A predictive MRI radiomics model for histologic differentiation in soft tissue sarcomas

Abstract

Background/Objectives: The aim of this study was to develop and validate a robust, radiomics-based classification model that uses pre-treatment MRI to non-invasively differentiate among major soft tissue sarcoma (STS) subtypes and a benign mimic. Methods: In this retrospective study, a cohort of 332 patients with biopsy-proven leiomyosarcoma, myxofibrosarcoma, myxoid liposarcoma, dedifferentiated liposarcoma, and undifferentiated pleomorphic sarcoma, along with the benign mimic intramuscular myxoma, was analyzed. Pre-treatment T1-weighted fat-saturated contrast-enhanced and T2-weighted fat-saturated MRI sequences were used for analysis. Following manual tumor segmentation, 1240 three-dimensional radiomic features were extracted. An XGBoost classifier was trained and validated using a robust 250-iteration bootstrap framework with nested cross-validation to ensure rigorous feature selection and unbiased performance evaluation. The model’s performance was assessed independently on T1-only, T2-only, and combined T1+T2 feature sets. Results: The combined T1 and T2 model achieved superior performance with an accuracy of 0.68 ± 0.04 and an AUC of 0.92 ± 0.02. At the subtype level, balanced accuracy was highest for intramuscular myxoma (0.91 ± 0.05), dedifferentiated liposarcoma (0.84 ± 0.06), and leiomyosarcoma (0.83 ± 0.05). SHAP analysis identified key features driving predictions, such as low T2 GLSZM Zone Size Entropy for myxoma and high T2 GLSZM Gray-Level Variance for leiomyosarcoma, which aligns with known pathological characteristics. Misclassifications predominantly occurred between subtypes with overlapping radiomic profiles. Conclusions: Radiomics applied to pre-treatment MRI enables robust, non-invasive classification of STS subtypes, demonstrating strong clinical potential for improving diagnostic confidence and informing triage strategies.