Development of the interpretable typing prediction model for osteosarcoma and chondrosarcoma based on machine learning and radiomics: a multicenter retrospective study

Background: Osteosarcoma and chondrosarcoma are common malignant bone tumors, and accurate differentiation between these two tumors is crucial for treatment strategies and prognosis assessment. However, traditional radiological methods face diagnostic challenges due to the similarity in imaging between the two.

Methods: Clinical CT images and pathological data of 76 patients confirmed by pathology from January 2018 to January 2024 were retrospectively collected from Guizhou Medical University Affiliated Hospital and Guizhou Medical University Second Affiliated Hospital. A total of 788 radiomic features, including shape, texture, and first-order statistics, were extracted in this study. Six machine learning models, including Random Forest (RF), Extra Trees (ET), AdaBoost, Gradient Boosting Tree (GB), Linear Discriminant Analysis (LDA), and XGBoost (XGB), were trained and validated. Additionally, the importance of features and the interpretability of the models were evaluated through SHAP value analysis.

Results: The RF model performed best in distinguishing between these two tumor types, with an AUC value close to perfect at 1.00. The ET and AdaBoost models also demonstrated high performance, with AUC values of 0.98 and 0.93, respectively. SHAP value analysis revealed significant influences of wavelet-transformed GLCM and First Order features on model predictions, further enhancing diagnostic interpretability.

Conclusion: This study confirms the effectiveness of combining machine learning with radiomic features in improving the accuracy and interpretability of osteosarcoma and chondrosarcoma diagnosis. The excellent performance of the RF model is particularly suitable for complex imaging data processing, providing valuable insights for the future.