The diagnostic accuracy of AI-driven opportunistic osteoporosis screening based on routine non-contrast CT

The diagnosis of osteoporosis is of great clinical significance for the prevention of fracture. To evaluate performance of a commercially available artificial intelligence (AI) solution for opportunistic osteoporosis screening using non-contrast computed tomography (NCCT) compared to dual-energy X-ray absorptiometry (DXA). This retrospective study included 518 patients who underwent both DXA and lumbar NCCT (LNCCT). Bone quality was classified into three groups—normal, osteopenia, and osteoporosis—based on DXA. Commercially available AI software was used to automatically segment vertebrae and extract volumetric bone mineral density (vBMD) values from T12 to L2 (thoracic vertebrae 12 to lumbar vertebrae 2) on LNCCT. Four classification methods were devised for AI-based vBMD assessment: method1 (average (avg) vBMD (T12+L1+L2) ), method2 (avg vBMD (T12+L1) ), method3 (avg vBMD (T12+L2) ), and method4 (avg vBMD (L1+L2) ). Agreements among AI-based methods and DXA were analyzed using intraclass correlation coefficients (ICCs), Bland-Altman analysis, and Linear Cohen’s weighted kappa statistics. Multi-categorical logistic regression and receiver operating characteristic (ROC) curves were employed to estimate the diagnostic performance of the four AI-based bone quality classification methods. A p-value of less than 0.05 was considered statistically significant. The AI-based methods from NCCT showed reasonable agreement with one another (ICC [95% confidence interval, CI]: 0.909[0.893–0.923]). The agreement between AI-based methods 1–4 and DXA was good (ICC [95%CI]: 0.689[0.641,0.732], 0.649[0.594,0.698], 0.666[0.616,0.712], 0.680[0.631,0.724], respectively). The decision function exhibited promising performance, with precision of 0.834 and recall of 0.735 for diagnosing osteoporosis in patients compared to those with normal and osteopenic conditions. The diagnostic performance of all four AI-based bone quality classification methods was found to be good when compared with DXA. Among these methods, avg vBMD (T12+L1+L2) (method 1) showed the best diagnostic performance. Not applicable.