Transvaginal ultrasound-based deep learning radiomics for preoperative identification of non-endometrioid endometrial carcinoma: A multicenter study.
Authors
Affiliations (3)
Affiliations (3)
- Department of Ultrasound, West China School of Public Health and West China Fourth Hospital, Sichuan University, Chengdu, Sichuan, 610000, China; Department of Ultrasound, Beijing Anzhen Nanchong Hospital of Capital Medical University & Nanchong Central Hospital, Nanchong, Sichuan, 637000, China.
- Department of Ultrasound, Peking Union Medical College Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, 100005, China.
- Department of Ultrasound, West China School of Public Health and West China Fourth Hospital, Sichuan University, Chengdu, Sichuan, 610000, China. Electronic address: [email protected].
Abstract
Preoperative identification of non-endometrioid endometrial carcinoma (NEEC) is clinically important because non-endometrioid histology may influence risk assessment and treatment planning. This study aimed to develop and externally validate a transvaginal ultrasound-based deep learning radiomics (DLR) model for preoperative identification of NEEC. A total of 663 patients with pathologically confirmed endometrial carcinoma from 11 centers were retrospectively included. The training cohort consisted of 483 patients from eight centers, including 161 NEEC and 322 EEC cases. An independent external test cohort included 180 patients from three additional centers, including 60 NEEC and 120 EEC cases. Radiomics features were extracted using PyRadiomics, and deep learning features were extracted using ResNet-101. Radiomics, deep learning, and combined DLR models were developed using machine learning classifiers and evaluated by receiver operating characteristic analysis, calibration curves, decision curve analysis, and external validation. The combined DLR model achieved the best diagnostic performance among the three models. In the external test cohort, the DLR model yielded an AUC of 0.837, with an accuracy of 0.756, sensitivity of 0.800, and specificity of 0.733. The DLR model outperformed the radiomics model and the deep learning model in both the training and external test cohorts. The transvaginal ultrasound-based DLR model showed potential as a noninvasive tool for preoperative identification of NEEC. This model may provide complementary information for individualized preoperative assessment, although prospective validation is required before clinical implementation.