This research was authorized by our institution’s ethical committee. Written informed consent was obtained from each patient before the MRI examination. We selectively recruited ninety-two patients who were pathologically diagnosed with STTs between January 2017 and September 2022 (51 males and 41 females, 16 to 86 years old with mean age 51.24 years). Inclusion criteria include: (1) all patients had undergone a 3.0T DCE MRI scan; (2) no chemotherapy or radiotherapy before surgery; (3) patients with histopathologically proven STTs. Exclusion criteria include: (1) poorly vascularized tumors like lipoma and well-differentiated liposarcoma; (2) inadequate image quality due to motion artifacts; (3) intermediate tumors such as myoepithelioma. All lesions were divided into groups of benign or malignant tumors based on the pathological categorization of soft-tissue tumors by the World Health Organization (2020) [21].

The MRI examinations were obtained using a 3.0 T MRI scanner (MAGNETOM Skyra; Siemens Healthcare, Erlangen, Germany). To identify lesions and assess morphological traits, we first administered conventional MRI series such as spin-echo T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), and fat-suppressed T2WI. We used multi-flip angle T1-weighted imaging technology to obtain T1 relaxation times (TRs) at three different flip angles (5°, 10°, and 15°) before contrast injection. Table S1 displays the MRI sequence parameters. T1WI three-dimensional volumetric interpolated breath-hold examination sequence was employed for DCE - MRI scans. The DCE images were taken on axial plane. The total acquisition time of VIBE sequence is 320 s, and the time resolution of each scan is 8 s; a total of 40 scans are obtained. In order to maintain a stable injection rate, we used an auto-injector for intravenous injection of gadoteridol (ProHance; Eisai, Tokyo, Japan) at a dose of 0.1mmol/kg at a rate of 2mL/s. Following that, at the same pace, we administered 20mL of physiological saline.

MR images on the picture archiving and communication system (PACS) (m-view v5.4.10.71, INFINITT Healthcare) were independently evaluated by two experienced radiologists (with 5 and 8 years of diagnostic practice) who were blinded to clinical information and histopathological reports. Consensus was reached on the MRI morphological data. The following was recorded: (1) size(maximum diameter of tumor); (2) location(head and neck, trunk, upper limb, lower limb); (3) shape (non-multilobulated or multilobulated); (4) margin (well-defined or ill-defined); (5) lesion internal enhancement pattern (homogeneous or heterogeneous); (6) tumor necrosis; (7) peri-tumoral edema. The margin of a mass that was clearly separated from surrounding structures, regardless of neighboring peritumoral edema, was called a “well-defined” tumor. Heterogeneous, defined as the presence of areas of low, intermediate, and high signal intensity in ≥ 50% of the tumor volume. High signal on T2-weighted imaging without enhancement was considered evidence of tumor necrosis. Peritumoral edema is characterized as a fluid-like, high signal in the peritumoral region on T2WI that can be distinguished from the tumor entity. Labels 6 and 7 were classified as ‘yes’ or ‘no’.

We included time-signal intensity curve (TIC) types and quantitative evaluations in the DCE-MRI data analysis. Based on the previously described approach, the TIC types were characterized as having no evident upward trend or consistently increasing (type I), rapidly increasing then flattening (type II), or rapidly increasing and dropping (type III) [22]. The region of interest (ROI) of the DCE-MRI image was positioned in the solid portion of the lesion based on regular MRI findings and enhancing features of the lesion. According to the size of the solid portions of the lesion, the ROIs spanned areas of change ranging from 1 to 5 cm2. Necrotic, cystic, and hemorrhagic areas were avoided when drawing the ROIs. For each patient, two experienced radiologists manually chose ROIs in four typical slices, and the average value was determined as the parameter value. Image processing was conducted using commercially accessible software (MItalytics, FITPU Healthcare, Singapore). The software allows for the selection of a unique arterial input function for each patient case and employs a constrained nonlinear optimization approach to match the different models. In total, we obtained 25 independent and derived qDCE parameters. The TOFTS model was used to derive the following parameters: Ktrans (min− 1), reverse reflux rate constant (Kep; min− 1), and extravascular extracellular volume (Ve; mL/mL). Similarly, the EXTOFTS model was used to calculate Ktrans, Kep, Ve, and volume fraction of plasma (Vp; mL/mL). The ATH, CC, and DP model were used to obtain the parameters of Ve, Vp, F (mL/min/mL), PS (mL/min/mL), mean transit time (MTT; s), and extraction fraction (E; %). Supplementary Material A1 mathematically describes the parameter fitting with the equations of the tissue concentration-time curve Ctiss(t) used to fit pharmacokinetic models.

All statistical analyses were conducted using R4.2.1 (www.r-project.org) and SPSS (version 25.0; SPSS, Chicago, III, USA).

MR morphologic characteristics and TIC types: The Shapiro-Wilk test was used to assess the normal distribution. The variables relating to benign and malignant lesions were assessed using the Mann-Whitney U test, χ2 test and univariate logistic regression analysis. Subsequently, we introduced the MRI morphological features and TIC with P < 0.05 into a multivariate logistic regression. The results of P < 0.05 were regarded as significant, and the findings were incorporated in the construction of traditional imaging model and subsequent studies.

Construction of qDCE models: We constructed the TOFTS, EXTOFTS, ATH, CC, and DP models using multivariate logistic regression based on the qDCE parameters derived from each TK model.

Development of top-parameter model: The Mann-Whitney U test, t test, and univariate logistic regression analysis were used to evaluate the qDCE parameters related to benign and malignant STTs. To filter the optimal parameters, we firstly attempted to incorporate all qDCE parameters of TK models into multifactor logistic regression. The variance inflation factor(VIF) was used to analyze multicollinearity. Some TK models have the same parameters or different TK model parameters interact with each other, multicollinearity is high, and some VIF values exceed 10. Therefore, the qDCE parameters with univariate logistic regression P < 0.05 were included in multiple logistic regression analysis with each TK model as the unit. The qDCE parameters with P < 0.05 were considered the top-parameters associated with the differentiation of benign and malignant STTs. There is no multicollinearity between top-parameters, and VIF < 10. Finally, we used the top-parameters to create a top-parameter model.

Construction and evaluation of the DP + Traditional imaging model and the comprehensive model: (1) The DP + Traditional imaging model was constructed using all parameters of the DP model and statistically significant traditional imaging features. (2) For the purpose of building the comprehensive model, we employed the top-parameters and statistically significant morphological features as input components. The area under the curve (AUC) was used to evaluate diagnostic performance. We also evaluated the accuracy, sensitivity, and specificity of each TK model and qDCE parameter.