Transvaginal ultrasound-based radiomics and integrated clinical indicators via multimodal deep learning for prediction of endometrial polyp recurrence after hysteroscopic surgery

OBJECTIVES: To evaluate a multimodal deep learning model integrating preoperative transvaginal ultrasound (TVUS)-based radiomics features and clinical indicators for predicting 1-year postoperative recurrence of endometrial polyps (EP) after hysteroscopic polypectomy. METHODS: A total of 116 patients with pathologically confirmed EP were assigned to a training cohort (n=81) and validation cohort (n=35). Radiomics features were extracted from TVUS images, and deep learning features were obtained using ResNet-based networks. These features, with clinical variables, were combined to build a multimodal model. Feature selection in the training cohort used reproducibility filtering (intraclass correlation coefficient [ICC] >0.80), univariate analysis, Pearson correlation (|r|>0.90), and least absolute shrinkage and selection operator (LASSO) regression. Model performance was evaluated by receiver operating characteristic (ROC) curves, area under the curve (AUC), calibration, and decision curve analysis (DCA). RESULTS: The multimodal model achieved AUCs of 0.941 (95 % CI: 0.897-0.985) and 0.922 (95 % CI: 0.852-0.992) in training and validation cohorts, outperforming clinical (0.812, 0.791) and radiomics-only models (0.871, 0.843). DeLong tests were significant (p<0.05). DCA showed higher clinical net benefit. CONCLUSIONS: This multimodal model effectively predicts 1-year recurrence after hysteroscopic EP resection, supporting individualized postoperative management.