Patients

This retrospective study received institutional review board approval, and the requirement for informed consent was waived for all participating institutions.

The training set consisted of 168 patients from the hospital 1. Two validation sets were created: the first set (validation set) was composed of 123 patients from two other hospitals (the hospital 2 and hospital 3), and the second set (TCIA set), used to assess generalizability, was generated using a publicly available dataset from TCIA (https://doi.org/10.7937/K9/TCIA.2015.7GO2GSKS), from which we obtained a collection of images from 44 patients with STS [21].

Annex A1 supplements the criteria for inclusion and exclusion of patients.

Clinical data including age, gender, and FNCLCC grade were collected.

Follow-up and survival analysis

All patients were followed up every 3–6 months with MRI or CT scanning during the 2 years following surgery and semiannually thereafter. Training and validation set data were censored in November 2021 and June 2020, respectively, and TCIA set data were censored in November 2011. PFS was defined as the time between surgery and radiographic detection of metastasis or recurrence, the day of death without evidence of progression, or the last negative follow-up.

MRI semantic features acquisition

A total of 335 patients underwent preoperative T1-weighted imaging (T1WI) and fat-suppressed T2-weighted imaging (FS-T2WI). Supplementary A2 displays the inspection equipment information and Table S1 displays the MRI scan parameters.

After drawing on previous studies, we selected six features from the MRI semantic features (Supplementary A3).

Tumor region delineation and radiomics feature extraction

The study flow chart is depicted in Fig. 1. ITK-SNAP (version 3.8.0; http://www.itksnap.org ) was employed to segment the region of interest (ROI) and evaluate the tumors in three dimensions. After segmentation, RIAS (version 0.2.1; https://pewter-papyrus-421.notion.site/RIAS-916ad7256e1e472985d4b11c8ebf0fe0) [22] was used to create peritumoral masks at a radial distance of 10 mm from the lesions in transverse and AP. Normal tissue, large arteries and veins, bronchi, and surrounding air were manually excluded. As Fig. 1(a) shows, the IT ROI corresponded to the maximum tumor area, the PT ROI to a radial distance of 10 mm from the lesion, and the WT ROI to the IT and PT regions combined.

Fig. 1

Preprocessing procedures of features extraction were shown in Supplementary A4. Radiomics features were then extracted using 3D Slicer software (version 4.10.2; https://www.slicer.org/). Finally, radiomics features, including first-order statistical, shape-based, textural, and wavelet decomposition features, were extracted from each three-dimensional ROI of the FS-T2WI and T1WI sequences. Textural features were included five classes (gray-level run-length matrix gray-level run-length matrix, gray-level dependence matrix, gray-level cooccurrence matrix, and neighborhood gray-tone difference matrix). On the basis of the ROIs, the following features combination were created: (1) IT features, consisting of radiomics features in the IT ROIs of T1WI and FS-T2WI; (2) PT features, consisting of radiomics features in the PT ROIs of T1WI and FS-T2WI; (3) WT features, consisting of radiomics features in the WT ROIs of T1WI and FS-T2WI; and (4) Combined features, consisting of both IT and PT features.

The inter- and intraobserver performance of the radiomics feature extraction process was assessed by calculating intraclass correlation coefficients (ICCs). Images from 40 patients were randomly chosen for segmentation by multiple radiologists. Inter-observer correlation coefficients were calculated by manually segmenting ROIs, performed by Reader 1, and intraobserver correlation coefficients were calculated by repeating the segmentation after 1 month, performed by Reader 2. Features with an ICC of < 0.80 were removed because they were deemed to have poor agreement. Among them, 40 T1WI features and 66 T2WI features were removed in the IT features; 10 T1WI features and 56 T2WI features were removed in the PT features; 11 T1WI features and 7 T2WI features were removed in the WT features.

“Combat compensation” method

The scanner effect is a major confounding factor in multi-center and multi-scheme studies that affects the extraction of radiomics features from MRI images [23]. Therefore, the combat compensation method was employed to eliminate the scanner effect.

Construction of radiomics signature

To remove the effect of varying gray values, all extracted radiomics features were normalized using z-scores. Because our feature pool had a high degree of dimensionality, feature selection was used to prevent overfitting. First, the 30 features with the strongest correlations and the least redundancy were selected by the minimum redundancy maximum relevance(mRMR) algorithm. Next, the feature parameters were further filtered using the least absolute shrinkage and selection operator (LASSO) regression algorithm (Fig. 2). Then, the following three machine learning classifiers were investigated: decision tree (DT), support vector machine (SVM), and logistic regression (LR). Three machine learning-predicted signatures were constructed for each feature subset, and a total of 12 machine learning-predicted signatures were built. Finally, the machine learning-predicted signature with the most accurate prediction results was selected, and its prediction score was inputted into Cox regression analysis to create the radiomics signature, which was used to obtain the radiomics score.

Fig. 2

(a) MRI feature selection using the least absolute shrinkage and selection operator regression algorithm. (b) The seven selected MRI features and their coefficients

Development of a clinical model and radiomics nomogram

Clinical information and MRI semantic features associated with STS progression were analyzed using univariate Cox regression, with factors significant at p < 0.05 considered significant independent risk factors for disease progression of STS patients. Such factor was included in the clinical model. Moreover, we integrated radiomics scores with selected clinical risk factors to develop a radiomics nomogram.

Validation and performance evaluation of the different models

The prognostic performance of machine learning-predicted signatures was evaluated on the basis of area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive, and negative predictive values. The ability of the clinical model, radiomics signature, and nomogram to predict the progression of STS patients was evaluated by the concordance index (C index) and the time-dependent receiver operating characteristic curve (T-ROC). The calibration curve was used to evaluate calibration ability. The integrated Brier score (IBS) was calculated using the “Boot632plus” splitting method to estimate the prediction error of the models. Decision curve analysis (DCA) was used to evaluate clinical usefulness. We used X-tile software (version 3.6.1; https://medicine.yale.edu/lab/rimm/research/software/) to identify optimal thresholds to classify patients into low- and high-risk groups on the basis of survival outcomes [24]. The Kaplan–Meier method and log-rank test were used to estimate the probability of PFS of the low- and high-risk groups.

Statistics

The baseline data were compared using Fisher’s exact test, the chi-square test (for categorical variables) and the Mann-Whitney U test, Student’s t test (for continuous variables). SPSS (version 25.0; IBM Corp., Armonk, NY, USA) and R software (version 4.2.2; www.r-project.org) were used for the statistical analyses. Two-sided p-values < 0.05 indicated statistical significance.