Prediction of tumor grade in endometrioid carcinoma using a deep learning radiomics model from ultrasound images: a multicenter study.

Background: Endometrial endometrioid carcinoma (EEC) tumor grade is a critical prognostic factor, but its accurate preoperative non-invasive assessment remains challenging due to the limitations of conventional imaging and biopsy. Transvaginal ultrasound (TVUS) is the primary imaging modality but offers limited quantitative insights for grading. Deep learning radiomics (DLR), which combines the strengths of deep learning (DL) for automatic feature extraction and radiomics for quantifying tumor heterogeneity, holds promise for uncovering prognostic information from routine ultrasound images. This study aimed to develop and validate a DLR model based on preoperative TVUS images for the non-invasive differentiation of EEC tumor grades. Methods: A total of 297 EEC cases with confirmed histological grades, including grade 1 (G1), grade 2 (G2), and grade 3 (G3), were selected from 1,258 endometrial cancer patients who underwent hysterectomy across eight centers. Radiomics features were extracted from TVUS images, and a radiomics model was constructed using the extreme gradient boosting (XGBoost) algorithm. Simultaneously, DL features were extracted using ResNet-50 to establish a DL model. A combined DLR model was then developed by integrating both feature sets, employing five-fold cross-validation for internal validation. An external testing cohort comprising 129 cases with corresponding grading data was collected from three independent centers. The performance of the three models in identifying EEC differentiation grade was compared using receiver operating characteristic (ROC) curve analysis to evaluate their diagnostic accuracy. Results: In differentiating EEC grades, the DLR model outperformed both the single radiomics and DL models. In the identification of G3 and G1/G2, the AUC of the DLR model was 0.871 and 0.843 in the training cohort and the external testing cohort, respectively. The AUC of the identification of G2 and G1 was 0.856 and 0.816 in the training cohort and the external testing cohort, respectively. Decision curve analysis confirmed the clinical utility of the DLR model. Conclusions: The DLR model based on TVUS images shows potential value for the non-invasive differentiation of EEC tumor grading and provides a useful supplement for non-invasive clinical staging of endometrial carcinoma prior to surgery.