prior fragility fractures, their anatomical site of fracture, and FRAX comorbidities. A range of convolutional neural networks and vision transformer models were tested under four conditions: full images with augmentation, full images without augmentation, isolated vertebrae with augmentation, and isolated vertebrae without augmentation. Experiments also used either image normalisation, or unmodified images. Model performance was evaluated using the area under the curve (AUC), precision-recall AUC (PR-AUC), sensitivity, specificity, and predictive values. Results Out of 4,161 lumbar spine DEXA scans 667 patients reported vertebral fracture (16.0%). Patients with fractures were significantly older (68.3±11.3 vs 64.4±12.7 years, p 0.001), had higher rates of previous fractures (47.1% vs 19.1%, p 0.001), and lower bone mineral density scores (L1-L4 T-score: 1.00±0.20 vs 1.10±0.21, p 0.001). The top-performing AI model based on AUC was a Vision Transformer Large using unmodified images without augmentation on isolated spine regions, achieving an AUC of 0.743, sensitivity of 21.2%, and specificity of 95.6%. This model correctly identified 32 of 89 patients with vertebral fractures while maintaining low false positive rates (32 false positives from 687 fracture-free patients). The highest-performing model had a precision-recall AUC of 0.357 (baseline PR-AUC = 0.16) was a Vision Transformer on unmodified images with augmentation and isolated spines, detecting 54 true positives with sensitivity of 47.8% and specificity of 80.0%, demonstrating improved fracture detection at the cost of increased false positives (144 vs 32). Across all architectures tested—including DenseNet and multiple Vision Transformer configurations—models consistently demonstrated high specificity (80%) but limited sensitivity (48%). The precision-recall AUCs ranged from 0.284 to 0.357, representing a 1.8- to 2.2-fold improvement over the baseline of 0.16, suggesting meaningful signal detection above random chance. Conclusion Although limitations exist such as potential ground truth errors when fractures occur outside the scanned lumbar region, our study provides proof of concept that AI based image classifiers can detect vertebral fractures from a single view, routine, AP lumbar spine DXA scans. This approach could help identify patients with previously undiagnosed vertebral fractures and enable timely referral for further assessment and management. With validation in larger multicentre cohorts, such models could be integrated into clinical workflows to enhance opportunistic vertebral fracture detection. Disclosure H. Amin: None. A. Marriucci: None. P. Angelov: None. J. Kerns: None. M. Bukhari: None.
Amin et al. (Wed,) studied this question.