Interpretable Deep Learning Radiomics Model for Preoperative Prediction of High-Grade Soft Tissue Sarcomas: A Multicenter MRI Study.

This study aims to develop a preoperative fat-suppressed T2-weighted imaging (FS-T2WI)-based deep learning radiomics (DLR) model for predicting high-grade soft tissue sarcomas (STSs). 129 patients from the Cancer Imaging Archive (TCIA) database and one center were assigned to the training set, with 74 patients from the other center as the external validation set. Intratumoral and peritumoral regions of interest (ROIs) were manually segmented on FS-T2WI, and corresponding radiomic features were extracted. Deep learning features were extracted from whole-image FS-T2WI using a pre-trained 3D ResNet18. Feature selection was performed using recursive feature elimination (RFE) followed by least absolute shrinkage and selection operator (LASSO), after which intratumor, combined intratumoral-peritumoral, and DLR models were constructed respectively. Model performance was evaluated in the validation set using the area under ROC curve (AUC) and F1-score. Interpretability was assessed using gradient-weighted class activation mapping (Grad-CAM) and Shapley additive explanations (SHAP). The DLR model achieved the best predictive performance in the external validation set (AUC = 0.939, F1-score = 0.89), with no significant performance differences across subgroups stratified by age, sex, or tumor site. Grad-CAM revealed that extended peritumoral and broader contextual regions also contributed substantially to high-grade STSs discrimination beyond intratumoral areas, while SHAP analysis of the DLR model showed that peritumoral features-particularly those reflecting tumor heterogeneity-were crucial for identifying high-grade STSs. The interpretable DLR model represents a highly accurate, reliable, and clinically applicable preoperative approach for predicting high-grade STSs.