This retrospective study obtained the approval of the local institutional review board (No. 2020-10), and the requirement for informed consent was waived. The multicenter study was performed jointly by five centers: centers A, B, C, D, and E. Three hundred sixty-nine patients from the five centers were collected consecutively during January 2016 to December 2020, and all the patients had MR imaging examination before the treatment for EC. The included criteria were as follows: 1) histological diagnosed stage I EC; 2) basic MRI sequences including T1-weighted imaging (T1WI), T2-weighted imaging (T2WI) with fat saturation, diffusion-weighted imaging (DWI), apparent diffusion coefficient (ADC), and contrast-enhanced (CE) T1WI sequences; 3) the interval time between MRI examination and surgical less than 30 days. The exclude criteria were as follows: 1) patients with a history of other cancer (n = 2); 2) the imaging had apparent motion artifacts, or the included sequences could not match well (n = 16); 3) insufficient clinical information (n = 12). The patients from all the five centers were randomly divided into a training group (n = 226) and a test group (n = 113) in a ratio of 2:1. Clinical parameters, including age, tumor size, and cancer antigen 125 (CA125), were obtained through the review of clinical data. The workflow is shown in Figure 1 .

MRI was performed by using 3.0/1.5-T scanners with a phased-array abdominal coil. The patients lay in a supine position and breathed freely during the acquisition. The following sequences were obtained: T1WI, T2WI, DWI, and CE-T1WI after intravenous administration of gadopentetate dimeglumine at a dose of 0.2 mmol/kg of body weight in a rate of 2 to 3 ml/s. The scanning details are shown in Supplementary Table 1 .

All imaging was delivered from the PACS system and stored in DICOM format. An ROI was drawn around the visible tumor boundary manually in MITK Workbench (http://mitk.org/wiki/The_Medical_Imaging_Interaction_Toolkit(MITK)). The ROIs were drawn based on T2WI cross-sectionally and then referred to the T1WI, DWI (the tumor area shown as high signal in high b value sequences), ADC (the tumor area shown as low signal), and CE-T1WI (delayed phase). All ROIs were drawn by one radiologist (with a 6-year experience in pelvic imaging), who was blinded to the pathological diagnosis, and were rechecked by the other two radiologists (with 32 and 14 years of experience in pelvic imaging, respectively). Thirty days later, 50 patients were randomly selected and the tumors were redrawn by radiologist 1 and by radiologist 2 independently. The ROIs were drawn on each cross-sectional area to generate the volumetric region of interest (VOI, resampled to as isotropic voxels [3 × 3 × 3 mm]) ( Figure 2 ).

All the radiomics feature-extracting processes were performed in the Pyradiomics software (https://pypi.org/project/pyradiomics/). Radiomics features based on the VOIs of T1WI T2WI, DWI, ADC, and CE-T1WI were extracted. The images were normalized by subtracting from mean values and dividing by the standard deviation. A voxel array shift of 300 was used to make the gray-level values within a 0–600 range.

To eliminate the effect of different MRI scanning protocols and to improve the classification efficiency of the diagnostic models, a compensation method named “Combat” was used to realign feature distributions computed from different MRI equipment and protocols (16).

The inter- and intraclass correlation coefficients (ICCs) of the extracted features were calculated to assess the reproducibility of radiomics features. Features with both inter- and intraclass ICCs less than 0.75 were considered non-stable features and eliminated. Pearson’s correlation was used to identify redundant features. If two features had a Pearson’s correlation coefficient 0.9, the one with a larger mean absolute coefficient was eliminated.

The Synthetic Minority Oversampling Technique (SMOTE) method was used because of unbalance of positive/negative LVSI samples in the training group. Positive LVSI (minority class) was oversampled and negative LVSI (majority class) was undersampled to balance the training set to improve the classification performance.

Then, a Least Absolute Shrinkage and Selection Operator (LASSO) was used to select non-zero coefficient features associated with LVSI with 10-fold cross-validation in early-stage EC patients. A radscore was calculated for each patient by linear combination of the final selected features.

Clinical parameters of LVSI including age, tumor size, and CA125 were analyzed by using multivariate logistic regression. Clinical risk factors were identified if the parameters were statistically significant in the analysis. A radiomics nomogram was constructed by using multivariable logistic regression analysis which combined the selected radiomics features and the independent clinical risk factors. To assess the discrimination performance of the radiomics nomogram, the area under the curve (AUC) of the receiver operating characteristic (ROC) curve was calculated. Calibration curves were plotted to evaluate the calibration performance of the radiomics nomogram.

The performances of the radiomics signature, clinical risk factors, and radiomics nomogram were validated in the test group. The sensitivity, specificity, and AUC were also calculated.

By quantifying the net benefits at different threshold probabilities in the training group, decision curve analysis (DCA) was performed to determine the clinical usefulness of the radiomics nomogram.

Statistical analysis was performed by using R software (version 4.0.5; http://www.Rproject.org). An independent sample t-test was used to compare the differences in continuous variables (age, tumor size, and CA125). Chi-squared test was used to compare the differences in categorical variables (FIGO stage, LVSI status). The association between LVSI and clinical risk factors was assessed using Spearman’s correlation. The “glmnet” package was used for LASSO and logistic regression, the “ComBatHarmonization” package was used for Combat, the “DMwR” package was used for SMOTE, the “rms” package was used for nomogram calculation, “pROC” package was used for AUC, and the “dca.R” package was used for DCA. ROC curve analysis was performed to calculate the AUC and corresponding 95% confidence interval. DeLong test was used to compare the performance of clinical risk factors, radscore, and nomogram. A P value less than 0.05 was considered statistically significant.

Three hundred thirty-nine patients were finally included with 54 positive LVSI and 285 negative LVSI (mean age, 56.8 years; range, 25–89 years). All patients underwent total hysterectomy and bilateral salpingo-oophorectomy. Among these, 216 were G1, 75 were G2, and 27 were G3; 277 were superficial or no myometrial invasion, and 62 were deep myometrial invasion. The patients’ clinical risk factors as well as the stage and tumor characteristics of each subgroup are presented in Table 1 .

A total of 358 radiomics features were extracted, including 14 shape features, 72 first-order features, and 272 texture features. After removing features with either inter- and intraclass ICC 0.75 and Pearson’s correlation coefficients 0.9, 231 and 96 features were retained, respectively. LASSO analysis finally included 15 radiomics features, which were defined as the radiomics signatures ( Figure 3 ). The radscore calculation was as follows:

Radscore = 0.16814 + 0.16823×shape_LAL + -0.05128×shape_MAL + 0.00899×T2WI_firstorder_10P + 0.01112×T2WI_firstorder_TotalEnergy + 0.03609×T2WI_glcm_Imc1 + -0.09119×T2WI_glszm_GLNU + -0.01004×DWI_firstorder_10P + -0.0668×DWI_glcm_CP + -0.04433×DWI_glrlm_RLNUN + -0.03675×DWI_glrlm_RunVariance + 0.09361×DWI_glszm_GLV + 0.01592×DWI_glszm_ZP + 0.04461×CE_firstorder_10P + -0.04113×CE_firstorder_IR + -0.06052×ADC_firstorder_Median

Multivariate logistic regression analysis showed that age, CA125, and tumor size were risk factors of LVSI in the early-stage EC. On considering that the selected feature “shape_LAL” and “shape_MAL” were the same as tumor size, we did not include tumor size in the nomogram for avoiding overfitting. Therefore, a radiomics nomogram was constructed by integrating the radiomics signatures, age, and CA125 ( Figure 4 ).

The prediction performance of radscore was 0.88 (95% CI: 0.81–0.94) and 0.82 (95% CI: 0.72–0.93) in the training group and test group, respectively. The prediction performance of clinical risk factors was 0.72 (95% CI: 0.63–0.81) and 0.71 (95% CI: 0.57–0.86) in the training group and test group, respectively. The prediction performance of the radiomics nomogram was 0.89 (95% CI: 0.83–0.95) and 0.85 (95% CI: 0.75–0.94) in the training group and test group, respectively. The AUC, specificity, sensitivity, positive predictive value, and negative predictive values are shown in Table 2 .

The DCA showed that both the radiomics signatures and radiomics nomogram could add net benefit to the patients to predict the presence of LVSI preoperatively ( Figure 5 ). When the threshold probability was within a range from 20% to 100%, the net benefit of using the nomogram to predict LVSI was more than the treat-all or treat-none scheme.