Multi-ethnic evaluation and calibration of 3D-DXA femoral shape, density and FEA strength
TL;DR
This is the first study to independently evaluate 3D-DXAs on multi-ethnic data across a wide age, including previously unstudied cohorts, and more than doubling the number of subjects used in prior validation works, providing detailed insights into the accuracy and limitations of DXA-based 3D reconstruction methods.
Abstract
Assessments of 3D geometry, bone density distribution and bone strength improve hip fracture risk prediction but currently require the use of computed tomography (CT). 2D-to-3D statistical shape and appearance model (SSAM) reconstructions from dual-energy X-ray absorptiometry (DXA) projections have been proposed as surrogates for CT-based models. Several studies have used 3D-DXAs reconstructed by 3D-Shaper’s software method, but no study has independently cross-validated 3D-DXAs on in vivo clinical images. The first aim of this study was to evaluate the extent 3D-DXA is applicable across a diverse population. 120 paired DXA and CT images from three ethnicities (age: 20–85; aBMD:0.611–1.214 g/cm 2 ) were analysed for differences in bone volume (BV), bone mineral content, vertex-to-surface distance, volumetric bone mineral density (vBMD) distribution and predicted bone strength. No subject or parameter groups showed outlier behaviour, as correlations for all results were above R 2 > 0.76. However, differences were observed for volumetric measurements (BV: slope = 0.89, bias = 9.73 cm 3 ; vBMD: slope = 0.88, bias = 6.27 g/cm 3 ) and bone strength. To improve the performance, we introduced an image-processing pipeline (named as 3D-DXA*) that calibrated the linear-regression slopes and intercepts for BV, total vBMD and the minimum fall strength (N = 96). Validation results (N = 24) in 11 fall orientations showed successful correction of proportional bias to unitary slope and zero intercept (before: slopes = 0.62–0.79, bias = 0.44–0.86 kN; after: slopes = 0.88–1.03, bias = 0.12–0.70 kN). However, Pearson correlations degraded by 0.01–0.02 for 2 cases and prediction scatter (RMSE) expanded by 3–29% for 10 fall orientations. Sub-analysis showed 3D-DXA* reduced differences in vBMD distribution with the CT models, but cohort-specific differences remained. This is the first study to independently evaluate 3D-DXAs on multi-ethnic data across a wide age, including previously unstudied cohorts, and more than doubling the number of subjects used in prior validation works. These results provide detailed insights into the accuracy and limitations of DXA-based 3D reconstruction methods.