Preventing fragility fractures: AI-based opportunistic osteoporosis detection in routine CT .
Authors
Affiliations (3)
Affiliations (3)
- Floy GmbH, Ludwigstraße 9, 80539, Munich, Germany.
- Department of Medicine II, University Hospital, LMU Munich, Munich, Germany.
- Floy GmbH, Ludwigstraße 9, 80539, Munich, Germany. [email protected].
Abstract
This study shows that an artificial intelligence tool can identify patterns of reduced bone mineral density in routine CT scans that would otherwise remain unreported. Compared to radiologists, the AI system flagged substantially more cases with potential signs of reduced bone mineral density, offering a scalable approach to support earlier identification of individuals who may benefit from further assessment. Osteoporosis often goes unrecognized in routine computed tomography (CT) scans performed for unrelated indications, leading to diagnosis only after a fragility fracture occurs. We present an artificial intelligence (AI)-based tool for opportunistic bone mineral density (BMD) assessment on contrast-enhanced and native CT scans of the entire spine to enable detection of patterns suggestive of reduced BMD. We trained a neural network to classify BMD of individual vertebrae as normal, osteopenic, or osteoporotic, using labels derived from Dual-Energy X-ray Absorptiometry (DXA) T-scores as the reference standard. The training dataset included 1084 sagittal spine CT series from 526 patients across multiple US and German centers, encompassing cervical, thoracic, and lumbar regions. The model was evaluated on an independent holdout set of 270 patients. Additionally, in a retrospective analysis of 354 patients who later sustained fragility fractures, the AI model was applied to contrast or native CT scans acquired 12-24 months prior to the fracture event to assess how often the AI model would flag patterns suggestive of reduced BMD in patients who subsequently sustained fragility fractures. On the independent holdout set, the algorithm achieved a specificity of 84% and a sensitivity of 67%. In the retrospective study, the AI model flagged patterns of reduced BMD in 184 of 354 patients (87 osteoporotic, 97 osteopenic), while radiologists had reported reduced BMD in only 41 cases. To the best of our knowledge, this AI model is the first to classify BMD across the entire spine on both contrast and native CT scans. It demonstrates robust real-world performance and offers a promising approach for scalable opportunistic BMD screening in routine CT imaging.