Bad science: Structural Parameters with 3D-DXA

In a recent article by Galgo Medical and Cetir Grup Mèdic1, the authors claim that “accurate estimates of structural parameters for the femur can be obtained from 3D-DXA models”. They report a correlation coefficient of 0.86 for the femoral neck axis length and 0.71 for the femoral neck shaft angle. Let me explain to you why these values should be taken with a grain of salt. The orientation of the hip is not known prior to a 3D-DXA analysis and has to be estimated in the same way as the shape and the bone mineral density distribution. In the below figure I illustrate that two models with different neck shaft angles can produce a similar shape outline when rotated differently. This means that for a certain bone projection (DXA image), 3D-DXA can produce femur models with completely different neck shaft angles, depending on what rotation it has converged to.

Why the neck axis length could not possibly have been estimated correctly is even more blatantly obvious. Gaglo Medical is still using my old software whereby the pelvis is not modeled, but is simply removed from the mask, and thus from the reconstruction process. This means that there is no information at all to guide the femoral head (and sometimes part of the femoral neck). It is simply impossible to get the neck axis length.

Why am I so certain about all this you ask me? Well, we actually already did this evaluation. In fact, it was Ludovic Humbert who did the experiments while he was still at UPF. He subsequently sent an abstract to the 2012 European Congress on Osteoporosis & Osteoarthritis (IOF-ECCEO12) in which he states that “The geometrical parameters showed correlation coefficients of r=0.94 (femoral neck axis length) and r=0.90 (femoral neck shaft angle). All the correlations were statistical significant (p-value”. The abstract can be found here. These results seemed suspiciously positive to me, knowing how the variations in the angular rotation of the femur affect the recovery of the neck shaft angle and how the limited field of view prevents the reconstruction of the femoral head. I therefore asked him to give me the raw data. I received the data for all measurements except the neck shaft angle and neck axis length. After a lot of nagging he finally sent me an e-mail stating: “There was a bug on the Femoral Neck shape angle. The true correlation coeff is poor (0.35), however significant (p=0.05).”

It’s nice if you are able to correlate measurements with a ground truth (QCT), but no conclusions can be derived from this as I explained in the evaluation of the cortical thickness here. At the very least you should compare it with what is currently available or another simple method such as manually measuring the parameters from DXA images directly. In an article by Ramamurthi et al.2, measurements from 2D DXA images were compared with measurements from QCT where a correlation coefficient of 0.90 was reported for the neck axis length. For other hip structure analysis (HSA) measurements, similar and some better correlations were found when measuring from 2D DXA 2,3 compared to the results presented by Clotet et al. Clearly 3D-DXA is not better than simply measuring the structural parameters from a 2D DXA image. Claiming the contrary for financial gain constitutes fraud in my opinion.

  1. Clotet J, Martelli Y, Di Gregorio S, Del Río Barquero LM, Humbert L. Structural Parameters of the Proximal Femur by 3-Dimensional Dual-Energy X-ray Absorptiometry Software: Comparison With Quantitative Computed Tomography. J Clin Densitom. 2017 Jun 14. pii: S1094-6950(17)30017-3.
  2. Ramamurthi K, Ahmad O, Engelke K, Taylor RH, Zhu K, Gustafsson S, Prince RL, Wilson KE. An in vivo comparison of hip structure analysis (HSA) with measurements obtained by QCT. Osteoporos Int. 2012 Feb;23(2):543-51.
  3. Ohnaru K, Sone T, Tanaka K, Akagi K, Ju YI, Choi HJ, Tomomitsu T, Fukunaga M. Hip structural analysis: a comparison of DXA with CT in postmenopausal Japanese women. Springerplus. 2013 Jul 20;2:331.