Automated proximal femur CT radiomics identifies age- and sex-related radiomic patterns of bone structure: a nationwide multicohort validation study

Age and sex are major determinants of bone structure, yet subtle skeletal modifications remain difficult to quantify on routine CT. Radiomics may capture these changes, but automated proximal femur radiomics has not been validated in large nationwide cohorts. This retrospective nationwide multicohort study included 4152 participants who underwent noncontrast routine CT. A subset of proximal femurs was manually annotated to train an nnU‑Net‑based automated segmentation model. Radiomic features were extracted from segmented volumes using PyRadiomics. Machine learning models identified radiomic patterns associated with age (regression) and sex (classification), evaluated in a development cohort ( n = 2,969) and an external validation cohort ( n = 1,183). Segmentation performance was assessed using the Dice similarity coefficient. Age‑related models were evaluated with R², mean absolute error (MAE), and root mean squared error (RMSE); sex‑related models with AUC, accuracy, and F1‑score. Femoral neck volumetric bone mineral density (vBMD) derived from quantitative computed tomography in the external validation cohort was used to examine the biological relevance of the age‑related radiomics patterns. The final cohort comprised 4,152 participants (1,947 men; median age 59 years; range 18–99 years). Automated segmentation achieved a Dice coefficient of 0.963 in the test set. The best age‑related model (support vector regression) yielded an external validation R² of 0.50, MAE of 8.84 years, and RMSE of 11.32 years. For sex-related classification, support vector classification (SVC) and XGBoost showed the strongest external performance; both achieved an AUC of 0.94, with SVC showing slightly higher accuracy and F1-score than XGBoost. The radiomics‑derived age‑related index showed a stronger association with femoral neck vBMD than chronological age (Spearman r = − 0.546 vs. −0.443; Steiger’s Z = 4.97, P < 0.001). Automated proximal femur CT radiomics identified radiomic patterns related to bone structure changes associated with age and sex across nationwide multicohort datasets. Its association with vBMD supports the biological relevance of these radiomics patterns, warranting further validation with clinical outcomes. ClinicalTrials.gov NCT07162168 08/29/2025 (Initial Release) Retrospectively registered