To develop a multiclass radiomics model for differentiating chondroid bone tumors using preoperative MRI.

This retrospective study included 120 patients (92 enchondromas, 16 low-grade chondrosarcomas, and 12 intermediate-to-high-grade chondrosarcomas) who underwent contrast-enhanced MRI between 2009 and 2019. Tumor segmentation was manually performed by a musculoskeletal radiologist and validated by a senior radiologist. We used least absolute shrinkage and selection operator (LASSO) and random forest (RF) for feature selection and classification, with and without synthetic minority oversampling technique (SMOTE). Model performance was evaluated using five-fold cross-validation with average precision, accuracy, area under the curve (AUC), and weighted kappa statistics.

The LASSO + RF model based on all sequences achieved the highest accuracy (0.826 ± 0.065) and AUC (0.967 ± 0.027). The highest mAP (0.750 ± 0.095) was observed in the SMOTE-enhanced T2WI-based model, highlighting the potential impact of class imbalance. Quadratic weighted kappa values ranged from 0.648 to 0.731 across models, indicating substantial agreement with pathological results.

Preoperative MRI-based radiomics provides a robust method for the classification of chondroid bone tumors, potentially enhancing clinical decision-making.