This retrospective study was approved by the Ethics Committees of the respective institutions, with informed consent waived due to its retrospective nature. Prior to analysis, all patient data was deidentified to ensure the confidentiality and anonymity of personal information.

We identified a cohort of 311 patients from four medical centers who underwent gynecological surgery, including 111 with USC and 200 with EEC. The participating centers were as follows: Shantou Central Hospital (Institution I), Sun Yat-Sen Memorial Hospital (Institution II), Sun Yat-Sen University Cancer Center (Institution III), and Cancer Hospital of Shantou University Medical College (Institution IV). The specific data collection timelines for each institution and histological subtype are detailed in Supplementary Table 1 . The inclusion criteria required (a) USC and EEC confirmed surgically and pathologically; (b) a pelvic MRI conducted within 14 days before gynecological surgery. The exclusion criteria encompassed: (a) maximum tumor diameter under 1 cm; (b) incomplete MRI examination; (c) incomplete pathology report; (d) presence of mixed cellular components and (e) history of neoadjuvant therapy. Ultimately, a total of 210 patients were included in the study, comprising 68 with USCs and 142 with EECs. Patients from Institution I and II were randomly assigned to a training cohort (100 patients) and an internal test cohort (44 patients) in a 7:3 ratio. A total of 66 patients were included as an external test cohort by Institutions III and IV. Figure 1 illustrates the flowchart of the patient recruitment process.

MRI was performed using either a 3.0-T or 1.5-T scanner with a pelvic phased-array surface coil. Institutions I and II utilized Siemens Magnetom Verio (3.0-T) and Siemens Magnetom Area (1.5-T) scanners, while institutions III and IV employed Siemens Magnetom Avanto (1.5-T) and GE Medical System Discovery HD750 (3.0-T) scanners. The sequences obtained included axial and sagittal T2-weighted imaging(T2WI), diffusion-weighted imaging (DWI) with a b-value of 800 or 1000 s/mm², and axial and sagittal contrast-enhanced MRI (CE-MRI). CE-MRI was conducted following the administration of gadolinium chelate (Gadovist, Bayer) at a dosage of 0.2 mmol/kg body weight. The detailed MRI acquisition protocols are summarized in Supplementary Table 2 .

Clinical data were collected from medical records, encompassing age, body mass index (BMI), menopausal status, obstetric history, family history of malignancy, diabetes history, International Federation of Gynecology and Obstetrics (FIGO) stage (2023), tumor markers (CA-125, CA-199, CEA, HE4), and details of myometrial and cervical stromal invasion, adnexal involvement, parametrial invasion, lymph node metastasis, and presence of abnormal ascites. For subsequent modeling, tumor grade was categorized as follows: (a) low grade, comprising FIGO grades 1 and 2 endometrioid carcinoma, and (b) high grade, consisting of FIGO grade 3 endometrioid carcinoma or uterine serous carcinoma (8). Additionally, in accordance with European Society for Medical Oncology guidelines, FIGO stage was categorized into early (IA) and advanced (IB or higher) stages for risk stratification (28). For the purpose of baseline characterization and analysis in this study, FIGO stage and histopathologic grade were determined based on the preoperative endometrial biopsy or D&C results, reflecting the diagnostic information available at the time of initial clinical decision-making.

Two experienced radiologists, LP.L. (Reader 1) with 5 years of experience and Y.S. (Reader 2) with 8 years of experience in gynecologic imaging, independently assessed the multiparametric MR images without access to medical records or pathological data. They assessed lesion characteristics including location, borders, growth patterns, diffuse distribution, presence of necrosis and hemorrhage, tumor largest diameter, tumor volume (calculated as d1×d2×d3×π/6, where d1 and d2 are measured along and perpendicular to the uterine long axis in the sagittal plane, and d3 is the largest lateral diameter in the axial plane). Additionally, they assessed signal intensity ratios (SIR) of the tumor and gluteus maximus on T2WI, DWI, and CE-T1WI, enhancement patterns on CE-T1WI, homogeneity, and the ratios of endometrial thickness (ET) to the largest longitudinal and anteroposterior (AP) dimensions of the uterus on T2WI sagittal images (12, 29, 30)( Supplementary Figure 1 ). Features were evaluated independently by two radiologists, and any discrepancies were resolved by consensus. The inter-observer agreement for the qualitative clinical-radiological features was assessed using Cohen’s kappa (κ) statistic, and for continuous variables, the intraclass correlation coefficient (ICC) was used ( Supplementary Table 3 ).

Figure 2 provides an overview of the study’s pipeline. The region of interest (ROI) was manually delineated along the lesion’s edge using ITK-SNAP software on T2WI, DWI, and CE-T1WI at the delayed phase, ensuring minimal inclusion of normal tissue to acquire comprehensive tumor data. Each tumor’s volumetric region of interest (VOI) was segmented. All ROIs drawing were performed by two experienced radiologists (Reader 1 and Reader 2) blinded to the patients’ histopathology. With 3-month intervals, 30 patients were randomly selected for Reader 2 to repeat the tumor ROI drawing. The inter-/intra-observer variability of the extracted features was assessed by ICC test. ICC > 0.75 indicated satisfactory agreement.

Radiomics analysis was conducted using PyRadiomics version 3.0.1, employing VOIs from T2WI, DWI, and delayed phase CE-T1WI. Prior to feature extraction, each image sequence was normalized by centering the gray values at the mean and scaling them according to the standard deviation, which effectively minimized variations caused by different scanners, scanning parameters, and protocols. A total of 535 radiomics features were extracted from various MRI images (T2WI, DWI, CE-T1WI), comprising 70 shape features, 90 first-order histogram features, and texture features including 120 grey level cooccurrence matrix (GLCM), 80 grey level run length matrix (GLRLM), 80 grey level size zone matrix (GLSZM), 25 neighboring grey tone difference matrix (NGTDM), and 70 grey level dependence matrix (GLDM). The study design adhered to the reporting guidelines of the Image Biomarker Standardization Initiative (IBSI) (31).

DL features were extracted utilizing a pre-trained Resnet50 convolutional neural network (CNN) model. Before extracting DL features, the data undergoes processing through these steps: (1) select the mask with the largest ROI in the labeled MRI; (2) crop MRI images using minimal bounding rectangles; (3) resize the tumor patch to 224 × 224 pixels. The Resnet50 network was initially pre-trained on the ImageNet dataset, followed by transfer learning on the training set. Upon completing Resnet50 training, we extracted 2048 deep learning features from each patch using the penultimate average pooling layer of the model. The features were then compressed to a set of 64 features using principal component analysis (PCA). Eventually, a total of 320 DL features was extracted from all series. Gradient-weighted class activation mapping (Grad-CAM) was employed to enhance model transparency and explore interpretability through visualization.

We applied z-score normalization to all features and removed those with constant values. Radiomics signatures with an ICC greater than 0.75 were initially screened using the Spearman correlation test. We retained one feature for further analysis when the Spearman correlation coefficient between two features exceeded 0.9. These features were further screened using the least absolute shrinkage and selection operator (LASSO). The regularization parameter (λ) was tuned using the one-standard error of the minimum criteria (1-SE criteria) alongside tenfold cross-validation-based feature selection (see Supplementary Figure 2 ). Following feature selection, the synthetic minority oversampling technique (SMOTE) algorithm was employed on the training set, but using only the features selected by LASSO, to balance the minority class samples for the subsequent model training step.

A SVM (support vector machine) algorithm was employed to construct seven models including a clinical-radiological model utilizing clinical and radiological data, a radiomics model using radiomics features, a DL model leveraging deep learning features, a CR model combining clinical-radiological and radiomics features, a DLR model integrating radiomics and deep learning features, a CDL model combining clinical-radiological and deep learning features, and a comprehensive all-combined model incorporating all selected features. All feature integrations were performed through direct concatenation (feature-level fusion) to maximize information utilization.

The models were developed in the training set and validated with both internal and external test sets. Model predictive performance was evaluated via a receiver operating characteristic (ROC) curve, with results presented as the area under the curve (AUC) and corresponding 95% confidence interval (CI). The accuracy (ACC), sensitivity (SEN), specificity (SPEC), and F1 score were determined using the cut-off value that maximizes the Youden index from the ROC curve analysis.

Characteristics were compared using the independent t-test or Mann–Whitney U test for continuous variables, and Fisher’ s exact test or χ (2) test for categorical variables, with p-values adjusted via the Benjamini-Hochberg correction. The DeLong test was employed to compare the AUCs. Decision curve analysis (DCA) evaluated the models’ clinical utility by analyzing net benefit across various threshold probabilities in the testing sets. Statistical analyses were conducted using Python (version 3.9; https://www.python.org/), R (version 4.1.2; https://www.r-project.org/) and SPSS (version 26.0; https://www.ibm.com/). Statistical significance was defined as a two-sided p-value < 0.05. The Benjamini-Hochberg procedure was used to adjust for multiple testing. To assess the adequacy of the achieved sample size, a post hoc power analysis was conducted using G*Power software (version 3.1.9.7).

This study enrolled 210 patients: divided into a training set of 100, an internal-testing set of 44, and an external-testing set of 66. The post-hoc power analysis demonstrated a statistical power of 87%, confirming that our sample size is sufficiently. A comparison between preoperative biopsy and final surgical pathology revealed discordance in 7 of 210 cases (3.3%), wherein the final diagnosis was of a higher grade or more aggressive histologic subtype than initially determined by biopsy. Table 1 details patient characteristics within the USC and EEC groups across different cohorts. The age and proportion of postmenopausal patients were higher in the USC group compared to the EEC group (p < 0.05), patients with USC usually presented with higher HE4 level, FIGO staging and histopathologic grade (p < 0.05). Significant disparities were also observed between USC and EEC groups in terms of ET/AP ratio, tumor border, infiltrative growth pattern, diffuse distribution, presence of necrosis, inhomogeneity, heterogenous enhancement, deep myometrial invasion, cervical stromal invasion, adnexal involvement and pelvic lymph node metastasis (all p < 0.05).

Among the 17 clinical-radiological characteristics, histopathologic grade, FIGO staging, ET/AP ratio and diffuse distribution were identified as significant features using the LASSO algorithm ( Supplementary Figure 3A ). The mean inter- and intra-observer reliabilities were 0.821 (95% CI 0.726–0.896) and 0.859 (95% CI 0.773–0.912), indicating excellent consistency in radiomics features. A total of 194 radiomics features and 160 DL features of the tumor, each with Spearman correlation coefficients > 0.9, were retained for further selection. Using LASSO algorithms, 30 radiomics features and 14 DL features were selected to construct the radiomics, DL, and combined models. Supplementary Figure 3 provides additional information on the features chosen by the LASSO algorithm.

The SVM model was optimized using the training set and subsequently evaluated on both internal and external test sets. Figures 3A–C displays the predicted scores for patients, demonstrating the models’ strong classification capability. Table 2 presents the performance metrics of various models on both the training and testing datasets. The clinical-radiological model achieved AUCs of 0.861 (95% CI: 0.747-0.975) and 0.700 (95% CI: 0.552-0.848) in the internal and external testing set, respectively. The AUCs of the radiomics model were 0.934 (95% CI: 0.862-0.999) and 0.750 (95% CI: 0.632-0.868) in the internal and external testing set, respectively. The AUCs of the DL model were 0.869 (95% CI: 0.757-0.980) in the internal-testing set, and 0.704 (95% CI:0.572-0.835) in the external-testing set. The all-combined model showed excellent predictive performance. The all-combined model demonstrated superior classification performance in the internal-testing set with an AUC of 0.957 (95% CI: 0.904-1.000), accuracy of 0.886, sensitivity of 0.923, specificity of 0.833, and F1 score of 0.906, while in the external-testing set, these values were 0.880 (95% CI: 0.800-0.961), 0.742, 0.636, 0.955, and 0.767, respectively.

DeLong’s test indicated that the all-combined model demonstrated significantly superior discriminatory ability compared to both the clinical-radiological model (AUC = 0.880 vs. 0.700, p < 0.05) and DL model (AUC = 0.880 vs. 0.704, p < 0.05) in the external-testing set ( Figure 3E ; Supplementary Figure 4 ).

The all-combined model demonstrated significantly superior discriminatory power compared to the CR model (AUC = 0.880 vs. 0.810, p < 0.05) and CDL model in the external-testing set (AUC = 0.880 vs. 0.688, p < 0.05) (refer to Table 2 ; Supplementary Figure 4 ). The DLR model demonstrated superior predictive performance compared to the clinical-radiological model, although the differences were not statistically significant in both the internal-testing set (AUC = 0.908 vs. 0.861, p = 0.504), and the external-testing set (AUC = 0.767 vs. 0.700, p = 0.499) ( Figures 3D, E ; Supplementary Figure 4 ). Accuracy, sensitivity and specificity values varied across models, with the best performance in combined models such as DLR model (accuracy of 0.980, sensitivity of 0.972 and specificity of 1.000 in training) and all-combined model (accuracy of 0.742, sensitivity of 0.923 and specificity of 0.833 in the external test set). These models consistently outperformed individual models like R model (sensitivity of 0.652 in the external test set) and C model (specificity of 0.647 in the external test set). The all-combined model and DLR achieved the highest F1 scores, with the all-combined model attaining 0.979 during training and 0.906 in the internal test set. The decision curves ( Figures 3F, G ) demonstrated that the combined model provided a superior overall net benefit across most reasonable threshold probabilities in both the internal and external testing sets. Figure 4 illustrates the activation maps highlighting image regions that significantly contribute to the feature output recognized by the deep CNN. Overall, the use of a multiparametric model based on radiomics and DL had better predictive value in the preoperative differential diagnosis between USC and EEC.