Diagnostic Value of Magnetic Resonance Imaging Radiomics and Machine-learning in Grading Soft Tissue Sarcoma: A Mini-review on the Current State

Soft tissue sarcomas (STS) are a heterogeneous group of rare malignant tumors. Tumor grade might be underestimated in biopsy due to intratumoral heterogeneity. This mini-review aims to present the current state of predicting malignancy grades of STS through radiomics, machine learning, and deep learning on magnetic resonance imaging (MRI). Several studies investigated various machine-learning and deep-learning approaches in T2-weighted (w) images, contrast-enhanced (CE) T1w images, and DWI/ADC maps with promising results. Combining semantic imaging features, radiomics features, and deep-learning signatures in machine-learning models has demonstrated superior predictive performances compared to individual feature sources. Furthermore, incorporating features from both tumor volume and peritumor region is beneficial. Especially random forest and support vector machine classifiers, often combined with the least absolute shrinkage and selection operator (LASSO) and/or synthetic minority oversampling technique (SMOTE), did show high area under the curve (AUC) values and accuracies in existing studies.