This multicenter, retrospective study was conducted between 2022 and 2024 at two research centers: Center 1 (The First Affiliated Hospital of Guangzhou University of Chinese Medicine) and Center 2 (Guangzhou Red Cross Hospital). Center 1 contributed a total of 351 patients, including 81 EC cases and 270 non-EC cases, which were used as the training set. Center 2 provided a total of 260 patients for external validation, comprising 51 EC cases and 209 non-EC cases. Multimodal ultrasound images—including two-dimensional and color Doppler images—as well as clinical data were collected from patients with pathologically confirmed EC. This study was conducted in accordance with the Declaration of Helsinki, and approved by the institutional review board (IRB). The requirement for informed consent was waived due to the retrospective nature.

The inclusion criteria are as follows: (1) Patients who underwent endometrial aspiration biopsy, curettage, or hysterectomy with pathologically confirmed diagnoses between 2022 and 2024. (2) Preoperative transvaginal color Doppler ultrasound performed according to standardized protocols. Exclusion criteria are as follows: (1) Poor-quality ultrasound images or absence of preoperative transvaginal color Doppler ultrasound. (2) Incomplete clinical data. (3) History of prior radiotherapy, chemotherapy, or multiple endometrial surgeries. (4) Use of hormone therapy for endometrial hyperplasia or autoimmune diseases. (5) Diagnosis of cervical cancer or other malignancies. (6) Presence of intrauterine devices obstructing endometrial visualization. Figure 1 presents the flowchart of the study population selection.

In the training set, 81 EC cases were included, with 127 two-dimensional images and 138 color Doppler images, while 270 non-EC cases contributed 299 two-dimensional images and 281 color Doppler images, resulting in a total of 845 images. For the external validation set, 51 EC cases were included, with 56 two-dimensional images and 73 color Doppler images, while 209 non-EC cases contributed 232 two-dimensional images and 237 color Doppler images, amounting to 598 images. Overall, the dataset comprised a total of 1,443 ultrasound images.

Transvaginal ultrasound (TVU) examinations were performed by three experienced physicians using various scanner models, including GE Voluson E6, GE Voluson E8, GE Voluson E10, Hitachi Avius L, Philips EPIQ 5, and Toshiba Aplio500. All systems were equipped with high-frequency (5–14 MHz) transvaginal probes. The examining physicians possessed extensive experience (>15 years) in obstetric and gynecologic ultrasound and strictly adhered to standardized examination and measurement techniques as outlined in the IETA consensus statement (16). Standard two-dimensional TVU and color Doppler ultrasound images, specifically showing the uterine endometrium and any endometrial lesions, were acquired.

In this study, we developed DL models based on two imaging modalities: two-dimensional (2D) ultrasound and color doppler ultrasound. The training set included Research Center 1 (n=351; images=845), while Research Center 2 (n=260; images=598) was designated as the external validation set. We selected four distinct and well-established convolutional neural network (CNN) architectures: ResNet-50 (17), ResNet-152 (18), EfficientNet-B0 (19), and DenseNet-201 (20). These architectures are known for their strong performance on natural image classification tasks. To expedite the training process, we employed transfer learning by freezing the pre-trained convolutional layers and only training the fully connected layers. Each architecture was trained separately on both imaging modalities. The model demonstrating the best performance, as validated by an external validation set, was then selected as the final DL model.

Before training, preprocessing techniques such as image normalization, resizing (to 512×512 pixels), and data augmentation (including random vertical and horizontal flipping, rotation, grayscale transformation, and adjustments to brightness, contrast, saturation, and hue) were applied to reduce overfitting and enhance training performance. Each model was trained using five-fold cross-validation, and hyperparameters such as batch size (16), learning rate (3e-5), and the number of epochs (200) were optimized. The Adam optimizer, with β1 of 0.9, β2 of 0.999, epsilon of 1e-8, and weight decay set to 0.01 for L2 regularization, was employed to prevent overfitting. Considering the EC-to-non-EC ratio is significantly imbalanced, we employed a class-weighted cross-entropy loss function. Specifically, a higher weight was assigned to the EC class. By increasing the cost of misclassifying an EC case, the loss function effectively forces the model to pay more attention to these rare but critical instances, thus improving its ability to correctly identify. After training, the model’s performance was evaluated on the external validation set using key metrics such as the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and F1 score. The loss and accuracy curves for both imaging modalities are presented in Supplementary Figure S1 .

Since clinical diagnoses are typically made at the patient level, some patients in this study had multiple ultrasound images (ranging from 1 to 8 per patient), which could introduce bias. To mitigate this, we averaged the image-level risk scores to generate a patient-level score, thereby reducing potential bias associated with multiple images per patient.

In addition to imaging data, we collected baseline clinical variables, including age, body mass index (BMI), history of gestation, menopausal status, presence of irregular vaginal bleeding, and history of hypertension, diabetes, hypothyroidism, and polycystic ovary syndrome (PCOS). Univariate logistic regression was performed to identify statistically significant clinical predictors (P < 0.05), which were then subjected to multivariate regression analysis. Variables that remained significant in the multivariate analysis were incorporated into the final clinical model for EC prediction.

A combined model was developed by integrating the best-performing CNNs models from the two-dimensional ultrasound and color Doppler modalities with the clinical prediction model. This hybrid approach aimed to enhance the accuracy and reliability of EC prediction.

Statistical analysis was conducted using R software (version 4.2.2, https://www.r-project.org/). For continuous variables, the Shapiro-Wilk test was used for assessing distribution. If the data followed a normal distribution, parametric tests were applied; otherwise, non-parametric tests were used. Categorical variables were analyzed using the chi-square test or Fisher’s exact test (where applicable). Categorical and continuous variables were presented as frequencies (percentages), mean ± standard deviation, or median (interquartile range) as appropriate. Gradient-weighted Class Activation Mapping (Grad-CAM) is utilized for visualizing and understanding the decision-making process within DL. The SHapley Additive exPlanations (SHAP) method is employed to explain the influence and contribution of features on the fused model’s output. The construction, training, and validation of the deep convolutional neural network model were carried out using the PyTorch framework (version 1.13.0, https://pytorch.org/).

The diagnostic models were evaluated using a range of performance metrics, including the area under the AUC with its 95% confidence interval, accuracy, sensitivity, specificity, and F1 score. The Delong test was employed to compare the performance of different models and assess statistical significance. Model fitness was determined using the Hosmer-Lemeshow goodness-of-fit test and calibration curves. Additionally, decision curve analysis (DCA) was performed to evaluate the net clinical benefit of the predictive models. A nomogram was constructed to visualize the combined model. Furthermore, several ultrasound specialists were involved in assessing the clinical applicability and interpretability of the combined model in predicting EC. Figure 2 illustrates the workflow of the entire study.

Univariate and multivariate regression analyses were performed to identify statistically significant associations between various variables and the presence of EC. Odds ratios (OR) and corresponding p-values were calculated to assess the effects of these variables. Specifically, age, BMI, menopausal status, irregular vaginal bleeding, hypertension, and diabetes were statistically significant in the univariate analysis. In the multivariate analysis, BMI, menopausal status, irregular vaginal bleeding, and diabetes remained significant. These variables were subsequently selected as optimal features for constructing the clinical model ( Table 2 ).

The predictive performance of the DL model was evaluated at the image-level using multimodal ultrasound. On the training set, the 2D ultrasound model (2D-DLS) achieved an area under the curve (AUC) of 0.842 (95% CI, 0.805–0.879), with an accuracy of 0.751, a sensitivity of 0.689, and a specificity of 0.898. In contrast, the color Doppler (Doppler-DLS) model yielded an AUC of 0.990 (95% CI, 0.983–0.997), an accuracy of 0.959, a sensitivity of 0.957, and a specificity of 0.964. On the external validation set, the 2D ultrasound model showed an AUC of 0.785 (95% CI, 0.718–0.853), an accuracy of 0.688, a sensitivity of 0.651, and a specificity of 0.839, while the color Doppler model obtained an AUC of 0.838 (95% CI, 0.786–0.889), an accuracy of 0.797, a sensitivity of 0.802, and a specificity of 0.781 ( Figures 3C, D ).

The performance of the patient-level models on the training set and external validation set was as follows: On the training set, the clinical model achieved an AUC of 0.820 (0.765, 0.874) with sensitivity of 0.580 and specificity of 0.937, the 2D ultrasound model had an AUC of 0.863 (0.825, 0.902) with sensitivity of 0.951 and specificity of 0.693, and the color doppler model showed an AUC of 0.988 (0.978, 0.997) with sensitivity of 0.963 and specificity of 0.963. The combined model performed the best with an AUC of 0.993 (0.986, 0.999), sensitivity of 0.988, and specificity of 0.959. On the external validation set, the clinical model had an AUC of 0.772 (0.690, 0.854) with sensitivity of 0.667 and specificity of 0.823, the 2D ultrasound model achieved an AUC of 0.792 (0.719, 0.864) with sensitivity of 0.824 and specificity of 0.694, and the color doppler model showed an AUC of 0.813 (0.745, 0.881) with sensitivity of 0.784 and specificity of 0.789. The merged model exhibited superior performance, achieving an AUC of 0.892 (95% CI: 0.846–0.938), a sensitivity of 0.784, and a specificity of 0.842. It outperformed the individual models across all evaluated metrics, demonstrating the effectiveness of integrating clinical data with multimodal ultrasound imaging ( Table 3 ; Figure 4 ). To assess the robustness of the model, we conducted subgroup analyses stratified by menopausal status, age groups (>50 years vs. ≤50 years), and BMI categories (>24 kg/m² vs. ≤24 kg/m²). Detailed results are presented in Supplementary Figure S2 .

Calibration curves were employed to assess the agreement between the predicted and actual outcomes, and the merged model demonstrated excellent calibration ability. The predicted probabilities closely aligned with the actual occurrence probabilities, approaching an ideal calibration, which indicates that the model’s predictions are both reliable and generalizable ( Figure 5A ). We explored the merged model’s predictive capability in distinguishing between non-EC and different cancer stages. Specifically, we conducted separate binary classification analyses to calculate the AUC for the merged model in distinguishing non-EC from each of the subsequent stages (1-4). The results indicate good performance, showing an AUC range of 0.87–0.97 ( Supplementary Figure S3 ).

Decision curve analysis (DCA) further demonstrated the clinical utility of the models. The DCA showed that the net benefit of the clinical model (red line) decreased as the high-risk threshold increased, although it remained relatively stable across most of the range. The 2D DL system (green line) provided a higher net benefit than the clinical model within the moderate-risk threshold range (0.2–0.4), despite a slight performance decline at higher thresholds. The Doppler DL system (yellow line) performed well in the low-risk threshold range; however, its net benefit diminished at higher thresholds, eventually falling behind the merged model. The merged model (orange line) outperformed the other models, particularly in the low-to-moderate risk thresholds, by providing a higher standardized net benefit and demonstrating more substantial decision-making advantages across various thresholds. The “All” (gray) and “None” (black) lines represent the decision baselines for scenarios in which either all patients are considered high-risk or none are. Overall, the merged model provided the highest net benefit across most thresholds, underscoring its strong clinical utility for risk prediction in this population ( Figure 5B ). Finally, a nomogram was constructed based on the risk scores derived from the three models to facilitate visual assessment by clinicians ( Figure 6 ).

Grad-CAM visualized the DL model’s decision-making process as heatmaps, where hot areas indicated the model’s attention regions ( Figures 7A-D ). SHAP explained the ranked contributions of features to the fused model ( Figure 7E ) and relationship among their own feature value impact on the model’s output ( Figure 7F ).