Multicentre MRI-based machine learning model for noninvasive prediction of pulmonary metastasis in osteosarcoma...

To develop and externally validate a multicenter MRI-based machine learning model integrating intra-tumoral heterogeneity (ITH) index, conventional radiomics (C-radiomics) and clinical variables for predicting one-year pulmonary metastasis (PM) in osteosarcoma.

This retrospective study enrolled 320 patients with histologically confirmed osteosarcoma from four institutions, comprising internal (n = 254, Centers A-C) and external sets (n = 66, Center D). Pre-treatment contrast-enhanced T1-weighted fat-suppressed MRI was used for tumor segmentation and feature extraction. ITH features were obtained through supervoxel-based clustering, and C-radiomics features were derived conventionally. An XGBoost model integrating ITH index, C-radiomics, and clinical variables was developed. Model performance was evaluated using ROC, calibration, and decision curve analysis (DCA), with SHAP and subgroup analysis providing interpretability and robustness.

Within one year after surgery, 39.4% of patients developed PM. The combined model achieved the highest predictive performance across sets, with an AUC of 0.843 (95% CI: 0.823-0.869), 73.8% accuracy, 78.2% sensitivity, and 81.1% specificity on the independent external test set, outperforming all single- and dual-modality models. Calibration and DCA confirmed strong model reliability and clinical utility across a broad threshold range. The ITH index (OR = 6.723, p = 0.008) and C-radiomics score (OR = 7.962, p = 0.001) were independent predictors of PM. Subgroup analysis demonstrated consistent performance across age, sex, stage, and tumor site (AUC range: 0.801-0.853).

The MRI-based model integrating ITH, C-radiomics, and clinical variables enables accurate, noninvasive prediction of early PM in osteosarcoma, supporting personalized risk stratification and clinical decision-making.