An interpretable multimodal machine learning model integrating transvaginal ultrasound and clinical features for differentiating benign from malignant endometrial thickening in postmenopausal women
Aug 2026· Frontiers in Oncology· Vol 16· 0 citations· 41 references
Medicine
TL;DR
The combined image–clinical logistic regression model demonstrated promising discriminatory and calibration performance for risk stratification of postmenopausal endometrial thickening, achieving improved specificity over conventional thickness-based criteria.
Abstract
Objective To develop and validate an interpretable multimodal machine learning model that integrates transvaginal ultrasound (TVUS) deep learning features with clinical risk factors for differentiating benign from malignant endometrial thickening in postmenopausal women, with the aim of reducing unnecessary invasive procedures. Methods In this retrospective single-centre diagnostic study, 601 postmenopausal women with histopathologically confirmed endometrial thickening (509 benign, 92 malignant) were identified consecutively and allocated to a training set (n = 420) and a stratified held-out test set (n = 181) at a 7:3 ratio. Deep learning features were extracted from manually segmented mid-sagittal TVUS images via four pretrained ResNet architectures and compressed into a single imaging score (DL_score) through sequential mRMR filtering and LASSO regression. Independent clinical predictors were identified by multivariate logistic regression. Ten fusion classifiers were trained on the combined feature set and evaluated by AUC, calibration, decision curve analysis, and SHAP-based interpretability. Results In the test set, all image-based models outperformed the clinical-only model (AUC, 0.797; 95% CI, 0.707–0.888), with ResNet152 achieving the highest single-modality AUC (0.850; 95% CI, 0.782–0.917). Among the combined models, logistic regression (LR) performed best, with an AUC of 0.906 (95% CI, 0.842–0.970), an accuracy of 84.6%, a sensitivity of 75.9%, and a specificity of 92.1%. The LR model was well calibrated (Brier score, 0.08; Hosmer–Lemeshow P = 0.60) and, on DCA, provided the greatest net benefit across the clinically relevant range of threshold probabilities. SHAP analysis identified the DL_score, postmenopausal bleeding, and BMI as the three most influential predictors; Grad-CAM activation maps indicated that the DL_score captured spatially localised information related to endometrial bulk and junction integrity. Conclusion The combined image–clinical logistic regression model demonstrated promising discriminatory and calibration performance for risk stratification of postmenopausal endometrial thickening, achieving improved specificity over conventional thickness-based criteria. This interpretable multimodal approach may serve as a useful adjunct to support clinical decision-making in identifying postmenopausal women at lower risk of malignancy, potentially reducing the burden of unnecessary invasive procedures.
OBJECTIVE
Preoperative risk stratification of high-risk endometrial lesions remains a clinical challenge. This study aimed to preliminarily explore an integrated machine learning approach combining transvaginal ultrasound (TVUS) radiomics, clinical indicators, and ultrasound semantic attributes to assist in clinical triage.
METHODS
TVUS images (n = 956) from 239 patients were retrospectively analyzed across two centers. Clinical and ultrasound semantic features were evaluated blindly. Within a patient-level 5-fold cross-validation scheme repeated across five independent random seeds, 1,125 radiomics features were extracted and selected via nested LASSO regression to compute a Rad-score. Frameworks using nine machine learning algorithms were evaluated via Area Under the Curve (AUC), DeLong tests, and Decision Curve Analysis (DCA).
RESULTS
The optimal validation AUCs for the Radiomics, Clinical-Semantic, and Combined models were 0.7589, 0.8965, and 0.9077 (CatBoost), respectively. The Combined framework yielded a statistically significant increase in AUC values compared with the Radiomics model (P < 0.001) and showed significant differences over the Clinical-Semantic model across seven algorithms (all P < 0.05). Endometrial-myometrial junction appearance was identified as the primary predictor. DCA indicated that the combined model's soft-voting configuration optimized initial triage at a 0.20 threshold, while its CatBoost configuration sustained higher net benefit across medium-to-high risk thresholds (0.30-0.60).
CONCLUSION
Integrating clinical-ultrasound semantic features with radiomics offers an objective, non-invasive approach to potentially assist in individualized risk stratification. This combined framework may provide a potential stratified decision-support pathway for further clinical evaluation.
Xiaohong Yao, Xuandan Ye, Lixia Chen et al.· BMC Women's Health· 0 citations
OBJECTIVE
To develop a multimodal machine learning model that integrates clinical data and ultrasound features to improve the non‑invasive preoperative differentiation between endometrial polyps (EMPs) and non‑polypoid endometrial lesions (including polypoid hyperplasia and normal endometrial tissue).
METHODS
This single‑center retrospective cohort study included 913 patients with a preoperative diagnosis of endometrial polyp confirmed by postoperative pathology. Based on pathological findings, patients were categorized into a typical polyp group (n = 601) and a non‑polyp group (n = 312, consisting mainly of polypoid hyperplasia). Clinical baseline characteristics, transvaginal ultrasound (TVUS) features, and hysteroscopic observations were collected. Independent predictors were identified through multivariate analysis. Multiple prediction models-including logistic regression, decision tree, random forest, and gradient boosting decision tree (GBDT)-were constructed and evaluated using a 70:30 split into training and test sets.
RESULTS
No significant differences were observed between the two groups in baseline characteristics. Typical EMPs more frequently presented as focal masses on ultrasound (82.4% vs. 75.3%) and exhibited detectable intralesional blood flow (51.6% vs. 38.8%). Multivariate analysis identified endometrial thickness on ultrasound, the presence of ultrasound‑detected blood flow, and hysteroscopic visualization of vegetations as independent predictors. Among the tested models, the GBDT demonstrated the best discriminative performance, with an area under the receiver operating characteristic curve (AUC) of 0.727, outperforming the traditional logistic regression model (AUC = 0.599).
CONCLUSION
Multimodal machine learning models, particularly GBDT, offer a potentially useful auxiliary tool for distinguishing EMPs from polypoid hyperplasia. However, their current diagnostic performance remains insufficient to support standalone clinical decision‑making. These findings highlight the limitations of existing non‑invasive methods and reinforce the necessity of histopathological confirmation. Future efforts should focus on incorporating more objective quantitative radiomic features to advance preoperative risk stratification.
Danli Ma, Zhiying Ye, Chuhan Wang et al.· BMC Women's Health· 0 citations
BACKGROUND
The significant sonographic overlap between metastatic and ultrasound-atypical reactive hyperplastic lymph nodes remains a challenge in subjective, experience-based ultrasound diagnosis. Consequently, clinicians rely on biopsy, which is an invasive procedure associated with complications such as bleeding and infection, as well as the risk of false-negatives that exacerbate patient anxiety.
MATERIALS AND METHODS
In this multicenter retrospective study, patients with suspicious lymph nodes from two institutions were enrolled. Conventional and habitat-based radiomic features, capturing intratumoral heterogeneity, were extracted from ultrasound images. After feature selection via ElasticNet regression and multicollinearity removal, four machine learning models were developed. Hyperparameters were optimized using fivefold cross-validation. Model performance was assessed using receiver operating characteristic curves, calibration plots, and decision curve analysis (DCA), and the model was interpreted using SHapley Additive exPlanation (SHAP) analysis.
RESULTS
A total of 1,230 patients (training set: 702; internal set: 302; external set: 226) were included in this study. Eight independent risk factors (e.g., age, long-to-short axis ratio, and cortical morphology) were identified. The random forest fusion model achieved an area under the curve (AUC) of 0.910 and an F1 score of 0.806 in the external set, significantly surpassing the clinical model (AUC = 0.690). Compared with conventional radiomics, the fusion model showed superior reclassification (net reclassification improvement = 0.562, integrated discrimination improvement = 0.144). SHAP analysis linked malignancy risk to higher gray-level nonuniformity and lower elongation, ensuring clinical plausibility. DCA confirmed robust clinical net benefit across all cohorts.
CONCLUSION
The fusion model, integrating radiomic and habitat features, enables noninvasive suspicious lymph node prediction. It may reduce unnecessary biopsies in low-risk patients and provide incremental value for individualized preoperative management by quantifying spatial characteristics.
Jiafei Shen, Zhiyan Jin, Xiaoxian Li et al.· Journal of Cancer Research a...· 0 citations
Objective This study aimed to develop an ultrasound-based deep learning radiomics nomogram for differentiating endometrial hyperplasia with polyps from endometrial cancer in endometrial lesions and to explore its clinical diagnostic efficacy. Methods We retrospectively collected clinical data from patients who underwent transvaginal ultrasound examinations in Suzhou Ninth People’s Hospital and Jiangsu Shengze Hospital from January 2020 to October 2024. Various machine learning models were developed using clinical data, handcrafted radiomics features, and deep learning features, with the highest AUC model selected as optimal. Univariate and stepwise multivariate analyses identified significant clinical features, which were combined with deep learning radiomics to create a Combined Model. The model’s predictive performance was evaluated using ROC, calibration, and decision curves, along with a deep learning radiomics nomogram. Results This retrospective study of 340 patients from two hospitals included 149 endometrial cancer cases. Among them, 305 patients were from Suzhou Ninth People’s Hospital and 35 were from Jiangsu Shengze Hospital. The 305 patients from Suzhou Ninth People’s Hospital were divided into a training cohort and an internal validation cohort in a 7:3 ratio, comprising 213 and 92 patients, respectively. The 35 patients from Shengze Hospital served as the external testing cohort. Multivariate analysis confirmed four independent predictors: Testosterone, Estradiol, Abnormal Bleeding, and Menopausal Status, used to build the final Clinical Model. The effectiveness of the models was assessed by comparing the area under the receiver operating characteristic curve (AUC) based on the internal validation cohort. The results showed that the Combined Model achieved an AUC of 0.991 (95% confidence interval [CI]: 0.980 - 1.000) on the training cohort, 0.94 (95% CI: 0.892 - 0.987) on the internal validation cohort, and 0.955 (95% CI: 0.882 - 1.000) on the external testing cohort. Calibration curves and DeLong test confirmed the Combined Model’s superior accuracy and performance over single-modality models, with DCA showing highest net benefit, and nomogram enabling individualized risk stratification. Conclusion The ultrasound-based combined model has high clinical diagnostic value for diagnosing endometrial cancer and simple endometrial hyperplasia with polyps.
Min Xu, W. Xu, Zhaoyang Lu et al.· Frontiers in Oncology· 0 citations
Objective To develop and validate a diagnostic model for malignant (MPE) and benign pleural effusion (BPE) using non-contrast chest CT deep learning (DL) and radiomics features, and to explore its incremental value alongside conventional biochemical biomarkers. Methods We retrospectively enrolled 208 patients (Jan 2020–Sep 2024) as internal training/testing cohorts (7:3 ratio) and 52 patients (Oct 2024–Dec 2025) for internal temporal validation. Radiomics and DL features were extracted from non-contrast CTs to construct a radiomics score (Radscore). Multivariable logistic regression evaluated the Radscore’s independent predictive value. Net reclassification improvement (NRI) and integrated discrimination improvement (IDI) assessed the incremental diagnostic value when combined with clinical biomarkers. Results Four radiomics features were ultimately retained to construct the Radscore. Multivariable adjusted analysis confirmed that the Radscore was an independent risk factor for MPE (OR: 2.718–2.776, P < 0.05). The model demonstrated good predictive performance (AUC = 0.805–0.864). Exploratory subgroup analyses indicated consistent discriminative trends, though estimates in certain strata were inherently limited by small event counts. Correlation analysis revealed only a weak correlation between pleural effusion carcinoembryonic antigen (pCEA) and the Radscore (r = 0.270, P = 0.001). Following the integration of pCEA, serum CEA (sCEA), pleural effusion adenosine deaminase (pADA), and serum CA125 (sCA125) into a combined clinical model, continuous NRI and IDI analyses demonstrated that the Radscore quantitatively refined individual risk estimation by shifting predicted probabilities in the appropriate direction for 75.0% of MPE patients and 66.7% of BPE patients. Conclusion The DL-radiomics-based Radscore is a promising quantitative biomarker for differentiating MPE from BPE. It functions independently of conventional biochemical metrics and provides meaningful incremental value, refining the accuracy of risk probability estimation in patients with pleural effusion.
Chun Cao, Jiang Liu, Qingqing Fang et al.· Frontiers in Medicine· 0 citations
RATIONALE AND OBJECTIVES
To develop and validate a machine learning model combining multiparametric Magnetic Resonance Imaging (MRI) radiomics and clinical indicators for predicting lymphovascular space invasion (LVSI) in endometrial cancer (EC).
MATERIALS AND METHODS
This retrospective study enrolled EC patients who underwent preoperative MRI at two centers. Of 567 initially screened patients, 408 were included per inclusion/exclusion criteria, divided into training and validation sets by hospital. Clinical risk factors and intratumoral/peritumoral radiomic features were identified. Six machine learning algorithms were used to build models; the one with the highest validation Area Under the Curve (AUC) was optimal. Five additional models were developed, and performance was evaluated via AUC, calibration curves, and decision curve analysis (DCA).
RESULTS
Logistic regression identified CA125 and tumor diameter as independent LVSI risk factors. Six machine learning models were built with CA125, tumor diameter, Rad_Score1 and Rad_Score2; the NeuralNetwork performed best (validation AUC=0.803). The combined clinical-intratumoral-peritumoral radiomics model achieved the highest AUC(training AUC = 0.863, validation AUC = 0.803), with good calibration (Hosmer-Lemeshow test, P > 0.05) and favorable net clinical benefit (threshold 0.1-0.7). SHapley Additive exPlanations (SHAP) analysis enhanced model interpretability.
CONCLUSION
This study systematically compared the predictive performance of six machine learning models for LVSI in EC, identifying the NeuralNetwork model as the most optimal. The combined clinical-intratumoral-peritumoral radiomics model, alongside its SHAP visualization tool, enhanced the accuracy (ACC) of non-invasive preoperative LVSI prediction, demonstrating certain potential for clinical application.