This retrospective study was approved by the local ethics committee, and informed consent was not required. A total of 461 patients with EC confirmed by postoperative histology from June 2015 to July 2022 who underwent preoperative MR examination were enrolled. The inclusion criteria were as follows: (1) confirmation of EC through pathology; and (2) presence of clinical and histopathological characteristic data, including age, serum CA125 level, tumor grade, depth of myometrial invasion (MI), and cervical stromal invasion. The exclusion criteria were as follows: (1) total hysterectomy not performed within 2 weeks after the MRI examination; (2) preoperative treatment with chemoradiation; (3) tumors too small to be visible on MRI; (4) images with obvious motion artifacts; (5) contraindications for MRI examination; (6) incomplete clinical data; and (7) EC occurring simultaneously with other malignant tumors. Eighty-seven patients were excluded Figure 1 ), and the remaining 374 patients were included (mean age: 54.3 ± 8.1 years). Surgical staging of all patients included total hysterectomy with bilateral salpingo-oophorectomy, and lymph node (LN) assessment included pelvic lymphadenectomy and accompanying paraaortic lymphadenectomy. Compared with endometrioid adenocarcinoma (EEA), non-EEA (including carcinosarcomas, mucinous carcinoma, serous carcinoma, clear cell carcinoma, mixed carcinoma, and undifferentiated carcinoma) has a greater malignancy rate and is more prone to LNM (28). Therefore, we grouped non-EEA patients in similar proportions into training, test, and independent validation cohorts on the basis of the proportion of non-EEA patients in the overall cohort (10.7%, 40/374) ( Table 1 ). Furthermore, patients with different field strengths were evaluated, with those who underwent 1.5-T MR accounting for 62.3% of the total cohort. Therefore, in the independent validation cohort, a similar proportion (61.6%) was also used for the 1.5-T dataset. The steps for data segmentation were as follows. First, 60 patients (37 patients who underwent 1.5-T MR and 23 who underwent 3.0-T MR) were randomly selected as the independent validation group. The remaining 314 patients were randomly divided into a training cohort (n = 220) and a test cohort (n = 94) at a ratio of 7:3. All patients were recruited from a single center. Notably, the independent validation cohort was not involved in the model training and testing phases.

The MRI scans were carried out via a 1.5-T (EXCELART Vantage™ powered by Atlas, Canon Medical Systems Corp., Tochigi, Japan) or 3.0-T (Siemens Magnetom Skyra, Erlangen, Germany) scanner, along with an 8-channel phased-array abdominal coil. To minimize artifacts resulting from intestinal motility, 20 mg of raceanisodamine hydrochloride (Hangzhou People’s Livelihood Pharmaceutical Co., Hangzhou, Zhejiang, China) was intravenously injected prior to scanning. All MRI sequences were obtained following a standard protocol (refer to Supplementary Table 1 for specifics). For diffusion-weighted imaging (DWI), b = 0 and 650 s/mm² were used for the 1.5-T scanner, and b = 0 and 1000 s/mm² were used for the 3.0-T scanner. The ADC maps were automatically reconstructed by the postprocessing workstation.

Image preprocessing was performed via a standard workflow (see the Supplementary Materials for details). Image segmentation was performed by two experienced radiologists (T.Z. and B.Y.). One month after the initial segmentation, 100 patients were randomly selected for tumor segmentation by another radiologist (Y.D.) for interreader reliability evaluation. The details of the lesion segmentation are shown in the Supplementary Materials . 3D Slicer software (version 4.10.2; https://download.slicer.org/) was used for the manual whole-tumor segmentation. Three volumes of interest (VOIs) were chosen ( Figure 2 ) as follows: (1) intratumoral region—ROIs were delineated along the edge of the lesion slice-by-slice, including areas of hemorrhage and necrosis, and the normal anatomical structure was avoided; the ROIs were fused into a three-dimensional (3D) VOI; (2) peritumoral margin—according to the previous literature (29), the tumor contour was automatically expanded by 3 mm to generate dilated VOIs; when the dilated VOIs were beyond the scope of the uterus, they were manually corrected for their boundaries to be set to the uterus edge; when other lesions in the myometrium (like fibroids, adenomyosis, etc.) were involved in the dilated VOI, manual correction was used; the peritumoral region = dilated VOI - tumor VOI; and (3) combined intra- and peritumoral regions.

The preprocessing of images and extraction of features were carried out via Artificial Intelligence Kit (AK, Version 3.3.0, GE Healthcare) software. Radiomics features, comprising first-order, shape-based, and texture features, were obtained from each of the VOIs. The methodology for extracting the radiomics parameters is shown in Figure 2 .

TS was defined as the maximum tumor diameter measured in three orthogonal planes: the transverse (x) and anteroposterior (y) diameters on oblique axial T2W images and the craniocaudal (z) diameter. APsag was measured on sagittal T2W images ( Figures 3A, B ). The whole-tumor volume on T2W images for the TV analysis was automatically calculated via 3D Slicer software. The TAR was calculated (the measurement method is detailed in the Supplementary Material ) ( Figures 3C, D ) via the following equation from a previous study: TAR = (area of tumor/area of uterus) × 100% (18).

The statistical analysis was conducted via the R language (version 4.2.0, https://www.r-project.org) and Python language (version 3.9, https://www.python.org). A binary classification model was developed to predict the LNM status as “LNM-positive” or “LNM-negative”. Clinical data were subjected to both univariate and multivariate logistic regression (LR) analyses for filtration. Radiomic features were evaluated via t tests, Fisher’s exact tests, chi-square tests and, when applicable, the Mann−Whitney U test. P <0.05 was considered to indicate statistical significance. Least absolute shrinkage and selection operator (LASSO) regression was used to identify the most important radiomic features. LR was utilized for training the prediction models and was employed to create a nomogram incorporating clinical and tumor morphological parameters along with the radiomics score (Radscore). The independent validation cohort was used only to evaluate model performance. The sample size was calculated according to the previous literature (30); the number of events per variable was 10 or greater in the setting of the multivariate LR model. Before the initiation of data modeling, radiomic features were subjected to standardization via the standard scalar method. To determine the consistency among readers in evaluating MR morphological and radiomics features, intraclass correlation coefficients (ICCs) were calculated, with an ICC value exceeding 0.8 suggesting nearly perfect agreement. The R codes used for modeling and data analysis are provided in the Supplementary Materials .

Approximately 70% of EC patients are diagnosed with stage I disease (31). The risk of LNM in EC patients with low-grade and superficial MI is 3–5%, whereas the risk of LNM in high-grade patients is approximately 16–22% (32–34). Therefore, the proportion of LNM in EC is relatively low. In this study, approximately 10.7% (40/374) of patients had LNM. To address the imbalance between LNM-positive and LNM-negative patients in our training cohort, we employed the synthetic minority oversampling technique (SMOTE) to generate synthetic samples in the minority (positive) class. SMOTE works by selecting two or more similar instances (using a distance measure) in the minority class and generating new instances that lie between these instances in the feature space (35, 36). The SMOTE has been applied in previous studies on LNM in EC (24, 37).

The performance of the prediction model was evaluated via receiver operating characteristic (ROC) curve analysis. The DeLong test was employed to assess whether the difference in the ROC curves between the two models was statistically significant. The calibration curve was assessed via the Hosmer−Lemeshow (HL) test, with a P value greater than 0.05 indicating satisfactory predictive performance. Decision curve analysis (DCA) was used to compare the net benefits of the clinical models and radiomics nomogram models, with all ROC curve cutoff values determined by the maximum Youden index. The areas under the curve (AUCs), accuracy (ACC), sensitivity (SEN), and specificity (SPE) were then calculated.

Table 1 summarizes the clinical and MRI morphological findings of the EC patients in the training, test, and independent validation cohorts. The 374 EC patients (1.5-T, n = 233; 3.0-T, n = 141) included 40 LNM-positive patients (40/374, 10.7%) and 334 LNM-negative patients. Forty non-EEA patients were included in this study (mixed carcinoma = 16, carcinosarcoma = 14, undifferentiated carcinoma = 4, serous carcinoma = 3, and clear cell carcinoma = 3) and were assigned to the training (24/220, 10.9%), test (9/94, 9.6%), and independent validation (7/60, 11.7%) cohorts in similar proportions.

From each VOI, 1036 features were extracted, including ADC_Intratumoral, ADC_Peritumoral, T2WI_Intratumoral, and T2WI_Peritumoral. After feature selection, 9 features remained and were selected for evaluating LNM status in EC (forming the Radscore, Table 2 ) (refer to the Supplementary Material for a breakdown of the feature extraction process). There was no significant difference in the features between the training and test sets ( Table 3 ). After the sample imbalance was adjusted with the SMOTE, the training cohort included 120 LNM-positive patients and 196 LNM-negative patients ( Figure 4 ).

All the morphological parameters and radiomics features showed excellent interreader reliability, with ICC values ranging from 0.908 to 0.997. This high level of agreement ensures the repeatability of the model for future clinical use. The details are shown in Supplementary Tables 4 - 6 .

After univariate analysis, the CA125 level, tumor grade, TV, TS, APsag, and TAR were significantly different between LNM-positive and LNM-negative patients (all P < 0.05, Table 1 ). Among these features, the tumor grade is a postoperative pathological result that can be obtained through preoperative endometrial biopsy or diagnostic curettage. Multivariate binary LR analysis revealed that the TAR and CA125 level were independent predictors of LNM in patients with EC (all P < 0.05, Table 4 ). The AUCs of the clinical model for predicting LNM in the training, test, and independent validation cohorts were 0.674 (95% confidence interval [CI]: 0.587–0.763; SEN: 68.4%, SPE: 58.7%), 0.668 (95% CI: 0.636–0.741; SEN: 75.0%, SPE: 64.2%) and 0.618 (95% CI: 0.567–0.654; SEN: 54.6%, SPE: 70.0%), respectively.

The performance of ADC mapping and T2WI in distinguishing LNM-positive patients from LNM-negative patients is summarized in Supplementary Table 7 . Interestingly, after applying LASSO regression with the peritumoral features from the ADC map, all feature coefficients decreased to zero. This implies that under the LASSO constraint, none of the features were identified as having substantial predictive power for the outcome. Therefore, ADC_Peritumoral features were excluded. Among the four radiomics models (ADC_Intratumoral, T2WI_Intratumoral, T2WI_Peritumoral, and T2WI_Intratumoral+Peritumoral), the combined intratumoral and peritumoral features from T2WI achieved the best prediction performance (AUC_training = 0.819; AUC_test = 0.771; and AUC_independent-validation = 0.772).

We combined the features of ADC mapping (intratumoral region) and T2WI (intra- and peritumoral regions) to construct the best prediction model (named hybrid-feature, Table 5 ), which showed the best classification performance (AUC_training = 0.850; AUC_test = 0.838; and AUC_independent-validation = 0.836). In this study, the Radscore consisted of 9 features from the hybrid-feature model.

The clinical model and the Radscore were combined, and a clinical–radiomics mixed model was constructed. The CA125 level, TAR, and Radscore were used to develop the nomogram in the training cohort ( Figure 5A ). The AUCs of the nomogram model for predicting LNM in the training ( Figure 5B ), test ( Figure 5C ) and independent validation ( Figure 5D ) cohorts were 0.878 (95% CI: 0.823–0.907; ACC: 81.0%, SEN: 84.0%, and SPE: 77.9%), 0.877 (95% CI: 0.831–0.914; ACC: 72.0%, SEN: 83.3%, and SPE: 71.0%) and 0.864 (95% CI: 0.815–0.901; ACC: 85.3%, SEN: 75.0%, and SPE: 86.0%), respectively. The formula was as follows:

The calibration curves, with HL scores of 0.481, 0.346 and 0.226 for the training, test, and independent validation cohorts, respectively, revealed that the nomogram was reasonably accurate in predicting LNM in EC patients ( Figures 6A–C ). The DeLong test yielded a p value of 0.03 when the radiomics model was compared with the nomogram and a p value of 0.01 when the clinical model was compared with the nomogram. These results indicate that the differences in outcomes between the nomogram and both the radiomics and clinical models are statistically significant. DCA indicated that the radiomics nomogram produced greater net benefit than the clinical model for predicting LNM in EC patients in the training ( Figure 6D ), test ( Figure 6E ), and independent validation cohorts ( Figure 6F ). The reclassification measures confirmed that the nomogram performed better than the radiomics and clinical models did, with a net reclassification index (NRI) of 0.208 (95% CI, 0.112–0.287) when the nomogram and radiomics model were compared and an NRI of 0.386 (95% CI, 0.264–0.482) when the nomogram and clinical model were compared.