Critical evaluation of 3D-DXA and 3D-Shaper: methodological limitations and their implications

Abstract

3D-DXA, as implemented in the software tool 3D-Shaper, is a software method that generates a 3D reconstruction of the proximal femur from a single 2D DXA image by registering a statistical model. Implementations of 3D-DXA aim to provide estimates of trabecular, cortical, and structural parameters, similar to those derived from quantitative computed tomography (QCT). As the inventor and developer of the software methods upon which 3D-DXA is built, I have been observing its adoption and widespread use with increasing concern. This article provides a critical evaluation of the methodological limitations inherent to 3D-DXA and discusses their implications for research and patient care. The primary issue is that the limited visibility of the cortex in a DXA image prevents 3D-DXA from accurately deriving cortical parameters. Instead, the software relies on predictions based on overall BMD rather than direct cortical measurements. This may lead to results that do not reflect actual cortical measurements. Additional concerns include the population bias due to the statistical model being derived from a specific demographic, and limited reconstruction accuracy by using single-view DXA images. These limitations have likely resulted in incorrect measurements and research outcomes, which have largely gone unrecognized due to the use of inappropriate performance assessment metrics and the absence of multiple comparison corrections in studies involving 3D-DXA. Despite these limitations, 3D-DXA has received regulatory approval in various countries, potentially compromising the accuracy of clinical diagnoses and treatment decisions. By highlighting these issues, this article aims to inform clinicians, researchers, and regulatory bodies about the significant limitations of 3D-DXA. It underscores the urgent need for a reevaluation of its use in research and clinical settings to prevent misinterpretation of results and to ensure patient safety.

Keywords: 3D-DXA; 3D-Shaper; bone mineral density; cortical parameter mapping; dual-energy X-ray absorptiometry; hip structure analsyis; quantitative computed tomography.

Figure 1

Flowchart illustrating the iterative process of 3D-DXA used to reconstruct the 3D proximal femur shape and density distribution from a single 2D DXA image. At each iteration, a projection is generated from the deformed density model instance, known as a digitally reconstructed radiograph (DRR). The DRR is then compared with the real DXA image by calculating a metric of similarity like the mean squared error. The model parameters and similarity transform (rotation, translation, and scale) are then adjusted in a way that attempts to increase the similarity in the next iteration. This process continues until the process has converged whereby the similarity between the DRR and the DXA image does not change beyond a predetermined threshold. Adapted from Whitmarsh et al.

Figure 2

Example DXA hip scan of a young adult acquired using the GE iDXA scanner. This figure was previously published in my earlier work.

Figure 3

The mean and the first 3 principal components (PCs) of the shape model (A) and projections of the density model (B), varying between −3 and +3 standard deviations (). This figure was adapted from Whitmarsh et al.

Figure 4

Example of generating an instance of a density model and measuring the cortical parameters from this volume. The top panels show cross-sections of the density model with the mean, PC1, PC2, PC3, and the new model instance generated by a linear combination of these volumes. Below them we see the values of the voxels taken from a line in the volume indicated by the red line in the cross-section. In the far left we can see how a stair step model (red) is fitted to the voxel values by smoothing the model (blue). This results in cortical thickness and cortical vBMD values that can be combined into a surface BMD (sBMD) value (green). When this is done at every point on the bone surface a color coded cortical surface map can be generated as shown above. Abbreviation: PC, principal component.

Figure 5

The color coded map of the mean (top) and maximum (bottom) cortical thickness estimation errors of 3D-DXA with respect to the same subject QCT scans. Adapted from Figure 8 in Humbert et al.

Figure 6

Reference data for a Spanish (A) and Argentinian (B) population, derived from and respectively. indicates significant differences compared with decade 20–30.