This study investigated the numerical convergence characteristics of specimen-specific "voxel-based" finite element models of 14 excised human cadaveric lumbar vertebral bodies (age: 37-87; M = 6, F = 8) that were generated automatically from clinical-type CT scans. With eventual clinical applications in mind, the ability of the model stiffness to predict the experimentally measured compressive fracture strength of the vertebral bodies was also assessed. The stiffness of "low"-resolution models (3 x 3 x 3 mm element size) was on average only 4% greater (p = 0.03) than for "high"-resolution models (1 x 1 x 1.5 mm) despite interspecimen variations that varied over four-fold. Damage predictions using low- vs high-resolution models were significantly different (p = 0.01) at loads corresponding to an overall strain of 0.5%. Both the high (r2 = 0.94) and low (r2 = 0.92) resolution model stiffness values were highly correlated with the experimentally measured ultimate strength values. Because vertebral stiffness variations in the population are much greater than those that arise from differences in voxel size, these results indicate that imaging resolution is not critical in cross-sectional studies of this parameter. However, longitudinal studies that seek to track more subtle changes in stiffness over time should account for the small but highly significant effects of voxel size. These results also demonstrate that an automated voxel-based finite element modeling technique may provide an excellent noninvasive assessment of vertebral strength.