Background and Aim: Dual-energy X-ray absorptiometry (DXA) is gold standard method for measuring bone mineral density (BMD) and diagnosing osteoporosis, but DXA is not widely available in low-resource settings. In this study, we developed an AI system to estimate BMD from plain radiographs for osteoporosis screening.
Methods: The development of the AI was based on 5134 plain X-rays, and testing was based on 1926 plain X-rays from the Vietnam Osteoporosis Study. Anteroposterior and oblige digital X-ray of hip and spine were taken by FCR Capsula XLII (Fujifilm Corp., Tokyo, Japan). We used seven Deep Convolution Neural Network to estimate areal BMD at the lumbar spine, total hip, and femoral neck. We termed the estimated BMD as 'xBMD'. We then compared xBMD with DXA-based areal BMD (aBMD) measured at femoral neck, total hip, and lumbar spine (Hologic Horizon, Hologic Corp., Bedford, MA, USA). The concordance between xBMD and aBMD was assessed by the coefficient of correlation and the Bland-Altman approach.
Results: The correlation between xBMD and aBMD was 0.90 (CI 95%: 0.88 – 0.91), 0.91 (CI 95%: 0.89 – 0.92), and 0.87 (CI 95%: 0.85 – 0.88) for femoral neck, total hip and lumbar spine, respectively. The correlation was greater in women than men, but the difference was not statistically significant. When aBMD was used to classify into osteoporosis vs non-osteoporosis, the discrimination of xBMD was high, with area under the ROC curve being 0.94 (CI 95%: 0.92 – 0.96) for femoral neck, 0.95 (CI 95%: 0.93 – 0.97) for total hip and 0.93 (CI 95%: 0.89 – 0.96) for lumbar spine.
Conclusion: These results suggest that it is possible to accurately predict areal BMD from plain radiographs, and that the AI system developed here can be used as an effective tool for opportunistic screening osteoporosis in low-resourced and high-volume settings.