Effect of microcomputed tomography voxel size on the finite element model accuracy for human cancellous bone

The level of structural detail that can be acquired and incorporated in a finite element (FE) analysis might greatly influence the results of microcomputed tomography (microCT)-based FE simulations, especially when relatively large bones, such as whole vertebrae, are of concern. We evaluated the effect of scanning and reconstruction voxel size on the microCT-based FE analyses of human cancellous tissue samples for fixed- and free-end boundary conditions using different combinations of scan/reconstruction voxel size. We found that the bone volume fraction (BV/TV) did not differ considerably between images scanned at 21 and 50 microm and reconstructed at 21, 50, or 110 microm (-0.5% to 7.8% change from the 21/21 microm case). For the images scanned and reconstructed at 110 microm, however, there was a large increase in BV/TV compared to the 21/21 microm case (58.7%). Fixed-end boundary conditions resulted in 1.8% [coefficient of variation (COV)] to 14.6% (E) difference from the free-end case. Dependence of model output parameters on scanning and reconstruction voxel size was similar between free- and fixed-end simulations. Up to 26%, 30%, 17.8%, and 32.3% difference in modulus (E), and average (VMExp), standard deviation (VMSD) and coefficient of variation (COV) of von Mises stresses, respectively, was observed between the 21/21 microm case and other scan/reconstruction combinations within the same (free or fixed) simulation group. Observed differences were largely attributable to scanning resolution, although reconstruction resolution also contributed significantly at the largest voxel sizes. All 21/21 microm results (taken as the gold standard) could be predicted from the 21/50 (r2adj= 0.91-0.99;p<0.001), 21/110 (r2adj =0.58-0.99;p<0.02) and 50/50 results (r2adj=0.61-0.97;p<0.02). While BV/TV, VMSD, and VMExp/sigma(z) from the 21/21 could be predicted by those from the 50/110 (r2adj =0.63-0.93;p<0.02) and 110/110 (r2adj =0.41-0.77;p<0.05) simulations as well, prediction of E, VMExp, and COV became marginally significant (0.04