MRI-based deep learning model for differentiating endometrial cancer from atypical endometrial hyperplasia: A Multicenter Study

MRI-based deep learning model for differentiating endometrial cancer from atypical endometrial hyperplasia: A Multicenter Study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article MRI-based deep learning model for differentiating endometrial cancer from atypical endometrial hyperplasia: A Multicenter Study Dan Hong Chai, Feng Hua Ma, Yan Chen, Zi Jing Lin, Guo Fu Zhang, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9251059/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 6 You are reading this latest preprint version Abstract Background Differentiating endometrial carcinoma (EC) from atypical endometrial hyperplasia (AEH) before surgery is critical for guiding treatment decisions, yet current Magnetic resonance imaging (MRI) assessment is limited by overlapping appearances and inter-observer variability. Methods In this retrospective, two-center study, 297 patients (153 stage I EC and 144 AEH) who underwent multiparametric MRI were included. Radiomics features were extracted from T2-weight ed imaging with fat saturation (T2WI-FS), diffusion-weighted imaging (DWI), apparent diffusion coefficient (ADC) map, and contrast-enhanced T1-weighted imaging (CE-T1WI) sequences. Feature selections were performed using univariate filtering, correlation analysis, least absolute shrinkage and selection operator (LASSO) and multivariate stepwise regression. A Radiomics model, a Clin + Rad model, and a DenseNet121-2D model were constructed and compared for differentiating EC from AEH. Model performance was evaluated using receiver operating characteristic (ROC) curves, calibration curves, decision curve analysis (DCA). Results Seven radiomics features and eight deep-learning features were ultimately selected to construct the models. In external validation cohort, the DenseNet121-2D model achieved the highest diagnostic performance (area under the curve [AUC 0.854]), outperforming both the Radiomics model (AUC 0.803) and the Clin + Rad model (AUC 0.773). The DenseNet121-2D model also demonstrated better calibration and greater net clinical benefit across threshold probabilities. SHAP analysis confirmed that many discriminative features originated from ADC, consistent with its biological role in reflecting tumor cellularity and diffusion restriction. Conclusions The DenseNet121-2D model demonstrated superior generalizability and diagnostic accuracy compared with traditional radiomics approaches, providing a promising non-invasive tool to support individualized decision-making. Endometrial carcinoma Atypical endometrial hyperplasia Magnetic resonance imaging Radiomics Deep learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Endometrial cancer (EC) is the second most common malignancy of the female reproductive system worldwide [ 1 , 2 ] . Its incidence has increased by more than 132% since the 1990s, largely due to increasing obesity rates and population aging [ 3 ] . According to the International Federation of Gynecology and Obstetrics (FIGO), the 5-year survival rate exceeds 95% for stage I EC but drops to below 20% for stage IV disease [ 4 ] , underscoring the importance of early and accurate diagnosis. Atypical endometrial hyperplasia (AEH) is widely regarded as a precursor to EC [ 5 ] . Dilatation and curettage (D&C) is commonly used for diagnosis. However, the procedure is performed blindly, it is susceptible to sampling errors and may miss focal lesions [ 6 ] . Studies have shown that 37%–43% of patients diagnosed with AEH preoperatively are ultimately found to have EC after hysterectomy [ 7 , 8 ] . Nevertheless, up to now, reliable non-invasive tools for differentiating EC from AEH remain lacking [ 9 ] . MRI offers a non-invasive approach for evaluating endometrial abnormalities [ 10 ] . However, diagnostic accuracy is hampered by physiological variations and the marked overlap in imaging characteristics across endometrial lesions. [ 11 – 14 ] . For example, AEH frequently demonstrate signal intensity similar to normal endometrium on T2WI [ 15 ] . Moreover, MRI interpretation depends heavily on radiologist expertise, with reported accuracies ranging from 60% to 75% [ 14 ] . Therefore, a non-invasive adjunctive tool would be of significant clinical value for preoperative differentiation and treatment decision-making. Recent advances in radiomics and convolutional neural networks (CNNs) have created new possibilities for image-based diagnosis. Radiomics extracts quantitative imaging features to characterize disease phenotypes and assist in classification [ 16 ] . CNNs automatically learn hierarchical imaging representations to support feature extraction and classification [ 14 , 17 ] . Besides, CNNs capture subtle spatial features and have demonstrated strong performance across tasks [ 18 , 19 ] . Their effectiveness has been validated across multiple clinical applications, including breast cancer detection [ 20 ], cervical cancer segmentation [ 21 ] , and gastric disease classification [ 22 ] . Recently, several studies have explored non-invasive differentiation of endometrial diseases. Zhang et al. developed an MRI-based radiomics model to differentiate EC from AEH, reporting an accuracy of 93.2%; however, its single-center design and limited sample size restrict its generalizability [ 23 ] . Takahashi et al. prososed a hysteroscopy-based machine learning model that achieved accuracies of 85% for EC and 100% for AEH [ 24 ] . Despite these advances, no MRI-based DL model has been developed specifically to discriminate EC from AEH. Materials and method Data collection This retrospective study was approved by the Ethics Committee of Jinshan Hospital, Fudan University (No. JIEC 2024-S45), and was conducted in accordance with the Declaration of Helsinki. The requirement for informed consent was waived by the Ethics Committee owing to the retrospective nature of the study. A total of 648 patients from Center A (January 2021–December 2023), and 105 patients from the Center B (January 2020–December 2023), were initially screened. All patients had undergone diagnostic curettage or hysterectomy, and all diagnoses of endometrial-related diseases were pathologically confirmed. The inclusion criteria were as follows: (1) Histopathologically confirmed stage I EC or AEH; (2) Preoperative multiparametric pelvic MRI, including axial T2-weighted imaging with fat saturation (T2WI-FS), diffusion-weighted imaging (DWI), apparent diffusion coefficient (ADC) mapping, and contrast-enhanced T1-weighted imaging (CE-T1WI); and (3) No history of prior radiotherapy, chemotherapy, hormonal therapy, or comorbid gynecological tumors. Exclusion criteria were: (1) Lesions too small for reliable delineation; (2) Poor-quality images; (3) Patients not classified as stage I EC; (4) Incomplete imaging sequences or export failure; and (5) Presence of other concurrent malignancies. Ultimately, 297 patients met the eligibility criteria, including 251 patients from Center A (124 stage I EC and 127 AHE) and 46 patients from Center B (29 stage I EC and 17 AHE). The overall workflow is illustrated in Supplementary Figure S1 . MRI Acquisition MRI examinations were performed using a 1.5-T and 3.0-T scanner (Avanto; Siemens, Erlangens, Germany) equipped with a phased-array abdominal coil. The detailed MRI acquisition parameters are provided in Table 1 . All patients were scanned in the supine position under free-breathing conditions. The routine MRI protocol included axial T2WI-FS, DWI with corresponding ADC maps, and CE-T1WI. DWI was performed with b-values of 0 and 1000 s/mm 2 . CE-T1WI images were obtained during the arterial, venous, and delayed phases, acquired immediately, 90–120 seconds, and 150–180 seconds after intravenous administration of gadopentetate dimeglumine at 0.2 mmol/kg at an injection rate of 2–3 mL/s. Table 1 MRI Equipment, Scanning Protocols, and Parameters at Different Centers Center MRI field strength / model / brand MRI sequence b s/mm 2 Sequence type TR ms TE ms matrix slice thickness mm GAP mm A 1.5T Avanto Siemens T2WI-FS 0, 1000 SE 4000–6500 83 256*256 4.5 2 DWI EPI 2400–3300 83 176*176 5 1.5-2 ADC EP/RM 2900 83 176*176 5–6 1.5-2 CE-T1WI VIBE 4.9 2.4 256*256 5 0 B 1.5T MAGNETOM Aera Siemens T2WI-FS 0, 800 TSE 4200–5200 82 640*640 4.5-5 0.9-1 DWI EPI 5100 80 336*336 5 0.8 ADC EP 4200 65 336*336 4.5 0.8 CE-T1WI VIBE 4.5 2.2 640*640 3 0 (ADC: apparent diffusion coefficient mapping;CE-T1WI༚contrast-enhanced T1-weighted imaging༛DWI༚diffusion-weighted imaging༛ MRI༚magnetic resonance imaging༛T2WI-FS༚T2-weighted imaging with fat-saturation༛TE༚echo time༛TR༚repetition time༛) Clinical Data Analysis Preoperative clinicopathological data were collected, including age, body mass index (BMI), obstetric history, menopausal status, hypertension, diabetes, hyperlipidemia, and tumor markers such as CA125 and CA199. Potential clinical predictors were initially screened using univariate analysis ( p < 0.05) and were subsequently evaluated with multivariate stepwise logistic regression. Radiomics Feature Extraction, Selection, and Modeling The unified analytical pipeline is shown in Fig. 1 . All MRI data were imported into ITK-SNAP (version 4.2.0) for processing. DWI, ADC, and arterial-phase CE-T1WI images were rigidly registered to T2WI. Radiologist 1 manually delineated the regions of interest (ROIs) along the tumor boundaries on each T2WI slice, and applied the same segmentation method to the DWI, ADC, and CE-T1WI images while remaining blinded to all clinical information and pathological results. Radiologist 1 and radiologist 2 then independently performed a second round of segmentation on a randomly selected subset of 50 patients. During this process, they carefully excluded endometrial cavity fluid, hematoma, and adjacent normal myometrium, while allowing necrotic, hemorrhagic, or cystic components within the lesion to be included. Feature stability was calculated using both inter- and intraclass correlation coefficients (ICC), and only features with ICC value greater than 0.8 were retained. Segmentation of the volume of interest (VOI) encompassing the entire tumor was then performed. Before feature extraction, all original images were resampled to a voxel size of 1.5 × 1.5 × 1.5 mm³. A total of 6,752 radiomics features were extracted, and standardized using z-score normalization. Redundant or irrelevant features were removed using the following sequential steps: Mann–Whitney U test ( p < 0.01), correlation analysis (PyRadiomics, version 2.2.0), least absolute shrinkage and selection operator (LASSO) regression, and multivariate stepwise regression. The retained features were used to construct the Radiomics model. A combined clinical–radiomics (Clin + Rad) model was then generated by integrating independent clinical predictors. Spearman correlation analysis was further used to eliminate features with high collinearity (|r| ≥ 0.90). DL Feature Extraction, Selection, and Modeling A two-dimensional, slice-based strategy was adopted for DL feature extraction. Specifically, for each imaging sequence, the axial slice containing the largest cross-sectional tumor area was selected based on radiologist-annotated masks. The selected slice was then cropped using a bounding box tightly enclosing the lesion. Original images were resampled to a voxel size of 1.5 × 1.5 × 1.5 mm³ to standardize spatial resolution across subjects. Finally, all images were resized to 224 ×224 pixels and intensity-normalized for training, testing, and validating. A DenseNet121-2D CNN pre-trained on ImageNet was used for feature extraction. All DL features were standardized using z-score normalization derived from the training cohort. Feature selection followed the same sequential approach as in the Radiomics pipeline: univariate filtering, correlation analysis, LASSO regression, and multivariate stepwise regression. The retained key DL features were used to construct the final DL model. Statistical Analysis Statistical analyses were performed using R software (version 4.4.0). Categorical variables were compared using chi-square test or Fisher’s exact test. Continuous variables were first assessed for normality, and analyzed using either the independent samples t-test or the Mann–Whitney U test, as appropriate. Feature dimensionality reduction proceeded sequentially through inter- and intra-observer ICC assessment, univariate filtering, correlation analysis, LASSO regression, and multivariate stepwise regression. The extracted DenseNet121-2D features were incorporated into a logistic regression classifier to construct the DL-based model. Model performance was assessed using the area under the curve (AUC) of receiver operating characteristic (ROC) and calibration curves and decision curve analysis (DCA). Additional performance metrics included accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Confusion matrices (CM) were used to calculate true positive (TP), false positive (FP), false negative (FN), and true negative (TN). Differences in AUC values between models were assessed using DeLong’s test. A two-tailed P < 0.05 was considered statistically significant. Results Patient Characteristics and Feature Selection A total of 297 patients were included, comprising 153 patients with EC and 144 patients with AEH. The median age was 48 years (Q1-Q3: 34–58 years). Baseline characteristics of patient with EC or AEH in the training, internal testing, and external validation cohorts are summarized in Table 2 . Univariate and multivariate stepwise logistic regression analyses identified two age and menopausal status as independent clinical predictors. Serial feature screening yielded seven key radiomics features (two from T2WI and five from ADC images) and eight DL features, which were used to construct the corresponding models. LASSO regression results for radiomics feature selection and model calibration are presented in Fig. 2 . Table 2 Comparison of Clinical Characteristics Between EC and AEH Patients Across Cohorts Characteristics Overall (N = 297) (144 AEH, 153 EC) Training (N = 176) (89 AEH, 87 EC) p -value 1 Internal Testing (N = 75) (38 AEH, 37 EC) p -value 1 External Validation (N = 46) (17 AEH, 29 EC) p -value 1 Age, Median [IQR] 2 48.0 [34.0,58.0] 34.0 [30.0,40.0] 58.0 [51.0,65.0] < 0.001 33.5 [30.0,39.0] 56.0 [53.0,66.0] < 0.001 50.0 [45.0,58.0] 55.0 [48.0,59.0] 0.322 Obstetric history (Y/N) 3 99/194 58/28 6/80 < 0.001 22/16 6/31 < 0.001 4/13 3/26 0.184 Menopause (Y/N) 3 175/117 84/2 21/65 < 0.001 37/1 7/29 < 0.001 12/5 14/15 0.141 Menopause Age, Median [IQR] 2 50.0 [49.0,53.0] 50.0 [49.0,51.0] 50.0 [49.0,52.0] 0.837 50.0 [50.0,50.0] 51.0 [49.0,53.0] 0.683 0.948 BMI, Median [IQR] 2 24.8 [22.1,28.3] 26.1 [22.0,30.3] 24.3 [22.2,27.4] 0.130 25.4 [22.0,28.2] 25.3 [22.3,27.5] 0.423 NA [NA,NA] NA [NA,NA] Hypertension (Y/N) 3 218/77 78/10 56/30 < 0.001 35/3 23/14 0.002 11/6 15/14 0.391 Diabetes (Y/N) 3 254/41 81/7 66/20 0.005 35/3 30/7 0.191 0.619 Hyperlipidemia (Y/N) 3 281/14 85/3 81/5 0.494 37/1 33/4 0.200 16/1 29/0 0.370 CA125 (+/-) 4 239/31 74/7 75/9 0.653 31/4 26/7 0.274 12/1 21/3 1.000 CA199 (+/-) 4 137/28 23/1 53/14 0.105 11/1 18/8 0.223 10/2 22/2 0.588 HPV (Y/N) 3 126/13 36/5 47/3 0.460 13/1 17/3 0.627 3/0 10/1 1.000 1 Wilcoxon rank sum test; Pearson's Chi-squared test; Fisher's exact test; 2 Data are presented as median [Q1–Q3]; IQR denotes interquartile range (25th–75th percentiles); 3 (Y/N) represents yes or no of the corresponding clinical condition; 4 (+/−) represents positive or negative results, respectively. AEH: atypical endometrial hyperplasia; BMI: body mass index; EC: endometrial carcinoma; HPV: human papillomavirus; NA: not available (A-B) The coefficient paths demonstrate the effect of LASSO regularization on feature selection and model calibration. As λ (Log λ) increases, more coefficients shrink to zero, effectively removing redundant or weakly contributing features. (C-D) Cross-validation curves show that the binomial deviance initially decreases as λ increases—indicating improved model fit—then rises when λ becomes too large due to over-regularization. This pattern underscores the importance of selecting an optimal λ that balance overfitting and underfitting. (LASSO: least absolute shrinkage and selection operator) Model Performance The diagnostic performances of the models, along with the CMs and ROC curves for differentiating EC from AEH, are shown in Table 3 and Fig. 3 . ROC curve analysis demonstrated that although the DenseNet121-2D model showed slightly lower AUCs than the Radiomics and Clin + Rad models in the training cohort (0.882 vs. 0.913 vs. 0.973) and the internal testing cohort (0.821 vs. 0.919 vs. 0.979), it outperformed both models in the external validation cohort, achieving an AUC of 0.854, compared with 0.80 for the Radiomics model and 0.77 for the Clin + Rad model. In the external validation cohort, DenseNet121-2D model also demonstrated the highest sensitivity (55.2%) and highest specificity (94.1%), exceeding the performance of both the Radiomics model (sensitivity 31.0%, specificity 94.1%) and the Clin + Rad model (sensitivity 58.6%, specificity 88.2%). Table 3 Predictive Performance of Different Models in the Training, Internal Testing and External Validation Cohorts AUC Accuracy (%) Sensitivity (%) Specificity (%) PPV (%) NPV (%) Training (n = 176) Radiomics 0.913 (0.869–0.956) 85.2 (79.1–90.1) 80.5 (69.0-89.7) 89.9 (76.4–97.8) 88.6 (87.0-89.7) 82.5 (80.0-83.7) Clin + Rad 0.973 (0.953–0.993) 92.6 (87.7–96.0) 89.7 (81.6–96.6) 95.5 (75.3–100.0) 95.1 (94.7–95.5) 90.4 (88.2–90.8) DenseNet 121-2D 0.882 (0.831–0.932) 81.8 (75.3–87.2) 77.0 (58.6–86.2) 86.5 (71.9–93.3) 84.8 (81.0-86.2) 79.4 (76.2–80.6) Internal Testing (n = 75) Radiomics 0.919 (0.860–0.978) 81.3 (70.7–89.4) 78.4 (59.5–97.3) 84.2 (73.6–100.0) 82.9 (78.6–85.7) 80.0 (77.8–82.6) Clin + Rad 0.979 (0.950-1.000) 93.3 (85.1–97.8) 94.6 (62.2–100.0) 92.1 (84.2–100.0) 92.1 (88.5–92.5) 94.6 (94.1–95.0) DenseNet 121-2D 0.821 (0.726–0.915) 72.0 (60.4–81.8) 62.2 (40.5–86.5) 81.6 (60.5–97.4) 76.7 (68.2–82.1) 68.9 (62.1–72.5) External Validation (n = 46) Radiomics 0.803 (0.673–0.933) 54.3 (39.0-69.1) 31.0 (20.7–69.1) 94.1 (82.4–100.0) 90.0 (85.7–95.2) 44.4 (41.2–45.9) Clin + Rad 0.773 (0.626–0.919) 69.6 (54.2–82.3) 58.6 (10.3–75.9) 88.2 (46.9–100.0) 89.5 (60.0-91.7) 55.6 (39.9–58.6) DenseNet 121-2D 0.854 (0.743–0.965) 69.6 (54.2–82.3) 55.2 (34.5–86.3) 94.1 (82.4–100.0) 94.1 (90.9–96.2) 55.2 (51.9–56.7) AUC: area under the receiver operating characteristic curve; PPV: positive predictive value; NPV: negative predictive value; Clin + Rad: clinical-radiomics model. (A): CM of the Radiomics, Clin + Rad, and DenseNet 121 models in the training, internal testing, and external validation cohorts. The color intensity corresponds to the number of cases in each cell: higher values are represented by darker color. (B): ROC curves of three models for distinguishing EC from AEH in the training (a), internal testing (b), external validation cohorts (c). (CM: Confusion matrix; ROC: receiver operating characteristic; Clin + Rad: Clinical+Radiomics.) According to the DeLong test, the Clin + Rad model showed significantly better diagnostic performance than the Radiomics model (P = 0.001) and the DenseNet121-2D model (P < 0.001) in the training cohort, as shown in Table 4 . However, in the external validation cohort, the DeLong test showed no statistically significant differences in AUC among the models. Nevertheless, additional analyses based on net reclassification improvement (NRI) and integrated discrimination improvement (IDI) revealed that the IDI of DenseNet121-2D relative to the Radiomics model remained statistically significant (P = 0.022), whereas the corresponding NRI was not significant. Furthermore, no significant differences in either NRI or IDI were found between DenseNet121-2D and the Clin + Rad model. These results indicate that the incremental diagnostic benefit of DenseNet121-2D in the external validation cohort was limited, although its potential value should be further investigated in a larger sample. Detailed results are provided in Table 5 . Table 4 DeLong Test Results for Pairwise Comparisons of AUCs among Different Models in Different Cohorts Comparative performance of different models P value Training Internal Testing External Validation Radiomics vs Clin + Rad 0.001 0.038 0.668 Radiomics vs DenseNet121_2D 0.298 0.025 0.447 DenseNet121_2D vs Clin + Rad < 0.001 0.002 0.315 Table 5 Comparison of NRI and IDI for Incremental Diagnostic Performance among Different Models across Different Cohorts Different model Cohort NRI (95% CI) P value IDI (95% CI) P value DenseNet121_2D vs Radiomics Training -0.079 (-0.254, 0.097) 0.380 -0.092 (-0.172, -0.011) 0.025 Internal Testing -0.349 (-0.627, -0.070) 0.014 -0.186 (-0.324, -0.049) 0.008 External Validation 0.227 (-0.023, 0.477) 0.075 0.135 (0.019, 0.250) 0.022 DenseNet121_2D vs Clin + Rad Training -0.374 (-0.538, -0.210) < 0.001 -0.323 (-0.406, -0.239) < 0.001 Internal Testing -0.721 (-0.983, -0.459) < 0.001 -0.453 (-0.612, -0.294) < 0.001 External Validation 0.059 (-0.262, 0.380) 0.719 0.007 (-0.196, 0.211) 0.943 Radiomics vs Clin + Rad Training 0.351 (0.216, 0.488) < 0.001 0.231 (0.170, 0.291) < 0.001 Internal Testing 0.426 (0.201, 0.651) < 0.001 0.267 (0.147, 0.386) < 0.001 External Validation 0.099 (-0.179, 0.378) 0.484 0.128 (-0.047, 0.302) 0.151 Note: NRI and IDI were used to assess the incremental diagnostic value of the comparison model relative to the reference model. NRI and IDI additionally reflect changes in case reclassification and overall discriminative ability, and may therefore complement AUC-based comparisons. NRI/IDI > 0 indicates that the former model outperforms the latter in classification or discrimination ability, whereas NRI/IDI < 0 indicates poorer performance of the former model relative to the latter. (95% CI: 95% confidence Interval; NRI: Net Reclassification Improvement; IDI: Integrated Discrimination Improvement;) Calibration and Clinical Utility As shown in Fig. 4A-4C, the calibration curves for both the Radiomics and DenseNet121-2D models closely aligned with the ideal reference line in the training and internal testing cohorts, indicating good agreement between predicted and observed outcomes. However, in the external validation cohort, the DenseNet121-2D model demonstrated superior calibration, reflecting greater reliability and generalizability. This advantage was further supported by DCA (Fig. 4D-F), which demonstrated that the DenseNet121-2D model yielded the highest net clinical benefit for differentiating EC from AEH across a broad range of threshold probabilities. Figure 4. Model performance assessed by calibration curve and DCA (A-C) Calibration curves of the Radiomics, Clin + Rad, and DenseNet121-2D models in the training (A), internal testing (B), and external validation (C) cohorts. The DenseNet121-2D model demonstrats better calibration and generalizability in the external validation cohort compared with the Clin + Rad and Radiomics models. (D-F) DCA results for. the training (D), internal testing (E), and external validation (F) cohorts. The DenseNet121-2D model yields a consistently higher net benefit than the other models at threshold probabilities below 0.6. (DCA: decision curve analysis) Correlation Analysis and Model Robustness To further explore feature relationships and enhance model interpretability, we examined the linear associations between key features and the target variable using Pearson correlation and constructing a correlation matrix (Fig. 5 ). The DenseNet121-2D model demonstrated low inter-feature correlations (|r| 0.7), indicating potential multicollinearity, which may contribute to reduced generalizability and overfitting. This aligns with model performance: despite high AUCs in the training and internal testing cohorts (> 0.90), both the Radiomics and Clin + Rad models showed a markedly declined in the external validation cohort (AUC = 0.80 and 0.773), whereas the DenseNet121-2D model maintained more stable performance. The figure presents pairwise Pearson correlation coefficients for the Radiomics (A), Clin + Rad (B), and DenseNet121-2D (C) feature sets. Color intensity represents both the strength and direction of the correlations, with red indicating positive correlations and blue indicating negative correlations. A diagonal value of 1.00 reflects the perfect self-correlation of each feature. Panels correspond to the training (A, D), internal testing (B, E), and external validation (C, F) cohorts. (A-C) Bee-swarm plots illustrating the distribution and influence of each feature on model predictions based on SHAP values. (D-F) Bar plots showing the mean SHAP values, reflecting the overall importance of each feature. The y-axis lists the model features, and the x-axis displays the mean SHAP values, indicating the average contribution of each feature to the model output. (SHAP: shapley additive explanations) SHAP Interpretability Analysis To further evaluate model interpretability, SHAP analysis was applied to the DL model (Fig. 6 ). Among all extracted features, the most influential were DWI-F253, ADC-F961, ADC-F502, T2WI-FS-F87, and ADC-F16 in the training cohort. Notably, these same five features remained the top contributors in the external validation cohort, although with varying magnitude. In the internal testing cohort, DWI-F253, T2WI-FS-F87, ADC-F961, and ADC-F502 continued to rank among the most important features. This recurrent identification of key features across cohorts supports the robustness and potential generalizability of the DenseNet121-2D model. Discussion In this study, we constructed and validated a DL-based diagnostic model (DenseNet121-2D) that can automatically differentiate endometrial carcinoma (EC) from atypical endometrial hyperplasia (AEH) using datasets from two centers. The model demonstrated satisfactory performance in the training cohort (AUC, 0.882)) and internal testing cohort (AUC, 0.821). Importantly, in the external validation cohort, the DenseNet121-2D model outperformed both the Radiomics and Clin + Rad models (AUC, 0.854 vs. 0.803 vs. 0.773), confirming its generalizability and robust performance across datasets. This method may help reduce misdiagnosis and improve diagnostic efficiency, and potentially decrease unnecessary surgeries for patients with AEH. Given its underlying architecture, the proposed model could be further extended to support intelligent, image-based auxiliary diagnoses for other diseases. In the present study, we built both a Radiomics model and a DenseNet121-2D model to differentiate patients with EC and AEH. The reliability of image-derived features is critical for building robust radiomics and DL models [ 25 ] . However, processes such as target segmentation, feature extraction, feature selection, and model construction are inherently subject to inter- and intra-observer variability [ 26 ] . To address this limitation, manual segmentation was adopted due to the relatively limited sample size, and the intraclass correlation coefficient (ICC) was calculated to enhance feature robustness and reproducibility [ 27 ] . Furthermore, feature selection was initially conducted using logistic regression and subsequently refined through multivariate stepwise regression. A Pearson correlation heatmap was used to ensure selected features did not exhibit high collinearity. In addition, Shapley Additive Explanations (SHAP) analysis was used to improve the interpretability and transparency of the final models. Radiomics approaches have previously been applied to endometrial disease classification, generally showing strong performance for EC assessment. For example, Liu et al. constructed an MRI-based radiomics model and a visual nomogram, which effectively differentiated patients with benign and malignant endometrial lesions [ 26 ] . However, the diagnostic accuracy of radiomics has been reported to be limited to 78.0% in test cohorts [ 26 ] , and often suffer from limited generalizability and a lack of robustness across different datasets. Yan et al. developed a radiomics nomogram for fertility-preserving treatment candidate selection in EC and AEH, achieving an AUC of 0.94 in a prospective validation cohort—likely attributable to the reduced biases in prospective study [ 28 ] . The DenseNet121-2D model was trained using an axial slice with maximum tumor area from T2WI-FS, DWI, ADC, and CE-T1WI. Among these sequences, CE-T1WI and DWI provide essential functional information in the evaluation of EC and are frequently used as complementary tools to T2WI [ 27 ] . Tumor enhancement patterns on CE-T1WI primarily reflect vascularity: although most EC are hypovascular, some lesions may appear iso- or hypervascular relative to the surrounding myometrium [ 29 ] . In addition, DWI and ADC mainly reflect water diffusion characteristics, ADC values are negatively associated with tumor cell density [ 30 ] , and prior studies have shown that ADC values in EC are significantly lower than those observed in AEH [ 31 , 32 ] . Studies have demonstrated that CNN-based methods achieved their highest diagnostic performance when trained on single axial ADC maps [ 27 ] . Therefore, incorporating CE-T1WI, DWI, and ADC maps alongside T2WI facilitates a more accurate, comprehensive, and reliable diagnostic assessment of EC. Given the strong diagnostic contribution of DWI and ADC-derived information discussed above, it is not surprising that a substantial proportion of discriminative features originated from the ADC maps in our study. This finding is clinically plausible, as EC typically demonstrates restricted diffusion which appears as hyperintensity relative to the surrounding myometrium on DWI and corresponding hypointensity on ADC maps [ 32 ] . Previous studies have consistently underscored the value of ADC measurements in differentiating benign from malignant endometrial lesions [ 33 , 34 ] . For example, Tamai et al. reported that normal endometrium exhibited significantly higher ADC values compared with EC, supporting ADC as an initial imaging biomarker [ 32 ] . The robustness of ADC-derived radiomics features has also been validated across centers. In a multicenter study, Peerlings et al. reported that 122 ADC-based features demonstrated high reproducibility (ICC > 0.85) across different tissues, scanners, and vendors [ 35 ] . This observation is consistent with our findings and explains the preferentially retention of ADC features during feature selection. Biologically, ADC values reflect the extent of water diffusion restriction, which becomes more pronounced in tumors with high cellularity and reduced extracellular space. A meta-analysis by Surov et al. confirmed a significant negative correlation between mean ADC values and tumor cellularity [ 36 ] . Taken together, these biological and technical factors underscore the reliability and clinical relevance of ADC features in differentiating EC from AEH. The DenseNet121-2D model achieved the best predictive performance in the external validation cohort (AUC, 85.4%), representing a notable improvement over the Radiomics model (AUC 80.3%) and demonstrating its superior generalizability. As outlined earlier, numerous DL architectures have been successfully applied to medical imaging tasks. For example, Takahashi et al. developed a DL algorithm trained on more than 411,800 hysteroscopic images, achieving strong discriminatory capacity across five endometrial disease categories [ 24 ] . Other work has applied DL models to pathological image, with high PPV (82.94%) and NPV (96.53%) both demonstrating strong accuracy [ 37 ] . In contrast, the NPV of all models in our study was substantially lower than the PPV. This discrepancy may be attributable to sample imbalance, as positive cases were more prevalent—particularly in the external validation cohort, where EC cases was twice that of AEH cases. To date, no MRI-based DL model has been reported for differentiating EC from AEH, highlighting the novelty and clinical significance of our work. There are also several limitations in our study. First, the generalizability of our findings is limited by the relatively small sample size, especially in the external validation cohort. Second, prospective validation was not performed, which may restrict the strength of our conclusions. Further study is warrant in a prospective, multi-center studies with larger and more heterogeneous dataset. Third, manual lesion segmentation is both time-consuming and susceptible to inter-observer variability. Future studies should explore automatic or semi-automatic segmentation approaches that provide more efficient and reproducible delineation. Conclusions In this study, we developed the first MRI-based CNN model specifically designed to discriminate EC from AEH. The proposed DenseNet121-2D architecture demonstrated superior diagnostic performance compared with both Radiomics model and Clin + Rad model, particularly in the external validation cohort. Furthermore, the model offers objective, quantitative evidence to support radiologists' decision-making, highlighting its potential for further clinical application. Declarations All authors have approved the final version of the manuscript and agree with its submission to BMC MEDICAL IMAGING . Funding This study was supported by the National Natural Science Foundation of China (Grant No. 82471943) and the Shanghai Jinshan District Health Commission (JSZK2023A02). Ethics approval and consent to participate This retrospective study was approved by the Ethics Committee of Jinshan Hospital, Fudan University (No. JIEC 2024-S45), and was conducted in accordance with the Declaration of Helsinki. The requirement for informed consent was waived by the Ethics Committee owing to the retrospective nature of the study. Competing interests The authors declare that they have no competing interests. Consent for publication Not applicable. Availability of data and materials The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request. Authors’ contributions (I) Conception and design: Dan Hong Chai; Zi Jing Lin (II) Administrative support: Jin Wei Qiang (III) Provision of study materials or patients: Feng Hua Ma; Guo Fu Zhang; Zhao Xia Qian (IV) Collection and assembly of data: Dan Hong, Chai; Yan Chen; (V) Data analysis and interpretation: Jin Wei Qiang (VI) Manuscript writing: All authors (VII) Final approval of manuscript: All authors Acknowledgements Not applicable. References Ferlay J, Soerjomataram I, Dikshit R. Cancer incidence and mortality worldwide: sources, methods and major patterns in GLOBOCAN 2012. Int J Cancer. 2015;136(5):E359–86. https://doi.org/10.1002/ijc.29210 . Bosse T, Peters EE, Creutzberg CL. Substantial lymph-vascular space invasion (LVSI) is a significant risk factor for recurrence in endometrial cancer–A pooled analysis of PORTEC 1 and 2 trials. Eur J Cancer. 2015;51(13):1742–50. https://doi.org/10.1016/j.ejca.2015.05.015 . Crosbie EJ, Kitson SJ, McAlpine JN. Endometrial cancer. Lancet. 2022;399(10333):1412–28. https://doi.org/10.1016/S0140-6736(22)00323-3 . Bhardwaj V, Sharma A, Parambath SV. Machine Learning for Endometrial Cancer Prediction and Prognostication. Front Oncol. 2022. https://doi.org/10.3389/fonc.2022.852746 . 27;12:852746. Herrington CS, editor. WHO classification of tumours: female genital tumours. 5th ed. Lyon: International Agency for Research on Cancer; 2020. Svirsky R, Smorgick N, Rozowski U. Can we rely on blind endometrial biopsy for detection of focal intrauterine pathology? Am J Obstet Gynecol. 2008;199(2):115. https://doi.org/10.1016/j.ajog.2008.02.015 . .e1-3. Trimble CL, Kauderer J, Zaino R. Concurrent endometrial carcinoma in women with a biopsy diagnosis of atypical endometrial hyperplasia: a Gynecologic Oncology Group study. Cancer. 2006;106(4):812–9. https://doi.org/10.1002/cncr.21650 . Rakha E, Wong SC, Soomro I. Clinical outcome of atypical endometrial hyperplasia diagnosed on an endometrial biopsy: institutional experience and review of literature. Am J Surg Pathol. 2012;36(11):1683–90. https://doi.org/10.1097/PAS.0b013e31825dd4ff . Gentry-Maharaj A, Karpinskyj C. Current and future approaches to screening for endometrial cancer. Best Pract Res Clin Obstet Gynaecol. 2020;65:79–97. https://doi.org/10.1016/j.bpobgyn.2019.12.006 . Haldorsen IS, Salvesen HB. What Is the Best Preoperative Imaging for Endometrial Cancer? Curr Oncol Rep. 2016;18(4):25. https://doi.org/10.1007/s11912-016-0506-0 . Nalaboff KM, Pellerito JS, Ben-Levi E. Imaging the endometrium: disease and normal variants. Radiographics. 2001;21(6):1409–24. https://doi.org/10.1148/radiographics.21.6.g01nv211409 . Kierans AS, Bennett GL, Haghighi M, Rosenkrantz AB. Utility of conventional and diffusion-weighted MRI features in distinguishing benign from malignant endometrial lesions. Eur J Radiol. 2014;83(4):726–32. https://doi.org/10.1016/j.ejrad.2013.11.030 . Bakir B, Sanli S, Bakir VL. Role of diffusion weighted MRI in the differential diagnosis of endometrial cancer, polyp, hyperplasia, and physiological thickening. Clin Imaging. 2017;41:86–94. https://doi.org/10.1016/j.clinimag.2016.10.016 . Zhang Y, Wang Z, Zhang J. Deep learning model for classifying endometrial lesions. J Transl Med. 2021;19(1):10. https://doi.org/10.1186/s12967-020-02660-x . Takeuchi M, Matsuzaki K, Uehara H, Yoshida S, Nishitani H, Shimazu H. Pathologies of the uterine endometrial cavity: usual and unusual manifestations and pitfalls on magnetic resonance imaging. Eur Radiol. 2005;15(11):2244–55. https://doi.org/10.1007/s00330-005-2814-x . De Marchi F, Fabris AM, Tommasi L, Nappi L, Saccardi C, Litta P. Accuracy of hysteroscopy made by young residents in detecting endometrial pathologies in postmenopausal women. Eur J Gynaecol Oncol. 2014;35(3):219–23. Wiestler B, Menze B. Deep learning for medical image analysis: a brief introduction. Neurooncol Adv. 2021;2(Suppl 4):iv35–41. https://doi.org/10.1093/noajnl/vdaa092 . Litjens G, Kooi T, Bejnordi BE. A survey on deep learning in medical image analysis. Med Image Anal. 2017;42:60–88. https://doi.org/10.1016/j.media.2017.07.005 . Mazurowski MA, Buda M, Saha A, Bashir MR. Deep learning in radiology: An overview of the concepts and a survey of the state of the art with focus on MRI. J Magn Reson Imaging. 2019;49(4):939–54. https://doi.org/10.1002/jmri.26534 . Chougrad H, Zouaki H, Alheyane O. Deep Convolutional Neural Networks for breast cancer screening. Comput Methods Programs Biomed. 2018;157:19–30. https://doi.org/10.1016/j.cmpb.2018.01.011 . Rodríguez Outeiral R, González PJ, Schaake EE, van der Heide UA, Simões R. Deep learning for segmentation of the cervical cancer gross tumor volume on magnetic resonance imaging for brachytherapy. Radiat Oncol. 2013;18(1):91. https://doi.org/10.1186/s13014-023-02283-8 . Cho BJ, Bang CS, Park SW. Automated classification of gastric neoplasms in endoscopic images using a convolutional neural network. Endoscopy. 2019;51(12):1121–9. https://doi.org/10.1055/a-0981-6133 . Zhang J, Zhang Q, Wang T. Multimodal MRI-Based Radiomics-Clinical Model for Preoperatively Differentiating Concurrent Endometrial Carcinoma From Atypical Endometrial Hyperplasia. Front Oncol. 2022;12:887546. https://doi.org/10.3389/fonc.2022.887546 . Takahashi Y, Sone K, Noda K. Automated system for diagnosing endometrial cancer by adopting deep-learning technology in hysteroscopy. PLoS ONE. 2021;16(3):e0248526. https://doi.org/10.1371/journal.pone.0248526 . van Timmeren JE, Cester D, Tanadini-Lang S, Alkadhi H, Baessler B. Radiomics in medical imaging-how-to guide and critical reflection. Insights Imaging. 2020;11(1):91. https://doi.org/10.1186/s13244-020-00887-2 . Liu J, Li S, Lin H. Development of MRI-based radiomics predictive model for classifying endometrial lesions. Sci Rep. 2023;13(1):1590. https://doi.org/10.1038/s41598-023-28819-2 . Urushibara A, Saida T, Mori K. The efficacy of deep learning models in the diagnosis of endometrial cancer using MRI: a comparison with radiologists. BMC Med Imaging. 2022;22(1):80. https://doi.org/10.1186/s12880-022-00808-3 . Yan BC, Ma FH, Li Y. An MRI radiomics nomogram improves the accuracy in identifying eligible candidates for fertility-preserving treatment in endometrioid adenocarcinoma. Am J Cancer Res. 2022;12(3):1056–68. Whittaker CS, Coady A, Culver L, Rustin G, Padwick M, Padhani AR. Diffusion-weighted MR imaging of female pelvic tumors: a pictorial review. Radiographics. 2009;29(3):759–74. https://doi.org/10.1148/rg.293085130 . discussion 774-8. Funt SA, Hricak H. Ovarian malignancies Top Magn Reson Imaging. 2003;14(4):329–37. https://doi.org/10.1097/00002142-200308000-00005 Fujii S, Matsusue E, Kigawa J. Diagnostic accuracy of the apparent diffusion coefficient in differentiating benign from malignant uterine endometrial cavity lesions: initial results. Eur Radiol. 2008;18(2):384–9. https://doi.org/10.1007/s00330-007-0769-9 . Tamai K, Koyama T, Saga T. Diffusion-weighted MR imaging of uterine endometrial cancer. J Magn Reson Imaging. 2007;26(3):682–7. https://doi.org/10.1002/jmri.20997 . Rechichi G, Galimberti S, Signorelli M. Endometrial cancer: correlation of apparent diffusion coefficient with tumor grade, depth of myometrial invasion, and presence of lymph node metastases. AJR Am J Roentgenol. 2011;197(1):256–62. https://doi.org/10.2214/AJR.10.5584 . Petrila O, Nistor I, Romedea NS, Negru D, Scripcariu V. Can the ADC Value Be Used as an Imaging Biopsy in Endometrial Cancer? Diagnostics (Basel). 2024;14(3):325. https://doi.org/10.3390/diagnostics14030325 . Peerlings J, Woodruff HC, Winfield JM. Stability of radiomics features in apparent diffusion coefficient maps from a multi-centre test-retest trial. Sci Rep. 2019;9(1):4800. https://doi.org/10.1038/s41598-019-41344-5 . Surov A, Meyer HJ, Wienke A. Correlation between apparent diffusion coefficient (ADC) and cellularity is different in several tumors: a meta-analysis. Oncotarget. 2017;8(35):59492–9. https://doi.org/10.18632/oncotarget.17752 . Sun H, Zeng X, Xu T, Peng G, Ma Y. Computer-Aided Diagnosis in Histopathological Images of the Endometrium Using a Convolutional Neural Network and Attention Mechanisms. IEEE J Biomed Health Inf. 2020;24(6):1664–76. https://doi.org/10.1109/JBHI.2019.2944977 . Additional Declarations No competing interests reported. Supplementary Files supplementary.docx Cite Share Download PDF Status: Under Review Version 1 posted Reviewers agreed at journal 15 May, 2026 Reviewers invited by journal 20 Apr, 2026 Editor invited by journal 09 Apr, 2026 Editor assigned by journal 02 Apr, 2026 Submission checks completed at journal 02 Apr, 2026 First submitted to journal 28 Mar, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9251059","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":626404219,"identity":"ddaa2600-9a99-41b0-8df7-cd89264479f5","order_by":0,"name":"Dan Hong Chai","email":"","orcid":"","institution":"Jinshan Hospital of Fudan University","correspondingAuthor":false,"prefix":"","firstName":"Dan","middleName":"Hong","lastName":"Chai","suffix":""},{"id":626404220,"identity":"42dc1013-8ac7-4bb5-a9c2-3c9e70159207","order_by":1,"name":"Feng Hua Ma","email":"","orcid":"","institution":"Obstetrics and Gynecology Hospital of Fudan University","correspondingAuthor":false,"prefix":"","firstName":"Feng","middleName":"Hua","lastName":"Ma","suffix":""},{"id":626404221,"identity":"55af6a95-7315-4aef-bab3-0ac0ba8df580","order_by":2,"name":"Yan Chen","email":"","orcid":"","institution":"Jinshan Hospital of Fudan University","correspondingAuthor":false,"prefix":"","firstName":"Yan","middleName":"","lastName":"Chen","suffix":""},{"id":626404223,"identity":"4c3c380c-6429-45c4-8853-c4f81249e645","order_by":3,"name":"Zi Jing Lin","email":"","orcid":"","institution":"Jinshan Hospital of Fudan University","correspondingAuthor":false,"prefix":"","firstName":"Zi","middleName":"Jing","lastName":"Lin","suffix":""},{"id":626404225,"identity":"5b3d06df-8bcc-42f5-b392-e98c780666ac","order_by":4,"name":"Guo Fu Zhang","email":"","orcid":"","institution":"Obstetrics and Gynecology Hospital of Fudan University","correspondingAuthor":false,"prefix":"","firstName":"Guo","middleName":"Fu","lastName":"Zhang","suffix":""},{"id":626404228,"identity":"a83f91a4-b435-4468-aff0-41b7665f9496","order_by":5,"name":"Zhao Xia Qian","email":"","orcid":"","institution":"Shanghai International Peace Maternity and Child Health Hospital of Shanghai Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Zhao","middleName":"Xia","lastName":"Qian","suffix":""},{"id":626404231,"identity":"fc240e5d-8c1a-4234-bacc-577b1eae76e3","order_by":6,"name":"Jin Wei Qiang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA9UlEQVRIiWNgGAWjYDCCAwwMBglAmo2BsQFI2fDw8zeQpiVNRnLGAcJakMFhG4OGBPw6+G4fYCh4uOOwHB97cuPngl/neQwYDjB++JiDW4vkuQQGg8Qzh43ZeB42S8/su81jztzALDlzG24tBmeAOLHtdmKbRGKDNG/PbR7LhgNszLxEaKkHamn+zdtzjsfgQAJxWhLYJBLbpHl+HCCsRfIMYwNQy3/DNp6Hbda8Dck8kjMONuP1C98Z5mOGP9vS5OXb0x/f5vljZ8/P33zww0c8WhgYGNsMIIwEEBss0oBPPQgwP4BrYfhDSPEoGAWjYBSMRAAA549T8RUPSI8AAAAASUVORK5CYII=","orcid":"","institution":"Jinshan Hospital of Fudan University","correspondingAuthor":true,"prefix":"","firstName":"Jin","middleName":"Wei","lastName":"Qiang","suffix":""}],"badges":[],"createdAt":"2026-03-28 08:38:16","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9251059/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9251059/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108181314,"identity":"05088f75-24ae-4d15-b4b5-8a325b2f3b8e","added_by":"auto","created_at":"2026-04-30 08:58:32","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":555930,"visible":true,"origin":"","legend":"\u003cp\u003eThe workflow of radiomics model diagram\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-9251059/v1/d94680f738967018e2f8a009.png"},{"id":108182418,"identity":"7d42d16a-e074-4291-960f-cb2c3e027a7c","added_by":"auto","created_at":"2026-04-30 08:59:21","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":462386,"visible":true,"origin":"","legend":"\u003cp\u003eLASSO regression for feature selection and model evaluation\u003c/p\u003e\n\u003cp\u003e(A-B) The coefficient paths demonstrate the effect of LASSO regularization on feature selection and model calibration. As λ (Log λ) increases, more coefficients shrink to zero, effectively removing redundant or weakly contributing features. (C-D) Cross-validation curves show that the binomial deviance initially decreases as λ increases—indicating improved model fit—then rises when λ becomes too large due to over-regularization. This pattern underscores the importance of selecting an optimal λ that balance overfitting and underfitting. (LASSO: least absolute shrinkage and selection operator)\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-9251059/v1/6f6cd8ed0493492889d03995.png"},{"id":108072568,"identity":"aca3a700-c691-43f2-acde-3f7f302fce5a","added_by":"auto","created_at":"2026-04-29 06:14:21","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":279224,"visible":true,"origin":"","legend":"\u003cp\u003eCM and ROC curves for distinguishing EC from AEH across different models.\u003c/p\u003e\n\u003cp\u003e(A): CM of the Radiomics, Clin+Rad, and DenseNet 121 models in the training, internal testing, and external validation cohorts. The color intensity corresponds to the number of cases in each cell: higher values are represented by darker color. (B): ROC curves of three models for distinguishing EC from AEH in the training (a), internal testing (b), external validation cohorts (c). (CM: Confusion matrix; ROC: receiver operating characteristic; Clin+Rad: Clinical+Radiomics.)\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-9251059/v1/9affe8984f1e4f615b7ec5d3.png"},{"id":108182055,"identity":"5aac9f68-d4f1-4c22-b809-ba11ebf585ca","added_by":"auto","created_at":"2026-04-30 08:59:06","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":511839,"visible":true,"origin":"","legend":"\u003cp\u003eModel performance assessed by calibration curve and DCA\u003c/p\u003e\n\u003cp\u003e(A-C) Calibration curves of the Radiomics, Clin+Rad, and DenseNet121-2D models in the training (A), internal testing (B), and external validation (C) cohorts. The DenseNet121-2D model demonstrats better calibration and generalizability in the external validation cohort compared with the Clin+Rad and Radiomics models. (D-F) DCA results for. the training (D), internal testing (E), and external validation (F) cohorts. The DenseNet121-2D model yields a consistently higher net benefit than the other models at threshold probabilities below 0.6. (DCA: decision curve analysis)\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-9251059/v1/cc3a72d411f491cb3d3367d3.png"},{"id":108490982,"identity":"5d2f28ec-4cd3-4199-8544-560a53ecd6de","added_by":"auto","created_at":"2026-05-05 09:50:44","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":575043,"visible":true,"origin":"","legend":"\u003cp\u003eHeatmap of the correlation matrices across multiple feature sets\u003c/p\u003e\n\u003cp\u003eThe figure presents pairwise Pearson correlation coefficients for the Radiomics (A), Clin+Rad (B), and DenseNet121-2D (C) feature sets. Color intensity represents both the strength and direction of the correlations, with red indicating positive correlations and blue indicating negative correlations. A diagonal value of 1.00 reflects the perfect self-correlation of each feature.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-9251059/v1/025df3cd0b94b9af63cbac21.png"},{"id":108072572,"identity":"f5c90613-7d19-40f8-b74c-b0592c3bb51d","added_by":"auto","created_at":"2026-04-29 06:14:21","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":942850,"visible":true,"origin":"","legend":"\u003cp\u003eSHAP interpretation of the Densenet121-2D model\u003c/p\u003e\n\u003cp\u003ePanels correspond to the training (A, D), internal testing (B, E), and external validation (C, F) cohorts. (A-C) Bee-swarm plots illustrating the distribution and influence of each feature on model predictions based on SHAP values. (D-F) Bar plots showing the mean SHAP values, reflecting the overall importance of each feature. The y-axis lists the model features, and the x-axis displays the mean SHAP values, indicating the average contribution of each feature to the model output. (SHAP: shapley additive explanations)\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-9251059/v1/bd53f2f8c153d67cfb367956.png"},{"id":108804284,"identity":"3accd0e9-fb6d-452f-9756-d75da65e46df","added_by":"auto","created_at":"2026-05-08 15:18:54","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3844042,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9251059/v1/e34b7407-b015-4383-9394-c458ca3af890.pdf"},{"id":108072566,"identity":"47be50aa-be51-410a-8157-ff8e4dd4bb0c","added_by":"auto","created_at":"2026-04-29 06:14:21","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":227274,"visible":true,"origin":"","legend":"","description":"","filename":"supplementary.docx","url":"https://assets-eu.researchsquare.com/files/rs-9251059/v1/8f01b60e833cf6fc33789e76.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"MRI-based deep learning model for differentiating endometrial cancer from atypical endometrial hyperplasia: A Multicenter Study","fulltext":[{"header":"Introduction","content":"\u003cp\u003eEndometrial cancer (EC) is the second most common malignancy of the female reproductive system worldwide \u003csup\u003e[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]\u003c/sup\u003e. Its incidence has increased by more than 132% since the 1990s, largely due to increasing obesity rates and population aging \u003csup\u003e[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]\u003c/sup\u003e. According to the International Federation of Gynecology and Obstetrics (FIGO), the 5-year survival rate exceeds 95% for stage I EC but drops to below 20% for stage IV disease \u003csup\u003e[\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]\u003c/sup\u003e, underscoring the importance of early and accurate diagnosis. Atypical endometrial hyperplasia (AEH) is widely regarded as a precursor to EC \u003csup\u003e[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]\u003c/sup\u003e. Dilatation and curettage (D\u0026amp;C) is commonly used for diagnosis. However, the procedure is performed blindly, it is susceptible to sampling errors and may miss focal lesions \u003csup\u003e[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]\u003c/sup\u003e. Studies have shown that 37%\u0026ndash;43% of patients diagnosed with AEH preoperatively are ultimately found to have EC after hysterectomy \u003csup\u003e[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]\u003c/sup\u003e. Nevertheless, up to now, reliable non-invasive tools for differentiating EC from AEH remain lacking \u003csup\u003e[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eMRI offers a non-invasive approach for evaluating endometrial abnormalities \u003csup\u003e[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]\u003c/sup\u003e. However, diagnostic accuracy is hampered by physiological variations and the marked overlap in imaging characteristics across endometrial lesions. \u003csup\u003e[\u003cspan additionalcitationids=\"CR12 CR13\" citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]\u003c/sup\u003e. For example, AEH frequently demonstrate signal intensity similar to normal endometrium on T2WI \u003csup\u003e[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]\u003c/sup\u003e. Moreover, MRI interpretation depends heavily on radiologist expertise, with reported accuracies ranging from 60% to 75% \u003csup\u003e[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]\u003c/sup\u003e. Therefore, a non-invasive adjunctive tool would be of significant clinical value for preoperative differentiation and treatment decision-making.\u003c/p\u003e \u003cp\u003eRecent advances in radiomics and convolutional neural networks (CNNs) have created new possibilities for image-based diagnosis. Radiomics extracts quantitative imaging features to characterize disease phenotypes and assist in classification \u003csup\u003e[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]\u003c/sup\u003e. CNNs automatically learn hierarchical imaging representations to support feature extraction and classification \u003csup\u003e[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]\u003c/sup\u003e. Besides, CNNs capture subtle spatial features and have demonstrated strong performance across tasks \u003csup\u003e[\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]\u003c/sup\u003e. Their effectiveness has been validated across multiple clinical applications, including breast cancer detection \u003csup\u003e[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e],\u003c/sup\u003e cervical cancer segmentation \u003csup\u003e[\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]\u003c/sup\u003e, and gastric disease classification \u003csup\u003e[\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eRecently, several studies have explored non-invasive differentiation of endometrial diseases. Zhang et al. developed an MRI-based radiomics model to differentiate EC from AEH, reporting an accuracy of 93.2%; however, its single-center design and limited sample size restrict its generalizability \u003csup\u003e[\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]\u003c/sup\u003e. Takahashi et al. prososed a hysteroscopy-based machine learning model that achieved accuracies of 85% for EC and 100% for AEH \u003csup\u003e[\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]\u003c/sup\u003e. Despite these advances, no MRI-based DL model has been developed specifically to discriminate EC from AEH.\u003c/p\u003e"},{"header":"Materials and method","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eData collection\u003c/h2\u003e \u003cp\u003eThis retrospective study was approved by the Ethics Committee of Jinshan Hospital, Fudan University (No. JIEC 2024-S45), and was conducted in accordance with the Declaration of Helsinki. The requirement for informed consent was waived by the Ethics Committee owing to the retrospective nature of the study. A total of 648 patients from Center A (January 2021\u0026ndash;December 2023), and 105 patients from the Center B (January 2020\u0026ndash;December 2023), were initially screened. All patients had undergone diagnostic curettage or hysterectomy, and all diagnoses of endometrial-related diseases were pathologically confirmed.\u003c/p\u003e \u003cp\u003eThe inclusion criteria were as follows: (1) Histopathologically confirmed stage I EC or AEH; (2) Preoperative multiparametric pelvic MRI, including axial T2-weighted imaging with fat saturation (T2WI-FS), diffusion-weighted imaging (DWI), apparent diffusion coefficient (ADC) mapping, and contrast-enhanced T1-weighted imaging (CE-T1WI); and (3) No history of prior radiotherapy, chemotherapy, hormonal therapy, or comorbid gynecological tumors. Exclusion criteria were: (1) Lesions too small for reliable delineation; (2) Poor-quality images; (3) Patients not classified as stage I EC; (4) Incomplete imaging sequences or export failure; and (5) Presence of other concurrent malignancies. Ultimately, 297 patients met the eligibility criteria, including 251 patients from Center A (124 stage I EC and 127 AHE) and 46 patients from Center B (29 stage I EC and 17 AHE). The overall workflow is illustrated in Supplementary Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eMRI Acquisition\u003c/h3\u003e\n\u003cp\u003eMRI examinations were performed using a 1.5-T and 3.0-T scanner (Avanto; Siemens, Erlangens, Germany) equipped with a phased-array abdominal coil. The detailed MRI acquisition parameters are provided in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. All patients were scanned in the supine position under free-breathing conditions. The routine MRI protocol included axial T2WI-FS, DWI with corresponding ADC maps, and CE-T1WI. DWI was performed with b-values of 0 and 1000 s/mm\u003csup\u003e2\u003c/sup\u003e. CE-T1WI images were obtained during the arterial, venous, and delayed phases, acquired immediately, 90\u0026ndash;120 seconds, and 150\u0026ndash;180 seconds after intravenous administration of gadopentetate dimeglumine at 0.2 mmol/kg at an injection rate of 2\u0026ndash;3 mL/s.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eMRI Equipment, Scanning Protocols, and Parameters at Different Centers\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"10\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCenter\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMRI field strength / model / brand\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMRI sequence\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eb s/mm\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSequence type\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eTR ms\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eTE ms\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003ematrix\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003eslice thickness mm\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c10\"\u003e \u003cp\u003eGAP\u003c/p\u003e \u003cp\u003emm\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003e1.5T\u003c/p\u003e \u003cp\u003eAvanto\u003c/p\u003e \u003cp\u003eSiemens\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eT2WI-FS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003e0, 1000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e4000\u0026ndash;6500\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e256*256\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e4.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDWI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eEPI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2400\u0026ndash;3300\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e176*176\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e1.5-2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eADC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eEP/RM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2900\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e176*176\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e5\u0026ndash;6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e1.5-2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCE-T1WI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eVIBE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e4.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e256*256\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eB\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003e1.5T\u003c/p\u003e \u003cp\u003eMAGNETOM\u003c/p\u003e \u003cp\u003eAera\u003c/p\u003e \u003cp\u003eSiemens\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eT2WI-FS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003e0, 800\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTSE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e4200\u0026ndash;5200\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e640*640\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e4.5-5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.9-1\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDWI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eEPI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e5100\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e336*336\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.8\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eADC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eEP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e4200\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e336*336\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e4.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0.8\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCE-T1WI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eVIBE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e4.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e640*640\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"10\"\u003e(ADC: apparent diffusion coefficient mapping;CE-T1WI༚contrast-enhanced T1-weighted imaging༛DWI༚diffusion-weighted imaging༛ MRI༚magnetic resonance imaging༛T2WI-FS༚T2-weighted imaging with fat-saturation༛TE༚echo time༛TR༚repetition time༛)\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e\n\u003ch3\u003eClinical Data Analysis\u003c/h3\u003e\n\u003cp\u003ePreoperative clinicopathological data were collected, including age, body mass index (BMI), obstetric history, menopausal status, hypertension, diabetes, hyperlipidemia, and tumor markers such as CA125 and CA199. Potential clinical predictors were initially screened using univariate analysis (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05) and were subsequently evaluated with multivariate stepwise logistic regression.\u003c/p\u003e\n\u003ch3\u003eRadiomics Feature Extraction, Selection, and Modeling\u003c/h3\u003e\n\u003cp\u003eThe unified analytical pipeline is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. All MRI data were imported into ITK-SNAP (version 4.2.0) for processing. DWI, ADC, and arterial-phase CE-T1WI images were rigidly registered to T2WI. Radiologist 1 manually delineated the regions of interest (ROIs) along the tumor boundaries on each T2WI slice, and applied the same segmentation method to the DWI, ADC, and CE-T1WI images while remaining blinded to all clinical information and pathological results. Radiologist 1 and radiologist 2 then independently performed a second round of segmentation on a randomly selected subset of 50 patients. During this process, they carefully excluded endometrial cavity fluid, hematoma, and adjacent normal myometrium, while allowing necrotic, hemorrhagic, or cystic components within the lesion to be included. Feature stability was calculated using both inter- and intraclass correlation coefficients (ICC), and only features with ICC value greater than 0.8 were retained.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eSegmentation of the volume of interest (VOI) encompassing the entire tumor was then performed. Before feature extraction, all original images were resampled to a voxel size of 1.5 \u0026times; 1.5 \u0026times; 1.5 mm\u0026sup3;. A total of 6,752 radiomics features were extracted, and standardized using z-score normalization. Redundant or irrelevant features were removed using the following sequential steps: Mann\u0026ndash;Whitney U test (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.01), correlation analysis (PyRadiomics, version 2.2.0), least absolute shrinkage and selection operator (LASSO) regression, and multivariate stepwise regression. The retained features were used to construct the Radiomics model. A combined clinical\u0026ndash;radiomics (Clin\u0026thinsp;+\u0026thinsp;Rad) model was then generated by integrating independent clinical predictors. Spearman correlation analysis was further used to eliminate features with high collinearity (|r| \u0026ge; 0.90).\u003c/p\u003e\n\u003ch3\u003eDL Feature Extraction, Selection, and Modeling\u003c/h3\u003e\n\u003cp\u003eA two-dimensional, slice-based strategy was adopted for DL feature extraction. Specifically, for each imaging sequence, the axial slice containing the largest cross-sectional tumor area was selected based on radiologist-annotated masks. The selected slice was then cropped using a bounding box tightly enclosing the lesion. Original images were resampled to a voxel size of 1.5 \u0026times; 1.5 \u0026times; 1.5 mm\u0026sup3; to standardize spatial resolution across subjects. Finally, all images were resized to 224 \u0026times;224 pixels and intensity-normalized for training, testing, and validating.\u003c/p\u003e \u003cp\u003eA DenseNet121-2D CNN pre-trained on ImageNet was used for feature extraction. All DL features were standardized using z-score normalization derived from the training cohort. Feature selection followed the same sequential approach as in the Radiomics pipeline: univariate filtering, correlation analysis, LASSO regression, and multivariate stepwise regression. The retained key DL features were used to construct the final DL model.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eStatistical Analysis\u003c/h2\u003e \u003cp\u003eStatistical analyses were performed using R software (version 4.4.0). Categorical variables were compared using chi-square test or Fisher\u0026rsquo;s exact test. Continuous variables were first assessed for normality, and analyzed using either the independent samples t-test or the Mann\u0026ndash;Whitney U test, as appropriate. Feature dimensionality reduction proceeded sequentially through inter- and intra-observer ICC assessment, univariate filtering, correlation analysis, LASSO regression, and multivariate stepwise regression. The extracted DenseNet121-2D features were incorporated into a logistic regression classifier to construct the DL-based model. Model performance was assessed using the area under the curve (AUC) of receiver operating characteristic (ROC) and calibration curves and decision curve analysis (DCA). Additional performance metrics included accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Confusion matrices (CM) were used to calculate true positive (TP), false positive (FP), false negative (FN), and true negative (TN). Differences in AUC values between models were assessed using DeLong\u0026rsquo;s test. A two-tailed P\u0026thinsp;\u0026lt;\u0026thinsp;0.05 was considered statistically significant.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003ePatient Characteristics and Feature Selection\u003c/h2\u003e \u003cp\u003eA total of 297 patients were included, comprising 153 patients with EC and 144 patients with AEH. The median age was 48 years (Q1-Q3: 34\u0026ndash;58 years). Baseline characteristics of patient with EC or AEH in the training, internal testing, and external validation cohorts are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Univariate and multivariate stepwise logistic regression analyses identified two age and menopausal status as independent clinical predictors. Serial feature screening yielded seven key radiomics features (two from T2WI and five from ADC images) and eight DL features, which were used to construct the corresponding models. LASSO regression results for radiomics feature selection and model calibration are presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparison of Clinical Characteristics Between EC and AEH Patients Across Cohorts\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"11\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c11\" colnum=\"11\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCharacteristics\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOverall\u003c/p\u003e \u003cp\u003e(N\u0026thinsp;=\u0026thinsp;297)\u003c/p\u003e \u003cp\u003e(144 AEH, 153 EC)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c4\" namest=\"c3\"\u003e \u003cp\u003eTraining (N\u0026thinsp;=\u0026thinsp;176)\u003c/p\u003e \u003cp\u003e(89 AEH, 87 EC)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e-value\u003csup\u003e1\u003c/sup\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e \u003cp\u003eInternal Testing (N\u0026thinsp;=\u0026thinsp;75)\u003c/p\u003e \u003cp\u003e(38 AEH, 37 EC)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e-value\u003csup\u003e1\u003c/sup\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c10\" namest=\"c9\"\u003e \u003cp\u003eExternal Validation\u003c/p\u003e \u003cp\u003e(N\u0026thinsp;=\u0026thinsp;46)\u003c/p\u003e \u003cp\u003e(17 AEH, 29 EC)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c11\"\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e -value\u003csup\u003e1\u003c/sup\u003e\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge,\u003c/p\u003e \u003cp\u003eMedian [IQR] \u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e48.0\u003c/p\u003e \u003cp\u003e[34.0,58.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e34.0\u003c/p\u003e \u003cp\u003e[30.0,40.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e58.0\u003c/p\u003e \u003cp\u003e[51.0,65.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e33.5\u003c/p\u003e \u003cp\u003e[30.0,39.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e56.0\u003c/p\u003e \u003cp\u003e[53.0,66.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e50.0\u003c/p\u003e \u003cp\u003e[45.0,58.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e55.0\u003c/p\u003e \u003cp\u003e[48.0,59.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e \u003cp\u003e0.322\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eObstetric history (Y/N) \u003csup\u003e3\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e99/194\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e58/28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6/80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e22/16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e6/31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e4/13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e3/26\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e \u003cp\u003e0.184\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMenopause (Y/N) \u003csup\u003e3\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e175/117\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e84/2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e21/65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e37/1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e7/29\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e12/5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e14/15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e \u003cp\u003e0.141\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMenopause Age,\u003c/p\u003e \u003cp\u003eMedian [IQR] \u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e50.0\u003c/p\u003e \u003cp\u003e[49.0,53.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e50.0\u003c/p\u003e \u003cp\u003e[49.0,51.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e50.0\u003c/p\u003e \u003cp\u003e[49.0,52.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.837\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e50.0\u003c/p\u003e \u003cp\u003e[50.0,50.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e51.0\u003c/p\u003e \u003cp\u003e[49.0,53.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.683\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e \u003cp\u003e0.948\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBMI,\u003c/p\u003e \u003cp\u003eMedian [IQR] \u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e24.8\u003c/p\u003e \u003cp\u003e[22.1,28.3]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e26.1\u003c/p\u003e \u003cp\u003e[22.0,30.3]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e24.3\u003c/p\u003e \u003cp\u003e[22.2,27.4]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.130\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e25.4\u003c/p\u003e \u003cp\u003e[22.0,28.2]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e25.3\u003c/p\u003e \u003cp\u003e[22.3,27.5]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.423\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eNA\u003c/p\u003e \u003cp\u003e[NA,NA]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003eNA\u003c/p\u003e \u003cp\u003e[NA,NA]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c11\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHypertension (Y/N)\u003csup\u003e3\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e218/77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e78/10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e56/30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e35/3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e23/14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.002\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e11/6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e15/14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e \u003cp\u003e0.391\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiabetes (Y/N) \u003csup\u003e3\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e254/41\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e81/7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e66/20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.005\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e35/3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e30/7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.191\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e \u003cp\u003e0.619\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHyperlipidemia (Y/N)\u003csup\u003e3\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e281/14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e85/3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e81/5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.494\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e37/1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e33/4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.200\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e16/1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e29/0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e \u003cp\u003e0.370\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCA125 (+/-)\u003csup\u003e4\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e239/31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e74/7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e75/9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.653\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e31/4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e26/7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.274\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e12/1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e21/3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e \u003cp\u003e1.000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCA199 (+/-)\u003csup\u003e4\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e137/28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e23/1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e53/14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.105\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e11/1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e18/8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.223\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e10/2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e22/2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e \u003cp\u003e0.588\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHPV (Y/N) \u003csup\u003e3\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e126/13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e36/5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e47/3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.460\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e13/1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e17/3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.627\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e3/0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e10/1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e \u003cp\u003e1.000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003csup\u003e1\u003c/sup\u003eWilcoxon rank sum test; Pearson's Chi-squared test; Fisher's exact test; \u003csup\u003e2\u003c/sup\u003eData are presented as median [Q1\u0026ndash;Q3]; IQR denotes interquartile range (25th\u0026ndash;75th percentiles); \u003csup\u003e3\u003c/sup\u003e(Y/N) represents yes or no of the corresponding clinical condition; \u003csup\u003e4\u003c/sup\u003e(+/\u0026minus;) represents positive or negative results, respectively.\u003c/p\u003e \u003cp\u003eAEH: atypical endometrial hyperplasia; BMI: body mass index; EC: endometrial carcinoma; HPV: human papillomavirus; NA: not available\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e(A-B) The coefficient paths demonstrate the effect of LASSO regularization on feature selection and model calibration. As λ (Log λ) increases, more coefficients shrink to zero, effectively removing redundant or weakly contributing features. (C-D) Cross-validation curves show that the binomial deviance initially decreases as λ increases\u0026mdash;indicating improved model fit\u0026mdash;then rises when λ becomes too large due to over-regularization. This pattern underscores the importance of selecting an optimal λ that balance overfitting and underfitting. (LASSO: least absolute shrinkage and selection operator)\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eModel Performance\u003c/h2\u003e \u003cp\u003eThe diagnostic performances of the models, along with the CMs and ROC curves for differentiating EC from AEH, are shown in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. ROC curve analysis demonstrated that although the DenseNet121-2D model showed slightly lower AUCs than the Radiomics and Clin\u0026thinsp;+\u0026thinsp;Rad models in the training cohort (0.882 vs. 0.913 vs. 0.973) and the internal testing cohort (0.821 vs. 0.919 vs. 0.979), it outperformed both models in the external validation cohort, achieving an AUC of 0.854, compared with 0.80 for the Radiomics model and 0.77 for the Clin\u0026thinsp;+\u0026thinsp;Rad model. In the external validation cohort, DenseNet121-2D model also demonstrated the highest sensitivity (55.2%) and highest specificity (94.1%), exceeding the performance of both the Radiomics model (sensitivity 31.0%, specificity 94.1%) and the Clin\u0026thinsp;+\u0026thinsp;Rad model (sensitivity 58.6%, specificity 88.2%).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePredictive Performance of Different Models in the Training, Internal Testing and External Validation Cohorts\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAccuracy (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSensitivity (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSpecificity (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePPV (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNPV (%)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colspan=\"7\" nameend=\"c7\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTraining (n\u0026thinsp;=\u0026thinsp;176)\u003c/b\u003e\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRadiomics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.913\u003c/p\u003e \u003cp\u003e(0.869\u0026ndash;0.956)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e85.2\u003c/p\u003e \u003cp\u003e(79.1\u0026ndash;90.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e80.5\u003c/p\u003e \u003cp\u003e(69.0-89.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e89.9\u003c/p\u003e \u003cp\u003e(76.4\u0026ndash;97.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e88.6\u003c/p\u003e \u003cp\u003e(87.0-89.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e82.5\u003c/p\u003e \u003cp\u003e(80.0-83.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClin\u0026thinsp;+\u0026thinsp;Rad\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.973\u003c/p\u003e \u003cp\u003e(0.953\u0026ndash;0.993)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e92.6\u003c/p\u003e \u003cp\u003e(87.7\u0026ndash;96.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e89.7\u003c/p\u003e \u003cp\u003e(81.6\u0026ndash;96.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e95.5\u003c/p\u003e \u003cp\u003e(75.3\u0026ndash;100.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e95.1\u003c/p\u003e \u003cp\u003e(94.7\u0026ndash;95.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e90.4\u003c/p\u003e \u003cp\u003e(88.2\u0026ndash;90.8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDenseNet 121-2D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.882\u003c/p\u003e \u003cp\u003e(0.831\u0026ndash;0.932)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e81.8\u003c/p\u003e \u003cp\u003e(75.3\u0026ndash;87.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e77.0\u003c/p\u003e \u003cp\u003e(58.6\u0026ndash;86.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e86.5\u003c/p\u003e \u003cp\u003e(71.9\u0026ndash;93.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e84.8\u003c/p\u003e \u003cp\u003e(81.0-86.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e79.4\u003c/p\u003e \u003cp\u003e(76.2\u0026ndash;80.6)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"7\" nameend=\"c7\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eInternal Testing (n\u0026thinsp;=\u0026thinsp;75)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRadiomics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.919\u003c/p\u003e \u003cp\u003e(0.860\u0026ndash;0.978)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e81.3\u003c/p\u003e \u003cp\u003e(70.7\u0026ndash;89.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e78.4\u003c/p\u003e \u003cp\u003e(59.5\u0026ndash;97.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e84.2\u003c/p\u003e \u003cp\u003e(73.6\u0026ndash;100.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e82.9\u003c/p\u003e \u003cp\u003e(78.6\u0026ndash;85.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e80.0\u003c/p\u003e \u003cp\u003e(77.8\u0026ndash;82.6)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClin\u0026thinsp;+\u0026thinsp;Rad\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.979\u003c/p\u003e \u003cp\u003e(0.950-1.000)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e93.3\u003c/p\u003e \u003cp\u003e(85.1\u0026ndash;97.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e94.6\u003c/p\u003e \u003cp\u003e(62.2\u0026ndash;100.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e92.1\u003c/p\u003e \u003cp\u003e(84.2\u0026ndash;100.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e92.1\u003c/p\u003e \u003cp\u003e(88.5\u0026ndash;92.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e94.6\u003c/p\u003e \u003cp\u003e(94.1\u0026ndash;95.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDenseNet 121-2D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.821\u003c/p\u003e \u003cp\u003e(0.726\u0026ndash;0.915)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e72.0\u003c/p\u003e \u003cp\u003e(60.4\u0026ndash;81.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e62.2\u003c/p\u003e \u003cp\u003e(40.5\u0026ndash;86.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e81.6\u003c/p\u003e \u003cp\u003e(60.5\u0026ndash;97.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e76.7\u003c/p\u003e \u003cp\u003e(68.2\u0026ndash;82.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e68.9\u003c/p\u003e \u003cp\u003e(62.1\u0026ndash;72.5)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"7\" nameend=\"c7\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eExternal Validation (n\u0026thinsp;=\u0026thinsp;46)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRadiomics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.803\u003c/p\u003e \u003cp\u003e(0.673\u0026ndash;0.933)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e54.3\u003c/p\u003e \u003cp\u003e(39.0-69.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e31.0\u003c/p\u003e \u003cp\u003e(20.7\u0026ndash;69.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e94.1\u003c/p\u003e \u003cp\u003e(82.4\u0026ndash;100.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e90.0\u003c/p\u003e \u003cp\u003e(85.7\u0026ndash;95.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e44.4\u003c/p\u003e \u003cp\u003e(41.2\u0026ndash;45.9)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClin\u0026thinsp;+\u0026thinsp;Rad\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.773\u003c/p\u003e \u003cp\u003e(0.626\u0026ndash;0.919)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e69.6\u003c/p\u003e \u003cp\u003e(54.2\u0026ndash;82.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e58.6\u003c/p\u003e \u003cp\u003e(10.3\u0026ndash;75.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e88.2\u003c/p\u003e \u003cp\u003e(46.9\u0026ndash;100.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e89.5\u003c/p\u003e \u003cp\u003e(60.0-91.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e55.6\u003c/p\u003e \u003cp\u003e(39.9\u0026ndash;58.6)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDenseNet 121-2D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.854\u003c/p\u003e \u003cp\u003e(0.743\u0026ndash;0.965)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e69.6\u003c/p\u003e \u003cp\u003e(54.2\u0026ndash;82.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e55.2\u003c/p\u003e \u003cp\u003e(34.5\u0026ndash;86.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e94.1\u003c/p\u003e \u003cp\u003e(82.4\u0026ndash;100.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e94.1\u003c/p\u003e \u003cp\u003e(90.9\u0026ndash;96.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e55.2\u003c/p\u003e \u003cp\u003e(51.9\u0026ndash;56.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eAUC: area under the receiver operating characteristic curve; PPV: positive predictive value; NPV: negative predictive value; Clin\u0026thinsp;+\u0026thinsp;Rad: clinical-radiomics model.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e(A): CM of the Radiomics, Clin\u0026thinsp;+\u0026thinsp;Rad, and DenseNet 121 models in the training, internal testing, and external validation cohorts. The color intensity corresponds to the number of cases in each cell: higher values are represented by darker color. (B): ROC curves of three models for distinguishing EC from AEH in the training (a), internal testing (b), external validation cohorts (c). (CM: Confusion matrix; ROC: receiver operating characteristic; Clin\u0026thinsp;+\u0026thinsp;Rad: Clinical+Radiomics.)\u003c/p\u003e \u003cp\u003eAccording to the DeLong test, the Clin\u0026thinsp;+\u0026thinsp;Rad model showed significantly better diagnostic performance than the Radiomics model (P\u0026thinsp;=\u0026thinsp;0.001) and the DenseNet121-2D model (P\u0026thinsp;\u0026lt;\u0026thinsp;0.001) in the training cohort, as shown in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. However, in the external validation cohort, the DeLong test showed no statistically significant differences in AUC among the models. Nevertheless, additional analyses based on net reclassification improvement (NRI) and integrated discrimination improvement (IDI) revealed that the IDI of DenseNet121-2D relative to the Radiomics model remained statistically significant (P\u0026thinsp;=\u0026thinsp;0.022), whereas the corresponding NRI was not significant. Furthermore, no significant differences in either NRI or IDI were found between DenseNet121-2D and the Clin\u0026thinsp;+\u0026thinsp;Rad model. These results indicate that the incremental diagnostic benefit of DenseNet121-2D in the external validation cohort was limited, although its potential value should be further investigated in a larger sample. Detailed results are provided in Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDeLong Test Results for Pairwise Comparisons of AUCs among Different Models in Different Cohorts\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eComparative performance of different models\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c4\" namest=\"c2\"\u003e \u003cp\u003eP value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTraining\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eInternal Testing\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eExternal Validation\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRadiomics vs Clin\u0026thinsp;+\u0026thinsp;Rad\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.038\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.668\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRadiomics vs DenseNet121_2D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.298\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.025\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.447\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDenseNet121_2D vs Clin\u0026thinsp;+\u0026thinsp;Rad\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.002\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.315\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparison of NRI and IDI for Incremental Diagnostic Performance among Different Models across Different Cohorts\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDifferent model\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCohort\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNRI (95% CI)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eP value\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eIDI (95% CI)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eP value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eDenseNet121_2D vs Radiomics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTraining\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-0.079\u003c/p\u003e \u003cp\u003e(-0.254, 0.097)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.380\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e-0.092\u003c/p\u003e \u003cp\u003e(-0.172, -0.011)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.025\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eInternal Testing\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-0.349\u003c/p\u003e \u003cp\u003e(-0.627, -0.070)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.014\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e-0.186\u003c/p\u003e \u003cp\u003e(-0.324, -0.049)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.008\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eExternal Validation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.227\u003c/p\u003e \u003cp\u003e(-0.023, 0.477)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.075\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.135\u003c/p\u003e \u003cp\u003e(0.019, 0.250)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.022\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eDenseNet121_2D vs Clin\u0026thinsp;+\u0026thinsp;Rad\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTraining\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-0.374\u003c/p\u003e \u003cp\u003e(-0.538, -0.210)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e-0.323\u003c/p\u003e \u003cp\u003e(-0.406, -0.239)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eInternal Testing\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-0.721\u003c/p\u003e \u003cp\u003e(-0.983, -0.459)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e-0.453\u003c/p\u003e \u003cp\u003e(-0.612, -0.294)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eExternal Validation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.059\u003c/p\u003e \u003cp\u003e(-0.262, 0.380)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.719\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.007\u003c/p\u003e \u003cp\u003e(-0.196, 0.211)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.943\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eRadiomics\u003c/p\u003e \u003cp\u003evs Clin\u0026thinsp;+\u0026thinsp;Rad\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTraining\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.351\u003c/p\u003e \u003cp\u003e(0.216, 0.488)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.231\u003c/p\u003e \u003cp\u003e(0.170, 0.291)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eInternal Testing\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.426\u003c/p\u003e \u003cp\u003e(0.201, 0.651)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.267\u003c/p\u003e \u003cp\u003e(0.147, 0.386)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eExternal Validation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.099\u003c/p\u003e \u003cp\u003e(-0.179, 0.378)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.484\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.128\u003c/p\u003e \u003cp\u003e(-0.047, 0.302)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.151\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"6\"\u003eNote: NRI and IDI were used to assess the incremental diagnostic value of the comparison model relative to the reference model. NRI and IDI additionally reflect changes in case reclassification and overall discriminative ability, and may therefore complement AUC-based comparisons. NRI/IDI\u0026thinsp;\u0026gt;\u0026thinsp;0 indicates that the former model outperforms the latter in classification or discrimination ability, whereas NRI/IDI\u0026thinsp;\u0026lt;\u0026thinsp;0 indicates poorer performance of the former model relative to the latter. (95% CI: 95% confidence Interval; NRI: Net Reclassification Improvement; IDI: Integrated Discrimination Improvement;)\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eCalibration and Clinical Utility\u003c/h2\u003e \u003cp\u003eAs shown in Fig.\u0026nbsp; 4A-4C, the calibration curves for both the Radiomics and DenseNet121-2D models closely aligned with the ideal reference line in the training and internal testing cohorts, indicating good agreement between predicted and observed outcomes. However, in the external validation cohort, the DenseNet121-2D model demonstrated superior calibration, reflecting greater reliability and generalizability. This advantage was further supported by DCA (Fig.\u0026nbsp;4D-F), which demonstrated that the DenseNet121-2D model yielded the highest net clinical benefit for differentiating EC from AEH across a broad range of threshold probabilities.\u003c/p\u003e \u003cp\u003e \u003cb\u003eFigure 4.\u003c/b\u003e Model performance assessed by calibration curve and DCA\u003c/p\u003e \u003cp\u003e(A-C) Calibration curves of the Radiomics, Clin\u0026thinsp;+\u0026thinsp;Rad, and DenseNet121-2D models in the training (A), internal testing (B), and external validation (C) cohorts. The DenseNet121-2D model demonstrats better calibration and generalizability in the external validation cohort compared with the Clin\u0026thinsp;+\u0026thinsp;Rad and Radiomics models. (D-F) DCA results for. the training (D), internal testing (E), and external validation (F) cohorts. The DenseNet121-2D model yields a consistently higher net benefit than the other models at threshold probabilities below 0.6. (DCA: decision curve analysis)\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eCorrelation Analysis and Model Robustness\u003c/h2\u003e \u003cp\u003eTo further explore feature relationships and enhance model interpretability, we examined the linear associations between key features and the target variable using Pearson correlation and constructing a correlation matrix (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e5\u003c/span\u003e). The DenseNet121-2D model demonstrated low inter-feature correlations (|r| \u0026lt; 0.6), suggesting minimal redundancy among the DL features. In contrast, several radiomics and Clin\u0026thinsp;+\u0026thinsp;Rad features exhibited strong linear dependencies (|r| \u0026gt; 0.7), indicating potential multicollinearity, which may contribute to reduced generalizability and overfitting. This aligns with model performance: despite high AUCs in the training and internal testing cohorts (\u0026gt;\u0026thinsp;0.90), both the Radiomics and Clin\u0026thinsp;+\u0026thinsp;Rad models showed a markedly declined in the external validation cohort (AUC\u0026thinsp;=\u0026thinsp;0.80 and 0.773), whereas the DenseNet121-2D model maintained more stable performance.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe figure presents pairwise Pearson correlation coefficients for the Radiomics (A), Clin\u0026thinsp;+\u0026thinsp;Rad (B), and DenseNet121-2D (C) feature sets. Color intensity represents both the strength and direction of the correlations, with red indicating positive correlations and blue indicating negative correlations. A diagonal value of 1.00 reflects the perfect self-correlation of each feature.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003ePanels correspond to the training (A, D), internal testing (B, E), and external validation (C, F) cohorts. (A-C) Bee-swarm plots illustrating the distribution and influence of each feature on model predictions based on SHAP values. (D-F) Bar plots showing the mean SHAP values, reflecting the overall importance of each feature. The y-axis lists the model features, and the x-axis displays the mean SHAP values, indicating the average contribution of each feature to the model output. (SHAP: shapley additive explanations)\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eSHAP Interpretability Analysis\u003c/h2\u003e \u003cp\u003eTo further evaluate model interpretability, SHAP analysis was applied to the DL model (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e6\u003c/span\u003e). Among all extracted features, the most influential were DWI-F253, ADC-F961, ADC-F502, T2WI-FS-F87, and ADC-F16 in the training cohort. Notably, these same five features remained the top contributors in the external validation cohort, although with varying magnitude. In the internal testing cohort, DWI-F253, T2WI-FS-F87, ADC-F961, and ADC-F502 continued to rank among the most important features. This recurrent identification of key features across cohorts supports the robustness and potential generalizability of the DenseNet121-2D model.\u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we constructed and validated a DL-based diagnostic model (DenseNet121-2D) that can automatically differentiate endometrial carcinoma (EC) from atypical endometrial hyperplasia (AEH) using datasets from two centers. The model demonstrated satisfactory performance in the training cohort (AUC, 0.882)) and internal testing cohort (AUC, 0.821). Importantly, in the external validation cohort, the DenseNet121-2D model outperformed both the Radiomics and Clin\u0026thinsp;+\u0026thinsp;Rad models (AUC, 0.854 vs. 0.803 vs. 0.773), confirming its generalizability and robust performance across datasets. This method may help reduce misdiagnosis and improve diagnostic efficiency, and potentially decrease unnecessary surgeries for patients with AEH. Given its underlying architecture, the proposed model could be further extended to support intelligent, image-based auxiliary diagnoses for other diseases.\u003c/p\u003e \u003cp\u003eIn the present study, we built both a Radiomics model and a DenseNet121-2D model to differentiate patients with EC and AEH. The reliability of image-derived features is critical for building robust radiomics and DL models \u003csup\u003e[\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]\u003c/sup\u003e. However, processes such as target segmentation, feature extraction, feature selection, and model construction are inherently subject to inter- and intra-observer variability \u003csup\u003e[\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]\u003c/sup\u003e. To address this limitation, manual segmentation was adopted due to the relatively limited sample size, and the intraclass correlation coefficient (ICC) was calculated to enhance feature robustness and reproducibility \u003csup\u003e[\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]\u003c/sup\u003e. Furthermore, feature selection was initially conducted using logistic regression and subsequently refined through multivariate stepwise regression. A Pearson correlation heatmap was used to ensure selected features did not exhibit high collinearity. In addition, Shapley Additive Explanations (SHAP) analysis was used to improve the interpretability and transparency of the final models. Radiomics approaches have previously been applied to endometrial disease classification, generally showing strong performance for EC assessment. For example, Liu et al. constructed an MRI-based radiomics model and a visual nomogram, which effectively differentiated patients with benign and malignant endometrial lesions \u003csup\u003e[\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]\u003c/sup\u003e. However, the diagnostic accuracy of radiomics has been reported to be limited to 78.0% in test cohorts \u003csup\u003e[\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]\u003c/sup\u003e, and often suffer from limited generalizability and a lack of robustness across different datasets. Yan et al. developed a radiomics nomogram for fertility-preserving treatment candidate selection in EC and AEH, achieving an AUC of 0.94 in a prospective validation cohort\u0026mdash;likely attributable to the reduced biases in prospective study \u003csup\u003e[\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eThe DenseNet121-2D model was trained using an axial slice with maximum tumor area from T2WI-FS, DWI, ADC, and CE-T1WI. Among these sequences, CE-T1WI and DWI provide essential functional information in the evaluation of EC and are frequently used as complementary tools to T2WI \u003csup\u003e[\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]\u003c/sup\u003e. Tumor enhancement patterns on CE-T1WI primarily reflect vascularity: although most EC are hypovascular, some lesions may appear iso- or hypervascular relative to the surrounding myometrium \u003csup\u003e[\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]\u003c/sup\u003e. In addition, DWI and ADC mainly reflect water diffusion characteristics, ADC values are negatively associated with tumor cell density \u003csup\u003e[\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]\u003c/sup\u003e, and prior studies have shown that ADC values in EC are significantly lower than those observed in AEH \u003csup\u003e[\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e, \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]\u003c/sup\u003e. Studies have demonstrated that CNN-based methods achieved their highest diagnostic performance when trained on single axial ADC maps \u003csup\u003e[\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]\u003c/sup\u003e. Therefore, incorporating CE-T1WI, DWI, and ADC maps alongside T2WI facilitates a more accurate, comprehensive, and reliable diagnostic assessment of EC.\u003c/p\u003e \u003cp\u003eGiven the strong diagnostic contribution of DWI and ADC-derived information discussed above, it is not surprising that a substantial proportion of discriminative features originated from the ADC maps in our study. This finding is clinically plausible, as EC typically demonstrates restricted diffusion which appears as hyperintensity relative to the surrounding myometrium on DWI and corresponding hypointensity on ADC maps \u003csup\u003e[\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]\u003c/sup\u003e. Previous studies have consistently underscored the value of ADC measurements in differentiating benign from malignant endometrial lesions \u003csup\u003e[\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]\u003c/sup\u003e. For example, Tamai et al. reported that normal endometrium exhibited significantly higher ADC values compared with EC, supporting ADC as an initial imaging biomarker \u003csup\u003e[\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eThe robustness of ADC-derived radiomics features has also been validated across centers. In a multicenter study, Peerlings et al. reported that 122 ADC-based features demonstrated high reproducibility (ICC\u0026thinsp;\u0026gt;\u0026thinsp;0.85) across different tissues, scanners, and vendors \u003csup\u003e[\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]\u003c/sup\u003e. This observation is consistent with our findings and explains the preferentially retention of ADC features during feature selection. Biologically, ADC values reflect the extent of water diffusion restriction, which becomes more pronounced in tumors with high cellularity and reduced extracellular space. A meta-analysis by Surov et al. confirmed a significant negative correlation between mean ADC values and tumor cellularity \u003csup\u003e[\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]\u003c/sup\u003e. Taken together, these biological and technical factors underscore the reliability and clinical relevance of ADC features in differentiating EC from AEH.\u003c/p\u003e \u003cp\u003eThe DenseNet121-2D model achieved the best predictive performance in the external validation cohort (AUC, 85.4%), representing a notable improvement over the Radiomics model (AUC 80.3%) and demonstrating its superior generalizability. As outlined earlier, numerous DL architectures have been successfully applied to medical imaging tasks. For example, Takahashi et al. developed a DL algorithm trained on more than 411,800 hysteroscopic images, achieving strong discriminatory capacity across five endometrial disease categories \u003csup\u003e[\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]\u003c/sup\u003e. Other work has applied DL models to pathological image, with high PPV (82.94%) and NPV (96.53%) both demonstrating strong accuracy \u003csup\u003e[\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e]\u003c/sup\u003e. In contrast, the NPV of all models in our study was substantially lower than the PPV. This discrepancy may be attributable to sample imbalance, as positive cases were more prevalent\u0026mdash;particularly in the external validation cohort, where EC cases was twice that of AEH cases. To date, no MRI-based DL model has been reported for differentiating EC from AEH, highlighting the novelty and clinical significance of our work.\u003c/p\u003e \u003cp\u003eThere are also several limitations in our study. First, the generalizability of our findings is limited by the relatively small sample size, especially in the external validation cohort. Second, prospective validation was not performed, which may restrict the strength of our conclusions. Further study is warrant in a prospective, multi-center studies with larger and more heterogeneous dataset. Third, manual lesion segmentation is both time-consuming and susceptible to inter-observer variability. Future studies should explore automatic or semi-automatic segmentation approaches that provide more efficient and reproducible delineation.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eIn this study, we developed the first MRI-based CNN model specifically designed to discriminate EC from AEH. The proposed DenseNet121-2D architecture demonstrated superior diagnostic performance compared with both Radiomics model and Clin\u0026thinsp;+\u0026thinsp;Rad model, particularly in the external validation cohort. Furthermore, the model offers objective, quantitative evidence to support radiologists' decision-making, highlighting its potential for further clinical application.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eAll authors have approved the final version of the manuscript and agree with its submission to \u003cem\u003eBMC MEDICAL IMAGING\u003c/em\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was supported by the National Natural Science Foundation of China (Grant No. 82471943) and the Shanghai Jinshan District Health Commission (JSZK2023A02).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis retrospective study was approved by the Ethics Committee of Jinshan Hospital, Fudan University (No. JIEC 2024-S45), and was conducted in accordance with the Declaration of Helsinki. The requirement for informed consent was waived by the Ethics Committee owing to the retrospective nature of the study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026rsquo; contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e(I) Conception and design: Dan Hong Chai; Zi Jing Lin\u003c/p\u003e\n\u003cp\u003e(II) Administrative support: Jin Wei Qiang\u003c/p\u003e\n\u003cp\u003e(III) Provision of study materials or patients: Feng Hua Ma; Guo Fu Zhang; Zhao Xia Qian\u003c/p\u003e\n\u003cp\u003e(IV) Collection and assembly of data: Dan Hong, Chai; Yan Chen;\u003c/p\u003e\n\u003cp\u003e(V) Data analysis and interpretation: Jin Wei Qiang\u003c/p\u003e\n\u003cp\u003e(VI) Manuscript writing: All authors\u003c/p\u003e\n\u003cp\u003e(VII) Final approval of manuscript: All authors\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eFerlay J, Soerjomataram I, Dikshit R. Cancer incidence and mortality worldwide: sources, methods and major patterns in GLOBOCAN 2012. Int J Cancer. 2015;136(5):E359\u0026ndash;86. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/ijc.29210\u003c/span\u003e\u003cspan address=\"10.1002/ijc.29210\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBosse T, Peters EE, Creutzberg CL. Substantial lymph-vascular space invasion (LVSI) is a significant risk factor for recurrence in endometrial cancer\u0026ndash;A pooled analysis of PORTEC 1 and 2 trials. Eur J Cancer. 2015;51(13):1742\u0026ndash;50. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ejca.2015.05.015\u003c/span\u003e\u003cspan address=\"10.1016/j.ejca.2015.05.015\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCrosbie EJ, Kitson SJ, McAlpine JN. Endometrial cancer. Lancet. 2022;399(10333):1412\u0026ndash;28. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/S0140-6736(22)00323-3\u003c/span\u003e\u003cspan address=\"10.1016/S0140-6736(22)00323-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBhardwaj V, Sharma A, Parambath SV. Machine Learning for Endometrial Cancer Prediction and Prognostication. Front Oncol. 2022. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/fonc.2022.852746\u003c/span\u003e\u003cspan address=\"10.3389/fonc.2022.852746\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. 27;12:852746.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHerrington CS, editor. WHO classification of tumours: female genital tumours. 5th ed. Lyon: International Agency for Research on Cancer; 2020.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSvirsky R, Smorgick N, Rozowski U. Can we rely on blind endometrial biopsy for detection of focal intrauterine pathology? Am J Obstet Gynecol. 2008;199(2):115. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ajog.2008.02.015\u003c/span\u003e\u003cspan address=\"10.1016/j.ajog.2008.02.015\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. .e1-3.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTrimble CL, Kauderer J, Zaino R. Concurrent endometrial carcinoma in women with a biopsy diagnosis of atypical endometrial hyperplasia: a Gynecologic Oncology Group study. Cancer. 2006;106(4):812\u0026ndash;9. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/cncr.21650\u003c/span\u003e\u003cspan address=\"10.1002/cncr.21650\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRakha E, Wong SC, Soomro I. Clinical outcome of atypical endometrial hyperplasia diagnosed on an endometrial biopsy: institutional experience and review of literature. Am J Surg Pathol. 2012;36(11):1683\u0026ndash;90. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1097/PAS.0b013e31825dd4ff\u003c/span\u003e\u003cspan address=\"10.1097/PAS.0b013e31825dd4ff\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGentry-Maharaj A, Karpinskyj C. Current and future approaches to screening for endometrial cancer. Best Pract Res Clin Obstet Gynaecol. 2020;65:79\u0026ndash;97. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.bpobgyn.2019.12.006\u003c/span\u003e\u003cspan address=\"10.1016/j.bpobgyn.2019.12.006\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHaldorsen IS, Salvesen HB. What Is the Best Preoperative Imaging for Endometrial Cancer? Curr Oncol Rep. 2016;18(4):25. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s11912-016-0506-0\u003c/span\u003e\u003cspan address=\"10.1007/s11912-016-0506-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNalaboff KM, Pellerito JS, Ben-Levi E. Imaging the endometrium: disease and normal variants. Radiographics. 2001;21(6):1409\u0026ndash;24. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1148/radiographics.21.6.g01nv211409\u003c/span\u003e\u003cspan address=\"10.1148/radiographics.21.6.g01nv211409\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKierans AS, Bennett GL, Haghighi M, Rosenkrantz AB. Utility of conventional and diffusion-weighted MRI features in distinguishing benign from malignant endometrial lesions. Eur J Radiol. 2014;83(4):726\u0026ndash;32. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ejrad.2013.11.030\u003c/span\u003e\u003cspan address=\"10.1016/j.ejrad.2013.11.030\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBakir B, Sanli S, Bakir VL. Role of diffusion weighted MRI in the differential diagnosis of endometrial cancer, polyp, hyperplasia, and physiological thickening. Clin Imaging. 2017;41:86\u0026ndash;94. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.clinimag.2016.10.016\u003c/span\u003e\u003cspan address=\"10.1016/j.clinimag.2016.10.016\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang Y, Wang Z, Zhang J. Deep learning model for classifying endometrial lesions. J Transl Med. 2021;19(1):10. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s12967-020-02660-x\u003c/span\u003e\u003cspan address=\"10.1186/s12967-020-02660-x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTakeuchi M, Matsuzaki K, Uehara H, Yoshida S, Nishitani H, Shimazu H. Pathologies of the uterine endometrial cavity: usual and unusual manifestations and pitfalls on magnetic resonance imaging. Eur Radiol. 2005;15(11):2244\u0026ndash;55. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s00330-005-2814-x\u003c/span\u003e\u003cspan address=\"10.1007/s00330-005-2814-x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDe Marchi F, Fabris AM, Tommasi L, Nappi L, Saccardi C, Litta P. Accuracy of hysteroscopy made by young residents in detecting endometrial pathologies in postmenopausal women. Eur J Gynaecol Oncol. 2014;35(3):219\u0026ndash;23.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWiestler B, Menze B. Deep learning for medical image analysis: a brief introduction. Neurooncol Adv. 2021;2(Suppl 4):iv35\u0026ndash;41. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1093/noajnl/vdaa092\u003c/span\u003e\u003cspan address=\"10.1093/noajnl/vdaa092\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLitjens G, Kooi T, Bejnordi BE. A survey on deep learning in medical image analysis. Med Image Anal. 2017;42:60\u0026ndash;88. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.media.2017.07.005\u003c/span\u003e\u003cspan address=\"10.1016/j.media.2017.07.005\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMazurowski MA, Buda M, Saha A, Bashir MR. Deep learning in radiology: An overview of the concepts and a survey of the state of the art with focus on MRI. J Magn Reson Imaging. 2019;49(4):939\u0026ndash;54. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/jmri.26534\u003c/span\u003e\u003cspan address=\"10.1002/jmri.26534\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChougrad H, Zouaki H, Alheyane O. Deep Convolutional Neural Networks for breast cancer screening. Comput Methods Programs Biomed. 2018;157:19\u0026ndash;30. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.cmpb.2018.01.011\u003c/span\u003e\u003cspan address=\"10.1016/j.cmpb.2018.01.011\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRodr\u0026iacute;guez Outeiral R, Gonz\u0026aacute;lez PJ, Schaake EE, van der Heide UA, Sim\u0026otilde;es R. Deep learning for segmentation of the cervical cancer gross tumor volume on magnetic resonance imaging for brachytherapy. Radiat Oncol. 2013;18(1):91. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s13014-023-02283-8\u003c/span\u003e\u003cspan address=\"10.1186/s13014-023-02283-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCho BJ, Bang CS, Park SW. Automated classification of gastric neoplasms in endoscopic images using a convolutional neural network. Endoscopy. 2019;51(12):1121\u0026ndash;9. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1055/a-0981-6133\u003c/span\u003e\u003cspan address=\"10.1055/a-0981-6133\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang J, Zhang Q, Wang T. Multimodal MRI-Based Radiomics-Clinical Model for Preoperatively Differentiating Concurrent Endometrial Carcinoma From Atypical Endometrial Hyperplasia. Front Oncol. 2022;12:887546. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/fonc.2022.887546\u003c/span\u003e\u003cspan address=\"10.3389/fonc.2022.887546\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTakahashi Y, Sone K, Noda K. Automated system for diagnosing endometrial cancer by adopting deep-learning technology in hysteroscopy. PLoS ONE. 2021;16(3):e0248526. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1371/journal.pone.0248526\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0248526\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003evan Timmeren JE, Cester D, Tanadini-Lang S, Alkadhi H, Baessler B. Radiomics in medical imaging-how-to guide and critical reflection. Insights Imaging. 2020;11(1):91. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s13244-020-00887-2\u003c/span\u003e\u003cspan address=\"10.1186/s13244-020-00887-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu J, Li S, Lin H. Development of MRI-based radiomics predictive model for classifying endometrial lesions. Sci Rep. 2023;13(1):1590. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41598-023-28819-2\u003c/span\u003e\u003cspan address=\"10.1038/s41598-023-28819-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eUrushibara A, Saida T, Mori K. The efficacy of deep learning models in the diagnosis of endometrial cancer using MRI: a comparison with radiologists. BMC Med Imaging. 2022;22(1):80. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s12880-022-00808-3\u003c/span\u003e\u003cspan address=\"10.1186/s12880-022-00808-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYan BC, Ma FH, Li Y. An MRI radiomics nomogram improves the accuracy in identifying eligible candidates for fertility-preserving treatment in endometrioid adenocarcinoma. Am J Cancer Res. 2022;12(3):1056\u0026ndash;68.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWhittaker CS, Coady A, Culver L, Rustin G, Padwick M, Padhani AR. Diffusion-weighted MR imaging of female pelvic tumors: a pictorial review. Radiographics. 2009;29(3):759\u0026ndash;74. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1148/rg.293085130\u003c/span\u003e\u003cspan address=\"10.1148/rg.293085130\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. discussion 774-8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFunt SA, Hricak H. Ovarian malignancies Top Magn Reson Imaging. 2003;14(4):329\u0026ndash;37. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1097/00002142-200308000-00005\u003c/span\u003e\u003cspan address=\"10.1097/00002142-200308000-00005\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFujii S, Matsusue E, Kigawa J. Diagnostic accuracy of the apparent diffusion coefficient in differentiating benign from malignant uterine endometrial cavity lesions: initial results. Eur Radiol. 2008;18(2):384\u0026ndash;9. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s00330-007-0769-9\u003c/span\u003e\u003cspan address=\"10.1007/s00330-007-0769-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTamai K, Koyama T, Saga T. Diffusion-weighted MR imaging of uterine endometrial cancer. J Magn Reson Imaging. 2007;26(3):682\u0026ndash;7. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/jmri.20997\u003c/span\u003e\u003cspan address=\"10.1002/jmri.20997\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRechichi G, Galimberti S, Signorelli M. Endometrial cancer: correlation of apparent diffusion coefficient with tumor grade, depth of myometrial invasion, and presence of lymph node metastases. AJR Am J Roentgenol. 2011;197(1):256\u0026ndash;62. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.2214/AJR.10.5584\u003c/span\u003e\u003cspan address=\"10.2214/AJR.10.5584\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePetrila O, Nistor I, Romedea NS, Negru D, Scripcariu V. Can the ADC Value Be Used as an Imaging Biopsy in Endometrial Cancer? Diagnostics (Basel). 2024;14(3):325. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/diagnostics14030325\u003c/span\u003e\u003cspan address=\"10.3390/diagnostics14030325\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePeerlings J, Woodruff HC, Winfield JM. Stability of radiomics features in apparent diffusion coefficient maps from a multi-centre test-retest trial. Sci Rep. 2019;9(1):4800. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41598-019-41344-5\u003c/span\u003e\u003cspan address=\"10.1038/s41598-019-41344-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSurov A, Meyer HJ, Wienke A. Correlation between apparent diffusion coefficient (ADC) and cellularity is different in several tumors: a meta-analysis. Oncotarget. 2017;8(35):59492\u0026ndash;9. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.18632/oncotarget.17752\u003c/span\u003e\u003cspan address=\"10.18632/oncotarget.17752\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSun H, Zeng X, Xu T, Peng G, Ma Y. Computer-Aided Diagnosis in Histopathological Images of the Endometrium Using a Convolutional Neural Network and Attention Mechanisms. IEEE J Biomed Health Inf. 2020;24(6):1664\u0026ndash;76. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/JBHI.2019.2944977\u003c/span\u003e\u003cspan address=\"10.1109/JBHI.2019.2944977\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Endometrial carcinoma, Atypical endometrial hyperplasia, Magnetic resonance imaging, Radiomics, Deep learning","lastPublishedDoi":"10.21203/rs.3.rs-9251059/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9251059/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eDifferentiating endometrial carcinoma (EC) from atypical endometrial hyperplasia (AEH) before surgery is critical for guiding treatment decisions, yet current Magnetic resonance imaging (MRI) assessment is limited by overlapping appearances and inter-observer variability.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eIn this retrospective, two-center study, 297 patients (153 stage I EC and 144 AEH) who underwent multiparametric MRI were included. Radiomics features were extracted from T2-weight ed imaging with fat saturation (T2WI-FS), diffusion-weighted imaging (DWI), apparent diffusion coefficient (ADC) map, and contrast-enhanced T1-weighted imaging (CE-T1WI) sequences. Feature selections were performed using univariate filtering, correlation analysis, least absolute shrinkage and selection operator (LASSO) and multivariate stepwise regression. A Radiomics model, a Clin\u0026thinsp;+\u0026thinsp;Rad model, and a DenseNet121-2D model were constructed and compared for differentiating EC from AEH. Model performance was evaluated using receiver operating characteristic (ROC) curves, calibration curves, decision curve analysis (DCA).\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eSeven radiomics features and eight deep-learning features were ultimately selected to construct the models. In external validation cohort, the DenseNet121-2D model achieved the highest diagnostic performance (area under the curve [AUC 0.854]), outperforming both the Radiomics model (AUC 0.803) and the Clin\u0026thinsp;+\u0026thinsp;Rad model (AUC 0.773). The DenseNet121-2D model also demonstrated better calibration and greater net clinical benefit across threshold probabilities. SHAP analysis confirmed that many discriminative features originated from ADC, consistent with its biological role in reflecting tumor cellularity and diffusion restriction.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eThe DenseNet121-2D model demonstrated superior generalizability and diagnostic accuracy compared with traditional radiomics approaches, providing a promising non-invasive tool to support individualized decision-making.\u003c/p\u003e","manuscriptTitle":"MRI-based deep learning model for differentiating endometrial cancer from atypical endometrial hyperplasia: A Multicenter Study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-29 06:14:16","doi":"10.21203/rs.3.rs-9251059/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"213685809538932841905710212693944197550","date":"2026-05-16T00:53:08+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-04-20T13:08:40+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-04-09T15:03:38+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-02T10:51:06+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-02T10:50:19+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Imaging","date":"2026-03-28T08:31:54+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"99b16dc4-bc0c-479c-8c43-c6f267df08c9","owner":[],"postedDate":"April 29th, 2026","published":true,"recentEditorialEvents":[{"type":"reviewerAgreed","content":"213685809538932841905710212693944197550","date":"2026-05-16T00:53:08+00:00","index":84,"fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-04-29T06:14:16+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-29 06:14:16","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9251059","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9251059","identity":"rs-9251059","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}