Response to concerns regarding the critical evaluation of 3D-Shaper

I appreciate the opportunity to respond to the concerns raised¹ regarding my recent review of 3D-DXA and 3D-Shaper. I welcome this exchange. However, I do not agree that my review is unbalanced or misleading. I maintain that it identifies fundamental methodological limitations inherent to 3D-DXA.

I also contest the assertion that my description as “the inventor and developer” of the underlying methods is overstated. My statements, “As the inventor and developer of the software methods upon which 3D-DXA is built” and “Having developed the original software method”,² are factually accurate and reflect a first-hand understanding of the design decisions, assumptions, and limitations of 3D-DXA. In addition, I have extensive experience in measuring cortical parameters from QCT,^3,4 using methods equivalent to those later adopted in the extension of 3D-DXA, which further informs my review.

The authors now provide the number of model parameters used; citing my original study in which Cattell’s scree test was used. These values are almost identical to those in the models I developed, which is unexpected given the larger and more diverse dataset. It would be helpful to see the underlying scree plots, as well as a clarification on whether these numbers have been used consistently across all their studies.

The authors state that 3D-Shaper parameters are “estimated.” However, because the software uses a statistical model within a 3D-2D registration method, it functions as a traditional machine learning approach. Consequently, it is more technically accurate to state that the method predicts cortical bone parameters rather than estimating them. Given that the model is strictly constrained by a specific population, and the registration process is inherently prone to error, the resulting predictions cannot be assumed to provide accurate representations of cortical bone properties for clinical or research use.

To support claims of agreement between 3D-Shaper and QCT, the authors cite various correlation studies. However, they also acknowledge that cortical and trabecular bone correlate with aBMD “because of a well-established physiological principle: individuals with higher overall bone density, as indicated by aBMD, also tend to have higher cortical and trabecular bone densities.” These correlations therefore do not demonstrate the accuracy of the parameters predicted by 3D-DXA. Instead, they primarily reflect the underlying aBMD, rather than the ability of the method to recover true cortical bone properties.

In my review, I described a minimal evaluation to assess whether 3D-DXA can produce subject-specific cortical bone maps. This can be done by comparing the cortical thickness map obtained from 3D-DXA with that from a ground truth QCT scan, alongside a simple population mean template. If the 3D-DXA reconstruction does not outperform this mean estimator, it cannot be considered to recover the true subject-specific cortical bone distribution. Although comparisons of this type have been explored during the development of the method, to my knowledge, they have not been published. The absence of such validation in the literature remains a key limitation when assessing whether 3D-DXA provides truly subject-specific cortical information.

The authors continue discussing a precision study,⁵ which is essentially a repeatability study. Since 3D-DXA parameters are strongly dependent on aBMD, they are expected to reflect the high repeatability of aBMD measurements. However, this does not demonstrate that the derived parameters are accurate. A method may be precise yet systematically inaccurate, and precision alone is insufficient to establish clinical validity.

The preprint on cross-scanner compatibility,⁶ comparing the lower-resolution GE Healthcare Prodigy with the higher-resolution iDXA, is particularly revealing. Despite substantial differences in pixel size (0.6 × 1.05 mm for Prodigy and 0.3 × 0.25 mm for iDXA), the study reports excellent agreement between scanners for 3D-DXA parameters. If the reconstruction was driven by image detail, such a large difference in spatial resolution would be expected to affect the results. The absence of such an effect indicates that the method is largely insensitive to pixel information. Instead, the outputs are dominated by overall density (total aBMD). The reported agreement therefore reflects the consistency of aBMD across scanners, rather than the precision of parameters predicted by 3D-DXA.

The authors further argue that observing different changes in cortical and trabecular parameters demonstrates that 3D-Shaper provides information beyond aBMD. However, the presence of differing trends does not establish their validity.

Regarding the longitudinal assessment, I refer to the study by Winzenrieth et al.,⁷ where cortical vBMD derived from 3D-DXA is shown to increase in response to 18 mo of teriparatide treatment. Although this increase is not significant vs baseline, it is significantly different from the decrease observed in the placebo group. This analysis was performed in a relatively large cohort of 250 subjects. In contrast, three separate clinical trials using QCT over 18 mo of teriparatide treatment consistently report a significant decrease in cortical vBMD, even with a sample size as small as 19 subjects.⁴ The observation that cortical vBMD does not increase in proportion to integral vBMD in the 3D-DXA analysis can be explained by a fractional contribution from cortical bone that is visible at the inferior medial and lateral cortices of the shaft in the DXA images. While this may influence the cortical vBMD prediction, it does not reflect the true physiological changes in cortical bone.

I acknowledge that both research and clinical versions of the software exist. However, the underlying methodological limitations are the same. While certain features in the research version may not require regulatory approval, their use in scientific publications can still be problematic. In particular, the application of inappropriate statistical methods, as discussed in my review, risks misinterpreting physiological changes, especially in the context of drug trials.

Regulatory approval is subject to periodic review and may incorporate new evidence, including the concerns raised in my review. In particular, the use of a statistical model derived from a narrow Spanish population, together with associated T-score and Z-score reference curves, may warrant further scrutiny when applied to other populations. I therefore recommend that clinicians remain vigilant regarding ongoing regulatory assessments and any updates to the approval status and intended use of 3D-DXA and 3D-Shaper.

Conflict of interests

The author is involved in a legal dispute related to 3D-DXA and 3D-Shaper.