The radiomics fingerprint of cartilage tumours: radiomics-based MRI differentiation of enchondroma and atypical cartilaginous tumour

Abstract Purpose This study aimed to develop and validate machine learning models based on quantitative radiomics parameters extracted from T1-weighted MRI to differentiate enchondromas from atypical cartilaginous tumours (ACTs). Methods A retrospective cohort comprising 66 patients (35 with histopathologically confirmed enchondroma and 31 with ACT) was included in the study. T1-weighted MRI images were used for 2D segmentation, performed independently by two experienced observers on all visible slices of each lesion. A comprehensive set of 107 radiomics features was extracted from these segmented regions of interest. LASSO regression was applied for dimensionality reduction. Four distinct machine learning algorithms—Support Vector Machine (SVM), Random Forest Classifier (RFC), Extreme Gradient Boosting (XGBoost), and Decision Tree Analysis—were trained and validated using a 70:30 data split. Results The radiomics features demonstrated high inter- and intra-observer reproducibility. All evaluated machine learning models exhibited strong diagnostic performance, with Area Under the Curve (AUC) values exceeding 0.90. Specifically, SVM achieved an AUC of 0.922 (95% CI 0.893–0.951), RFC yielded an AUC of 0.920 (95% CI 0.881–0.963), and Decision Tree Analysis showed an AUC of 0.949 (95% CI 0.927–0.972). Notably, the XGBoost model achieved the highest diagnostic efficacy, boasting an impressive AUC of 0.987 (95% CI 0.976–0.999), coupled with a sensitivity of 89.35% and a specificity of 96.55%. Conclusion Our results indicate that the combination of MRI-based radiomics and machine learning algorithms, particularly XGBoost, offers a non-invasive and highly accurate method for distinguishing enchondroma from ACT.