Metastatic lesions in the spine change local bone mechanical properties and increase fracture risk. Computed tomography (CT) based finite element (FE) models are used to estimate vertebral strength and stiffness, but their predictive accuracy depends on how CT-derived material properties are mapped into the FE mesh. This study evaluated the effect of two CT-based material grouping strategies—Adaptive Clustering (AC) and Uniform fixed-width grouping—on material-mapping fidelity and on simulated vertebral strength and stiffness in vertebrae with different metastatic lesion types.
The analysis used 44 vertebrae from 11 donors. Vertebrae were categorized by imaging-observed lesion type: 8 osteolytic, 12 osteoblastic, 12 mixed, and 12 with no observed lesion (NOL). CT-based FE models were generated for each vertebra and material mapping was applied using grouping schemes ranging from 2 to 500 discrete material groups.
Two grouping strategies were compared:
Adaptive Clustering (AC): a data-driven grouping approach that adapts group boundaries to the underlying CT-derived material distribution.
Uniform fixed-width grouping: groups defined by evenly spaced material-value intervals across the CT-derived range.
Models across the range of group numbers were compared for material mapping error and for simulated mechanical outputs (strength and stiffness). An exploratory, lesion-stratified assessment compared simulated results with experimental mechanical measurements.
Material-mapping fidelity was quantified using the root mean square error (RMSE) of Young's modulus between the original CT-derived continuous mapping and the discretized grouped mapping. Across the tested range, Adaptive Clustering (AC) produced a significantly lower Young’s modulus RMSE than Uniform grouping (p < 0.05), indicating improved fidelity in representing the CT-derived material distribution when using AC.
Simulated vertebral strength and stiffness were examined across group counts. Both simulated outcomes tended to stabilise by approximately 50 material groups; increasing the number of groups beyond 50 had little additional effect on simulated strength or stiffness values.
At the 50-group level, simulations showed the following overall associations with experimental measurements:
These correlations indicate that numerical stabilisation of FE outputs with respect to material discretization does not necessarily equate to high experimental accuracy.
When stratified by lesion type at 50 material groups, simulated strength correlations with experimental strength varied by subgroup:
Simulated stiffness correlations also differed by subgroup:
These lesion-stratified comparisons were exploratory. The authors note limited subgroup sizes, which restricts robustness of these subgroup-specific conclusions.
Bland–Altman analyses comparing FE-predicted fracture load with experimental fracture load indicated a systematic underestimation by the simulations. The presence of consistent bias highlights that even with improved material-mapping fidelity (AC) and numerical stabilization of outputs, the models did not fully capture experimental fracture loads.
Key implications from the study are:
Material mapping accuracy benefits from Adaptive Clustering over Uniform fixed-width grouping, as demonstrated by lower Young’s modulus RMSE.
FE predictions of vertebral strength and stiffness numerically stabilise by around 50 material groups, so increasing grouping resolution above this threshold yields diminishing returns for simulated outputs.
Despite improved mapping fidelity and numerical stabilization, FE models still systematically underestimated experimental fracture load and showed variable agreement across lesion types, with best performance in vertebrae without observable lesions.
Lesion-bearing vertebrae present a greater challenge for accurate experimental prediction; lesion-type stratification showed lower correlations and, in some cases, negligible predictive value for stiffness.
The lesion-stratified results are exploratory and should be interpreted cautiously because subgroup sample sizes were limited. The study demonstrates that better numerical representation of CT-derived material properties (via AC) improves mapping fidelity, but that numerical convergence of FE outputs does not guarantee experimental accuracy. Researchers using CT-based FE models for vertebral strength prediction in the context of metastatic disease should consider both mapping strategy and the potential for systematic bias when interpreting predicted fracture loads.
The authors declared no competing interests. Ethical oversight and approvals were obtained (MGB IRB 2016P001582), and appropriate participant consent procedures were followed. All data produced in the study are available from the authors upon reasonable request.