<![CDATA[Introduction: Reliable preoperative discrimination of malignant from benign soft-tissue tumours (STTs) governs biopsy, surgical-margin and neoadjuvant decisions, yet conventional MRI interpretation is experience-dependent and biopsy is invasive and prone to sampling error. We aimed to develop and internally validate a machine-learning radiomics model from multiparametric MRI (mpMRI) for this task across multiple institutions. Methods: In this STARD 2015-compliant retrospective multi-institutional diagnostic-accuracy study, 215 patients (132 benign, 83 malignant) with histopathologically confirmed STTs imaged at three South Sumatran centres (2019–2023) were split 70:30 into training (n=150) and validation (n=65) cohorts. Radiomic features from T1W, T2W fat-suppressed and ADC maps underwent ICC-stability filtering and LASSO selection; SVM, Random Forest and XGBoost classifiers were compared against histopathology (reference standard) and blinded radiologist visual reads. Sensitivity, specificity, predictive values, likelihood ratios (95% CIs), ROC (DeLong), Cohen’s κ and multivariable logistic regression were computed. Results: A 14-feature signature was selected from 945 ICC-stable features. In internal validation, XGBoost achieved AUC 0.92 (95% CI 0.88–0.95), sensitivity 86.7% (70.3–94.7), specificity 91.4% (77.6–97.0), PPV 89.7%, NPV 88.9%, accuracy 89.2%, LR+ 10.1 and LR− 0.15. XGBoost exceeded SVM (AUC 0.84; DeLong p=0.012) and radiologist visual read (AUC 0.78; p<0.001; McNemar p=0.027). Inter-reader κ was 0.78 (0.63–0.94). The radiomics signature (adjusted OR 3.32, p<0.001) and lower ADC (OR 0.21, p<0.001) were independent malignancy predictors. Conclusion: An mpMRI XGBoost radiomics model provides accurate, non-invasive discrimination of malignant from benign STTs with high specificity and a clinically useful positive likelihood ratio, supporting its role as PACS-integrated decision support and as a triage tool in resource-variable settings.]]>
