How long does 3D-DXA take to produce its fake results
While I have already given many examples of the fraudulent nature of the 3D-DXA software sold by Galgo Medical and DMS-Apelem. I would like to highlight one more element of their quackery. What I am referring to is the computation time. In a recent publication by members of Galgo Medical1 they state that the 3D-DXA software generates a reconstruction in, on average, 1 minute and 30 seconds. In this publication Humbert et al. state: “A great reduction in computing time was obtained by implementing the 3D-DXA algorithms in C++ using the Insight Segmentation and Registration Toolkit (ITK)  and taking advantage of multicore system (4 cores, 8 threads) used in this study”. Let me explain something to you. The authors of this study used my software, which was already developed in C++ using the ITK library. I expect the statement “taking advantage of multicore system” should be taken very loosely. The bottleneck in the software is the computation of the thin plate spline transform, which can not be performed in parallel. Whatever they have been able to compute in paralleled will provide only a minimal improvement in the computation time. In contrast, I reported a computation time of approximately 1 hour in the original paper on the 3D-DXA technology2. Väänänen et al. even report a computation time of a whopping 40 hours3. So what is going on here? Is Galgo Medical lying again?
There are actually a few ways to reduce the computation time:
- Reducing the resolution: This includes the number of control points to deform the model, the number of voxels in the density model and the number of samples you take to compute the similarity between the DXA image and the projection of the model. Unfortunate, if you reduce any of these resolutions, you also reduce the accuracy of the reconstruction.
- Reducing the number of iterations: Running the reconstruction process with more iterations means you can come closer to an optimal solution. Unfortunately, this optimal solution might not be the correct one and can be a so-called local minimum. In order to avoid local minima, you have to rerun the optimization using a new set of starting parameters. But never is it guaranteed that you do not end up in a local minimum, the chance of this happening is just reduced with more iterations. Thus, reducing the number of iteration will reduce your chance to finding an accurate reconstruction.
- Reducing the number of model parameters: The statistical model is defined by a set of parameters which describe the way in which the model deforms. The more model parameters you retain in your model, the better the model reflects the input population. Unfortunately, the more model parameters, the longer it takes to compute a reconstruction. By not including sufficient model parameters you limit the ability of the model to form to other bones and forces the model to even more closely follow the relationships in the input population (Spanish postmenopausal women in this case).
It is likely that Galgo Medical significantly reduced some or all of these parameters to reduce the computation time to around 2 minutes, which coincidentally is also pretty much the maximum time a clinician might be willing to wait for some results. As a consequence, the accuracy of the reconstruction will be considerably diminished. As the developer of this technology I can tell you that there is simply no way that you can get a reasonable reconstruction in 2 minutes. Prof. Guoyan Zheng of the University of Bern, who has developed very similar technologies, agreed with me on this and expressed to me his great disbelieve about the claims made by the members of Galgo Medical.
But let’s be honest, all this is rather trivial. I have already shown that 3D-DXA does not work as Galgo Medical claims it does, which they clearly are aware of considering they chose not to publish an evaluation which would have shown just that. Altogether there is no difference between fake results in 2 minutes or fake results in one hour. Galgo Medical was not lying, they simply preferred to provide fake results in 2 minutes. Rather sensible don’t you think?
- L. Humbert et al. 3D-DXA: Assessing the Femoral Shape, the Trabecular Macrostructure and the Cortex in 3D from DXA images. IEEE Trans Med Imaging. 2017 Jan;36(1):27-39.
- T. Whitmarsh, et al., Reconstructing the 3D shape and bone mineral density distribution of the proximal femur from dual-energy X-ray absorptiometry, IEEE Transactions on Medical Imaging, vol. 30(12), pp. 2101-14, 2011.
- S.P. Vaananen et al. Generation of 3D shape, density, cortical thickness and finite element mesh of proximal femur from a DXA image. Med Image Anal, vol. 24, no. 1, pp. 125-134, Jun 19, 2015.