To assess the significance of regional quantitative computed tomography measurements of bone density with respect to mechanical strength in the human lumbar spine, 58 vertebrae (from 12 males, 10 females) were scanned in vitro with multiple-thin-slice quantitative computed tomography and then compressed to fracture. With computer graphics, 18 specific regions of physical density and 10 combination averages of density were identified within each vertebral body. To ensure the statistical independence of data, the individual vertebral specimens were assigned to one of three groups (T11-L1, L2-L3, or L4-L5). Use of best-subsets procedures resulted in regression models to predict fracture strength. These models used specific regional density values and often the age and sex of the donors. The correlation coefficients that resulted from the multiple regression models ranged from r = 0.88 to r = 0.95. When the density values were multiplied by the minimum cross-sectional area of the vertebral body, similar regional density averages were selected, and the predictive values were slightly improved (r = 0.94-0.97). The heterogeneity of the density samples (measured as standard deviation) in multiple regression fashion also produced strong correlation coefficients (r = 0.88-0.94). The bone density in an anterior cylinder of the midplane region, the location measured most often in quantitative computed tomography densitometry, was strongly correlated (r = 0.85) to fracture load for the T12-L1 group (N = 20), but was not significant for the other two groups of vertebrae. The cancellous bone density from the female data was not found to be significantly different from the male data set. These data suggest that when the density of specific regions in the vertebral main body are measured, most of the variance in the whole-bone strength could be predicted, and these optimal regions (a relative measure of cancellous heterogeneity) appeared to vary with spine level. Using the bone density from multiple regions of the cancellous interior may ultimately yield clinically useful information relevant to patient treatment decisions.