An interpretable multimodal machine learning model integrating transvaginal ultrasound and clinical features for differentiating benign from malignant endometrial thickening in postmenopausal women

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.