Fig. 1

A general flowchart of data analysis. A Imaging-derived features were extracted by the deep learning analysis and handcrafted radiomics analysis on multimodality medical images, including CT and MRI, respectively. B Predictive models on both deep learning and handcrafted radiomics features for classification of lipoma and WDLPS were approached by machine learning methods including features selection and model construction. The deep learning-based model with the optimal performance was chosen to generate a deep learning signature. An integrated differentiation model was constructed by the deep learning signature and independent clinical predictors. All differentiation models were evaluated by ROC curves, precision-recall plots, and calibration plots in both training and validation cohorts. CT, Computed tomography; MRI, Magnetic resonance imaging; WDLPS, Well-differentiated liposarcoma; ROC, Receiver operating characteristic; SVM, Support vector machine