Ultrasound-based radiomics combined with clinical features for predicting myometrial invasion depth in endometrial carcinoma

Endometrial carcinoma (EC) is one of the most prevalent gynecologic malignancies worldwide, with rising incidence. Preoperative assessment of myometrial invasion depth is crucial for treatment planning and prognostication. Deep invasion (> 50% myometrial thickness) is a well-recognized risk factor for lymph node metastasis, recurrence, and poor survival, often requiring extended surgery and adjuvant therapy. Transvaginal ultrasound is commonly used for evaluation, but subjective interpretation based on morphological features—such as endometrial thickening and myometrial disruption—is highly operator-dependent, limiting its reliability and reproducibility. To address this limitation, we developed an interpretable machine learning model integrating clinical and ultrasound radiomics data to predict the depth of myometrial invasion in EC. Retrospective data from 294 patients were analyzed, with 205 (70%)assigned to the training set and 89 (30%)to the test set. After feature extraction and selection using intraclass/interclass correlation coefficients and least absolute shrinkage and selection operator regression, radscores were calculated and combined with clinical variables—including demographics, tumor markers, and ultrasound characteristics—to build nine machine learning models. The random forest model demonstrated the best performance, achieving an area under the receiver operating characteristic curve (AUC) of 0.85 and an F1 score of 0.81 for predicting deep myometrial invasion, and an AUC of 0.77 and an F1 score of 0.74 for any myometrial invasion in the test set. The SHapley Additive exPlanations analysis revealed that maximum tumor diameter, age, carcinoembryonic antigen levels, and radscore1 were key predictors for deep invasion, while tumor diameter, age, radscore2, and blood flow signal grade contributed most to predicting any invasion. This explainable model provides a non-invasive, clinically relevant tool for preoperative risk stratification, potentially supporting personalized treatment decisions in endometrial carcinoma.