Combining transvaginal ultrasound radiomics with clinical-ultrasound semantic features for the identification of high-risk endometrial lesions.

As summarized in Table  3 , statistically significant differences ( P  < 0.05) were observed between the benign and high-risk groups regarding age, menopausal status, endometrial thickness, endometrial echogenicity, endometrial midline, endometrial–myometrial junction appearance, the “bright edge” sign, intracavitary fluid, Doppler color flow score, and vascular patterns. Table 3 Clinical data and ultrasound semantic features Analysis Parameters Benign ( n  = 177) High-Risk( n  = 62) Statistics P -value Menopausal status \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{\chi\:}}^{2}$$\end{document} =7.35 (Pearson) 0.007  postmenopausal 33 (18.6%) 22 (35.5%)  premenopausal 144 (81.4%) 40 (64.5%) Age (years) 43.19 ± 10.86 48.35 ± 9.72 t=-3.31 0.001 Endometrial thickness (mm) 10 (6.4, 14.3) 14.8 (9, 20.7) Z=-4.32 < 0.001 Echogenicity 26.96(Fisher) < 0.001  Uniform 28 (15.8%) 0  Homogeneous background with regular cystic areas 5 (2.8%) 0  Homogeneous background with irregular cystic areas 3 (1.7%) 1 (1.6%)  Heterogeneous background without cystic areas 114 (64.4%) 48 (77.4%)  Heterogeneous background with regular cystic areas 25 (14.1%) 6 (9.7%)  Heterogeneous background with irregular cystic areas 2 (1.1%) 7 (11.3%) Endometrial midline appearance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{\chi\:}}^{2}$$\end{document} =31.61 (Pearson) < 0.001  Linear 41 (23.2%) 0  Non-linear 26 (14.7%) 1 (1.6%)  Irregular 9 (5.1%) 2 (3.2%)  Not defined 101 (57.1%) 59 (95.2%) Endometrial–myometrial junction \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{\chi\:}}^{2}$$\end{document} =99.97 (Pearson) < 0.001  Regular 147 (83.1%) 12 (19.4%)  Irregular 25 (14.1%) 22 (35.5%)  Interrupted 5 (2.8%) 18 (29%)  Not defined 0 10 (16.1%) Bright edge \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{\chi\:}}^{2}$$\end{document} =19.61 (Pearson) < 0.001  Absent 122 (68.9%) 60 (96.8%)  Present 55 (31.1%) 2 (3.2%) Intracavitary fluid 12.62 (Fisher) 0.002  Absence of echogenicity 168(94.9%) 55(88.7%)  Anechoic or low-level echogenicity 8(4.5%) 1(1.6%)  Ground-glass appearance 1(0.6%) 2(3.2%)  Mixed echogenicity 0 4(6.5%) Doppler color score 28.53 (Fisher) < 0.001  1 (no color) 140(79.1%) 31(50%)  2 (minimal color) 32(18.1%) 16(25.8%)  3 (moderate color) 3(1.7%) 11(17.7%)  4 (abundant color) 2(1.1%) 4(6.5%) Vascular pattern 43.09 (Fisher) < 0.001  Absence of blood flow 140(79.1%) 31(50%)  Single dominant vessel without branching 11(6.2%) 0  Single dominant vessel with branching 2(1.1%) 3(4.8%)  Scattered vessels 19(10.7%) 15(24.2%)  Circular flow 3(1.7%) 0  Multiple vessels with focal origin 1(0.6%) 3(4.8%)  Multiple vessels with multifocal origin 1(0.6%) 10(16.1%) Clinical data and ultrasound semantic features Performance metrics were aggregated across five independent 5-fold cross-validation repetitions using five distinct random seeds (seed∈{42, 100, 200, 500, 1000}). As presented in Table  4 , data are reported as the mean±standard deviation across the five seeds. The standard deviations for all diagnostic models across all feature subsets did not exceed 0.007. The performance metrics for the SVM, Logistic Regression, XGBoost, LightGBM, and AdaBoost classifiers remained identical across all five random seeds. Table 4 Comparison of diagnostic performance among machine learning models Model Radiomics Clinical-Semantic Combined Accuracy AUC Accuracy AUC Accuracy AUC SVM 0.7617 ± 0.0000 0.7374 ± 0.0000 0.8201 ± 0.0000 0.8100 ± 0.0000 0.8284 ± 0.0000 0.8623 ± 0.0000 LogisticRegression 0.7549 ± 0.0000 0.7179 ± 0.0000 0.8579 ± 0.0000 0.8923 ± 0.0000 0.8599 ± 0.0000 0.8949 ± 0.0000 RandomForest 0.7436 ± 0.0051 0.7233 ± 0.0040 0.8200 ± 0.0052 0.8652 ± 0.0043 0.8508 ± 0.0048 0.8914 ± 0.0043 XGBoost 0.7551 ± 0.0000 0.7413 ± 0.0000 0.8285 ± 0.0000 0.8671 ± 0.0000 0.8389 ± 0.0000 0.8814 ± 0.0000 LightGBM 0.7594 ± 0.0000 0.7314 ± 0.0000 0.8158 ± 0.0000 0.8530 ± 0.0000 0.8387 ± 0.0000 0.8707 ± 0.0000 GradientBoosting 0.7494 ± 0.0058 0.7367 ± 0.0033 0.8200 ± 0.0042 0.8581 ± 0.0006 0.8362 ± 0.0035 0.8688 ± 0.0004 AdaBoost 0.7446 ± 0.0000 0.7122 ± 0.0000 0.8370 ± 0.0000 0.8965 ± 0.0000 0.8486 ± 0.0000 0.8923 ± 0.0000 CatBoost 0.7641 ± 0.0067 0.7457 ± 0.0049 0.8310 ± 0.0038 0.8863 ± 0.0027 0.8618 ± 0.0041 0.9077 ± 0.0059 Voting_Soft 0.7686 ± 0.0038 0.7589 ± 0.0024 0.8259 ± 0.0023 0.8880 ± 0.0007 0.8441 ± 0.0025 0.9014 ± 0.0006 The mean values represent the average results calculated across 5 independent random seed runs, and the standard deviation (SD) represents the deviation across these five seeds. AUC Area under the receiver operating characteristic curve Comparison of diagnostic performance among machine learning models The mean values represent the average results calculated across 5 independent random seed runs, and the standard deviation (SD) represents the deviation across these five seeds. AUC Area under the receiver operating characteristic curve Among the initially evaluated metrics, 700 + features were processed through the iterative LASSO pipeline (Supplementary Fig. 1). The framework dynamically retained a core subset of 150–208 features per iteration (averaging 167 features, reflecting an approximate 76% redundancy reduction), with the optimal regularization parameter α clustering within the 10 magnitude. Cross-fold evaluation confirmed that a core subset of 30 features demonstrated complete selection consistency across all 5 folds (Supplementary Fig. 2). This stable tier predominantly comprised multi-scale wavelet parameters (50%, 15/30) and first-order intensity statistics (27%, 8/30). As summarized in Table  4 , the CatBoost classifier implemented within the Combined framework achieved the highest overall diagnostic efficacy among the nine modeling pipelines, yielding a mean AUC of 0.9077 (± 0.0059) and a mean accuracy of 0.8618 (± 0.0041). Individual classifiers restricted to the standalone Radiomics framework demonstrated lower diagnostic performance, with mean AUC values ranging from 0.7122 to 0.7589. Loss convergence curves across the cross-validation folds were analyzed for the CatBoost combined model (Fig.  2 ). The training loss smoothly decreased toward 0 across all folds, indicating adequate model convergence. For the validation set, fold 1 demonstrated the optimal convergence profile, minimizing at a log-loss of 0.2569 at iteration 192 via the early-stopping mechanism before stabilizing. The validation loss curves for the remaining folds, particularly folds 3 and 4, exhibited varying degrees of rebound after reaching their respective minima, indicating a minor over-fitting tendency due to cross-validation partition variability. Fig. 2 Loss convergence curves for the CatBoost combined model under cross-validation. The left displays the training log-loss across all folds, and the right illustrates the corresponding validation log-loss comparison Loss convergence curves for the CatBoost combined model under cross-validation. The left displays the training log-loss across all folds, and the right illustrates the corresponding validation log-loss comparison The DeLong test was performed across all machine learning pipelines to compare diagnostic efficacy among the three frameworks. Pairwise comparisons revealed that differences in AUC values between the Combined model and the standalone Radiomics model were statistically meaningful across all nine classification algorithms (all P  < 0.001). Similarly, the standalone Clinical-Semantic model demonstrated higher diagnostic efficacy than the Radiomics model across all pipelines (all P  < 0.001). When comparing the Combined model against the standalone Clinical-Semantic model, differences in AUC values depended on the specific classification framework. For the Gradient Boosting ( P  = 0.1336) and AdaBoost ( P  = 0.6289) algorithms, no statistical differences in AUC values were observed between the combined and standalone clinical-semantic frameworks. Conversely, for the remaining seven classification algorithms, differences in diagnostic efficacy between the two frameworks were statistically meaningful (all P  < 0.05). To further evaluate the reliability of the predicted probability outputs from the combined framework, calibration curves and Brier scores were systematically analyzed across all nine machine learning models (Fig.  3 ). Among the evaluated algorithms, the CatBoost combined model achieved the lowest Brier score (0.1016). Fig. 3 Calibration curves of the nine machine learning models within the combined framework. The dashed line indicates perfect calibration, and labels denote the respective Brier scores Calibration curves of the nine machine learning models within the combined framework. The dashed line indicates perfect calibration, and labels denote the respective Brier scores The calibration curve of the CatBoost model aligned closely with the perfect calibration diagonal, demonstrating optimal performance particularly in the high-probability range (predicted probability > 0.6). A minor underestimation bias was observed in the low-to-medium probability range (0.2–0.4). Furthermore, the predicted probability distribution histogram (Fig.  4 ) demonstrated that the predicted probabilities for histopathologically true-positive cases were highly concentrated within the 0.8–1.0 interval, whereas true-negative cases were primarily distributed within the 0.0–0.1 interval, confirming robust probability separation and risk stratification efficacy. Fig. 4 Calibration and probability separation of the CatBoost combined model. The left shows the OOF calibration curve against the perfect reference line, and the right displays the sample distribution demonstrating robust discrimination between true-negative and true-positive cases Calibration and probability separation of the CatBoost combined model. The left shows the OOF calibration curve against the perfect reference line, and the right displays the sample distribution demonstrating robust discrimination between true-negative and true-positive cases Relative feature importance analysis within the top-performing CatBoost combined model demonstrated that Endometrial-Myometrial boundary parametric attributes, Endometrial Thickness, and Age were the leading clinical-semantic predictors, with the Rad-score serving as a foundational metric contributing substantially to the joint framework (Fig.  5 ). Fig. 5 Feature importance ranking (± SD) for the CatBoost combined framework across cross-validation partitions Feature importance ranking (± SD) for the CatBoost combined framework across cross-validation partitions Decision curve analysis (DCA) was utilized to evaluate the clinical utility of the combined machine learning models (Fig.  6 ). At a threshold probability of 0.2, the Voting_Soft combined algorithm achieved the highest net benefit (NB = 0.6891), slightly exceeding the “Treat All” baseline (NB = 0.6882), whereas the CatBoost model (NB = 0.6860) and the remaining algorithms yielded NBs below the baseline. Fig. 6 Decision curve analysis (DCA) of the combined machine learning models. Net clinical benefit is plotted against threshold probabilities. The legend indicates the specific net benefit (NB) quantified at the clinical risk threshold of 0.20 for each modeling strategy Decision curve analysis (DCA) of the combined machine learning models. Net clinical benefit is plotted against threshold probabilities. The legend indicates the specific net benefit (NB) quantified at the clinical risk threshold of 0.20 for each modeling strategy In the low-threshold range (0.0–0.2), the NBs of all combined models approximated or remained comparable to the “Treat All” baseline. Within the medium-threshold range (0.3–0.6), curve stratification emerged; most models, led by CatBoost, exceeded the “Treat All” baseline, while the AdaBoost curve fell below it in several segments. In the high-threshold range (0.6–0.8), all curves exhibited a downward trend, where the CatBoost combined model exhibited a relatively gradual decline, remaining superior to alternative combined models and baseline strategies.

A total of 239 patients met the selection criteria, including 177 (74.06%) with benign lesions and 62 (25.94%) with high-risk lesions, yielding 956 standardized transvaginal ultrasound (TVUS) images. The specific histopathological diagnoses within the cohort are listed in Table  1 . This two-center retrospective study utilized clinical and imaging data collected between January 2021 and October 2025 from The Third People’s Hospital of Longgang District (lead center) and Shenzhen Longgang District Central Hospital (participating center). The study was conducted in accordance with institutional ethical standards, with the requirement for informed consent waived. Table 1 Pathological distribution of endometrial lesions ( N  = 239) Benign ( n =177) High-Risk ( n =62) Histopathological Type No. of cases Percentage (%) Histopathological Type No. of cases Percentage (%) Endometrial polyps 85 48.0 Endometrioid adenocarcinoma 40 64.5 Endometrial simple hyperplasia 35 19.8 Atypical endometrial hyperplasia 19 30.7 Endometrial polypoid hyperplasia 22 12.4 Endometrial carcinosarcoma 1 1.6 Endometritis 16 9.0 Endometrial squamous cell carcinoma 1 1.6 Submucous myomas 14 7.9 Undifferentiated endometrial carcinoma 1 1.6 Endometrial complex hyperplasia 3 1.7 Adenofibroma 1 0.6 Adenomyomatous polyp 1 0.6 Pathological distribution of endometrial lesions ( N  = 239) Patients were evaluated for inclusion based on the following criteria: (1) availability of TVUS images clear enough for analysis; (2) histopathologically confirmed diagnosis via biopsy or surgical resection; (3) an interval between TVUS examination and pathological sampling of ≤ 30 days; and (4) complete clinical and ultrasound semantic records. Exclusion criteria consisted of: age < 18 years, pregnancy or lactation, a history of uterine surgery or hormone therapy, and severe adenomyosis or the presence of an intrauterine device causing prominent acoustic shadowing. TVUS examinations were performed by sonographers with ≥ 5 years of experience using LOGIQ E9, Voluson E10, or Resona 8 systems (5.0–10.0 MHz). Standardized views included the maximum longitudinal and transverse planes of the endometrium and color Doppler imaging. Static images were exported directly from the workstation and cropped by two specialists (with ≥ 10 years of experience) to remove peripheral artifacts and patient metadata while preserving the uterine contour. Based on the World Health Organization (WHO) classification of female genital tumors and established clinical management strategies [ 23 , 24 ], lesions were categorized into two cohorts: (1) Benign Group, and (2) High-Risk Group (comprising EC and AEH). To align with clinical management paths, AEH was incorporated into the high-risk group because approximately 40% of AEH cases coexist with or progress to EC [ 25 ], necessitating similar surgical or therapeutic interventions [ 3 ]. The technical workflow of the study is illustrated in Fig. 1 . Fig. 1 Flowchart of the study workflow Flowchart of the study workflow Clinical and ultrasonographic features were recorded by two investigators (each with 10 years of experience in gynecological ultrasound) blinded to the pathological outcomes. Disagreements were resolved by consensus with a senior specialist (with > 15 years of experience). The ultrasound semantic features were defined according to the International Endometrial Tumor Analysis (IETA) consensus and adapted from established parameters from previous studies [ 9 , 26 , 27 ]. These recorded parameters included: ① Age. ② Menopausal status (premenopausal or postmenopausal). ③ Endometrial thickness (ET): Double-layer thickness measured at the thickest part on the longitudinal plane. ④Endometrial echogenicity: Classified as uniform, homogeneous background with regular/irregular cystic areas, or heterogeneous background without/with regular or irregular cystic areas. ⑤ Endometrial midline (linear, non-linear, irregular, or not defined). ⑥ Endometrial-myometrial junction (EMJ): Assessed for integrity as regular, irregular, interrupted, or not defined. ⑦ “Bright edge” sign (presence or absence). ⑧ Intracavitary fluid (absent, anechoic, ground-glass, or mixed). ⑨ Color flow score: Score 1 (no flow), Score 2 (minimal flow), Score 3 (moderate flow), or Score 4 (abundant flow). ⑩ Vascular pattern: Classified into seven patterns: no vessels, single dominant vessel (with/without branching), scattered vessels, circular vessels, and multiple dominant vessels (focal or multifocal origin). ① Age. ② Menopausal status (premenopausal or postmenopausal). ③ Endometrial thickness (ET): Double-layer thickness measured at the thickest part on the longitudinal plane. ④Endometrial echogenicity: Classified as uniform, homogeneous background with regular/irregular cystic areas, or heterogeneous background without/with regular or irregular cystic areas. ⑤ Endometrial midline (linear, non-linear, irregular, or not defined). ⑥ Endometrial-myometrial junction (EMJ): Assessed for integrity as regular, irregular, interrupted, or not defined. ⑦ “Bright edge” sign (presence or absence). ⑧ Intracavitary fluid (absent, anechoic, ground-glass, or mixed). ⑨ Color flow score: Score 1 (no flow), Score 2 (minimal flow), Score 3 (moderate flow), or Score 4 (abundant flow). ⑩ Vascular pattern: Classified into seven patterns: no vessels, single dominant vessel (with/without branching), scattered vessels, circular vessels, and multiple dominant vessels (focal or multifocal origin). Manual delineation of the regions of interest (ROIs) was performed on the cropped TVUS images using the Darwin Research Platform( http://premium.darwin.yizhun-ai.com ) [ 28 ]. Segmentations were conducted independently by two specialists (each with 10 years of experience in gynecological ultrasound) blinded to the pathological reports. Any boundary discrepancies were reviewed and adjusted by a senior specialist (with > 15 years of experience) to ensure structural consistency. Clinical continuous variables (age and endometrial thickness) were normalized via Z-score transformation, while categorical ultrasound semantic features were converted using one-hot encoding. To minimize gray-scale intensity variations across different ultrasound systems, voxel intensity normalization to a standard [0,1] range was applied to the segmented ROI regions prior to feature extraction. A total of 1,125 radiomics features were extracted from each ROI using the Darwin Research Platform. These features comprised 216 first-order statistics, 9 two-dimensional shape features, and 900 texture matrix features, which specifically included 288 gray level co-occurrence matrix (GLCM), 192 gray level run length matrix (GLRLM), 192 gray level size zone matrix (GLSZM), 60 neighboring gray tone difference matrix (NGTDM), and 168 gray level dependence matrix (GLDM) features. Subsequent statistical modeling and feature analysis were performed using Python (v3.10.4) with the scikit-learn library (v1.2.2). To evaluate the stability of the diagnostic framework and prevent data leakage, dataset partitioning was conducted under a patient-level 5-fold cross-validation scheme. The specific distribution of benign and high-risk cases within the training and testing sets across all five folds is presented in Table  2 . This 5-fold cross-validation process was repeated independently five times using five distinct random seeds (seed∈{42, 100, 200, 500, 1000}).Within the training loop of each fold, low-variance and non-significant features were pre-filtered to narrow the pool to approximately 700 + variables. Subsequently, features were independently standardized, and Least Absolute Shrinkage and Selection Operator (LASSO) regression with L1 regularization was nested within each iteration to execute localized feature selection. The optimal regularization parameter (α) was determined via internal cross-validation. The Radiomics score (Rad-score) for each patient was calculated via a linear combination of the selected features weighted by their respective coefficients using the formula: Rad-score = Σ(β i × X i ), where β i represents the regression coefficient and X i represents the standardized feature value. Table 2 Sample distribution across the 5-fold cross-validation sets n Fold Dataset Group Total High-Risk Benign Fold 1 Training 50 141 191 Test 12 36 48 Fold 2 Training 50 141 191 Test 12 36 48 Fold 3 Training 49 142 191 Test 13 35 48 Fold 4 Training 49 142 191 Test 13 35 48 Fold 5 Training 50 142 192 Test 12 35 47 Sample distribution across the 5-fold cross-validation sets Three diagnostic strategies were evaluated: (1) a Radiomics model, (2) a Clinical-Semantic model, and (3) a Combined model integrating both feature sets. For each strategy, nine machine learning pipelines—comprising eight base classifiers (SVM, Logistic Regression, Random Forest, XGBoost, LightGBM, Gradient Boosting, AdaBoost, and CatBoost) and a Voting ensemble—were constructed and tested. Statistical analyses and model evaluations were performed using Python (v3.10.4) with the scikit-learn library (v1.2.2) and the Darwin Research Platform. Categorical variables were compared using the Chi-square test or Fisher’s exact test. Continuous variables were evaluated for normality using the Kolmogorov–Smirnov test; normally distributed data were analyzed using the independent-samples t -test, while non-normally distributed data were compared using the Mann–Whitney U test. Model diagnostic performance was evaluated using accuracy and the Area Under the Curve (AUC) obtained via a patient-level 5-fold cross-validation scheme. Pairwise comparisons of AUC values among the Radiomics, Clinical-Semantic, and Combined modeling frameworks were performed using the DeLong test. Decision Curve Analysis (DCA) was applied to evaluate clinical net benefit. Statistical significance was defined as a two-tailed P  < 0.05.

This study evaluated standalone and combined transvaginal ultrasound-based models for the risk stratification of endometrial lesions. The optimal validation Area Under the Curve (AUC) values for the radiomics, clinical-semantic, and combined models were 0.7589, 0.8965, and 0.9077 (CatBoost), respectively. While specific combined configurations showed statistically significant differences compared to standalone clinical-semantic metrics, this incremental gain was dependent on the choice of classifier algorithm. In decision curve analysis, the Voting_Soft configuration of the combined model was associated with favorable net benefits at lower thresholds (0.15–0.20), while the CatBoost combined framework maintained relatively stable net benefits at higher thresholds (0.30–0.60). In summary, this integrated framework serves as an exploratory tool that may provide complementary information to assist in clinical patient evaluation.

Endometrial cancer (EC) incidence is rising globally, and while early-stage prognosis is generally favorable, mortality remains high for advanced and high-risk histological types [ 29 , 30 ]. As a recognized precursor, atypical endometrial hyperplasia also carries considerable clinical risk, with a high probability of progressing to malignancy if not identified early [ 3 ]. Although transvaginal ultrasound remains the frontline screening tool [ 6 ], its efficacy is limited by inter-observer variability and the subjective nature of morphological assessment [ 7 , 8 ]. Accordingly, this study assessed the diagnostic efficacy of ultrasound-based radiomics and clinical-ultrasound semantic features via combined machine learning modeling. Our results suggested that the combined models exhibited superior performance compared to standalone unimodal approaches, highlighting that the integration of non-invasive clinical-ultrasound features may offer important complementary value in optimizing diagnostic precision. In the clinical and sonographic assessment, early identification of high-risk indicators remains foundational for refining diagnostic pathways. Our findings identified statistically meaningful differences in age, menopausal status, endometrial thickness, and specific morphological features between benign and malignant lesions, aligning with historical epidemiological data [ 27 , 31 – 34 ]. Pathophysiologically, advanced age and postmenopausal status correlate with prolonged, unopposed estrogen exposure or localized hormonal imbalances, which drive atypical endometrial hyperplasia and subsequent oncogenesis [ 31 ]. Sonographically, the observed loss of the endometrial midline and an interrupted or ill-defined endometrial-myometrial junction serve as key indicators of invasive borders and potential myometrial infiltration [ 32 , 33 ]. Furthermore, in terms of vascularity, the hypervascularity (color scores 3–4) and disordered multi-dominant vascular patterns prominent in the high-risk cohort may reflect aberrant pathological angiogenesis, a phenomenon characterized by irregular vessel calibers and chaotic endothelial cell proliferation that is frequently observed in malignant endometrial alterations [ 27 , 34 , 35 ]. In the present study, the standalone radiomics signature exhibited moderate efficacy in differentiating benign lesions from the high-risk cohort. Rather than macroscopic geometric configurations, the top-performing radiomic features were predominantly multi-scale wavelet and first-order statistical patterns. From a clinical perspective, high-risk and malignant endometrial lesions may harbor varying degrees of intratumoral heterogeneity across spatial, cellular, and microenvironmental dimensions [ 10 , 35 ]. Microscopic heterogeneities—such as asymmetric cell proliferation densities, disordered microvascular perfusion, and focal areas of stromal infiltration—may manifest as sub-visual pixel frequency fluctuations on digital ultrasound images [ 36 ]. Consequently, while macroscopic shape features offer limited diagnostic power in early-stage or localized disease due to geometric overlap with certain benign lesions, computer-assisted sub-visual texture analysis may provide non-invasive surrogates to quantify underlying tissue complexity. In our evaluation of combined modeling, the combined approach exhibited enhanced diagnostic capacity compared to standalone unimodal frameworks, with the CatBoost combined model demonstrating favorable classification performance. This performance improvement potentially underscores the non-linear synergy achieved by merging macroscopic clinical-ultrasound parameters with sub-visual radiomic metrics. From an algorithmic perspective, tree-based ensemble frameworks, such as CatBoost, are well-suited for capturing high-order interactions and multi-dimensional cross-talk among heterogeneous variables [ 37 ]. Clinically, while traditional indicators like endometrial thickness may demonstrate limited diagnostic specificity when evaluated in isolation [ 1 ], their diagnostic relevance appears to be enhanced when contextualized within a multi-dimensional data environment that accounts for micro-scale texture variations. Consequently, the integration of macro-morphological attributes and micro-radiomics profiles may allow the combined model to identify discrete pathological patterns that single-modality evaluations fail to resolve, thereby revealing latent complementary value within routine clinical metrics. The clinical pragmatic value of the combined models was evaluated via Decision Curve Analysis (DCA). Within the low-risk threshold decision interval (0–0.20), the net benefits of the ensemble algorithms closely aligned with the “treat-all” baseline, suggesting that universal intervention strategies may remain instructional in preliminary screening scenarios where non-detection must be strictly avoided. Notably, the clinical advantage of the combined models became apparent within the 0.30–0.60 threshold range, where frameworks like CatBoost provided higher net benefits than conventional benchmarks, potentially assisting clinicians in minimizing unnecessary invasive evaluations without compromising diagnostic safety. Furthermore, distinct performance divergence across mathematical architectures—as illustrated by the noticeable attenuation of AdaBoost beyond a 0.50 threshold—may support a stratified clinical decision pathway. Specifically, for initial outpatient triage (thresholds 0.15–0.20), referencing the Voting_Soft configuration may optimize early net benefit, whereas for further clinical evaluation and management decisions (thresholds 0.30–0.60), the more resilient CatBoost framework appears preferable, thereby providing a non-invasive, objective decision-support tool to individualize the management of endometrial lesions. Although this study provides a preliminary reference for the non-invasive evaluation of high-risk endometrial lesions, several limitations remain. First, this investigation relied on a retrospective design with a relatively modest cohort and lacked an independent external validation set. Concurrently, endometrial thickness fluctuates periodically in premenopausal individuals; although the IETA consensus recommends executing ultrasonographic evaluations during the early proliferative phase (days 4–6 of the menstrual cycle) [ 9 ], our retrospective data precluded uniform enforcement of this specific scanning window, potentially introducing chronological bias. Furthermore, consolidating endometrial atypical hyperplasia and endometrial cancer into a single high-risk cohort may obscure nuanced microscopic heterogeneities unique to distinct pathological entities. Second, although all manual regions of interest were co-reviewed and verified by senior sonologists, inter-observer and intra-observer reproducibility coefficients were not formally quantified, leaving a methodological gap in evaluating the stability of the extracted radiomic features. Additionally, confidence intervals and formal statistical comparisons were not computed for the decision curve analysis, which limits the statistical evaluation of variations in net clinical benefit. Furthermore, the diagnostic performance of the models was not compared with that of sonologists of different seniority levels. Recent studies demonstrate that validating AI frameworks within standardized reporting systems can enhance the diagnostic accuracy of general clinicians [ 38 , 39 ]. Future research leveraging prospective, multi-center recruitment should focus on the integration of multimodal ultrasound, shear-wave elastography, and molecular genetic profiles to build a more resilient, multi-dimensional decision-support system to individualize management paradigms.

Endometrial cancer (EC) represents the sixth most prevalent malignancy among women globally, with climbing incidence and mortality rates observed in younger demographics [ 1 , 2 ]. Atypical endometrial hyperplasia (AEH), a critical precancerous precursor that shares a high risk of coexisting or progressing to EC, similarly imposes a non-negligible healthcare burden [ 3 ]. Early and accurate stratification of these high-risk endometrial lesions is vital; while the 5-year survival rate for Stage I EC exceeds 95%, it drops below 20% in advanced or recurrent stages [ 4 ]. Transvaginal ultrasound (TVUS) serves as the primary non-invasive screening tool for endometrial lesions [ 5 , 6 ], although its diagnostic performance is limited by operator dependency and morphological overlaps between benign and high-risk conditions [ 7 , 8 ]. Standardized frameworks, such as the International Endometrial Tumor Analysis (IETA) consensus, aim to improve diagnostic objectivity [ 9 ]; however, inter-observer consistency remains variable [ 7 ], and individual ultrasound parameters often lack sufficient predictive value to reliably identify high-risk cases. Radiomics enables the high-throughput extraction of quantitative features to evaluate tissue heterogeneity less accessible to routine visual assessment [ 10 , 11 ]. This approach has demonstrated clinical utility in the characterization, staging, and prognosis prediction of various solid tumors, including breast, thyroid, lung, and liver cancers [ 12 – 14 ]. In endometrial lesions, prior radiomics investigations have focused primarily on magnetic resonance imaging (MRI) to evaluate biological aggressiveness [ 15 – 18 ] and molecular subtypes [ 19 ]. Regarding ultrasound modalities, recent studies have demonstrated the feasibility of artificial intelligence in endometrial applications, including the automated segmentation and radiomics-based classification of benign versus atypical lesions, the pre-operative evaluation of deep myometrial invasion, and multi-center deep learning categorization [ 20 – 22 ]. However, despite the role of TVUS as the preferred frontline screening modality, the incremental value of combining ultrasound image features with subjective semantic descriptors remains less established compared with magnetic resonance imaging (MRI) techniques. To address this gap, this study evaluated an integrated approach that couples clinical data, standardized ultrasound semantic features, and ultrasound radiomics features. The main contributions of this work include: (1)evaluating the incremental value of combining sub-visual radiomics with routine clinical-ultrasound descriptors, and (2)comparing independent risk assessment models with combined evaluation strategies to explore their potential to assist in the preoperative risk stratification of endometrial lesions.

Supplementary Material 1. Supplementary Material 1.