Preoperative Multi-model Images-based Radiomics Model for Distinguishing Spinal Osteosarcoma and Chondrosarcoma

This study aimed to develop and validate a radiomics fusion model based on CT and MRI for distinguishing between spinal osteosarcoma and chondrosarcoma, and to compare the performance of models derived from different imaging modalities.

A retrospective analysis was conducted on 63 patients with histologically confirmed spinal osteosarcoma (n=20) and chondrosarcoma (n=43). Radiomics features were extracted from CT and MRI (T1-weighted, T2-weighted, and T2-weighted fat-suppressed) sequences, followed by feature selection using univariate logistic regression and LASSO. Eight machine learning models were utilized to construct radiomics models, based on CT, MR, both CT and MR, and clinical information combined with CT and MR. Models were evaluated via five-fold cross-validation and compared against radiologists' interpretations using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, F1 score, and Matthews correlation coefficient.

The MRI-based radiomics model using linear discriminant analysis achieved the highest diagnostic performance (AUC=0.963, sensitivity=95.3%, specificity=80.0%), significantly outperforming both CT-based models (AUC=0.700) and radiologists' diagnosis (p<0.001). The CTMR and clinico-CTMR models did not show significant improvement over the MR model. The MR model demonstrated excellent calibration and clinical utility, with substantial net benefit across threshold probabilities.

The superior performance of the MRI-based model highlighted the value of MRI radiomics in tumor differentiation. This clinically practical tool may support preoperative diagnosis using routine MRI, potentially facilitating more timely treatment decisions.

In conclusion, the MRI-based radiomics model enabled accurate preoperative discrimination between spinal osteosarcoma and chondrosarcoma.