Abstract
BACKGROUND: Lymphovascular space invasion (LVSI) is a key prognostic factor in endometrial cancer and integral to the updated 2023 International Federation of Gynecology and Obstetrics (FIGO) staging system. However, its preoperative detection remains challenging. This study aimed to develop and validate an integrated model combining clinical variables, magnetic resonance imaging (MRI)-based radiomics, and deep learning features for predicting LVSI preoperatively. METHODS: This retrospective study enrolled 580 patients with pathologically proven endometrial cancer from Shenzhen People's Hospital and Shenzhen Second People's Hospital. Radiomics and deep learning features were extracted via T2-weighted imaging (T2WI), apparent diffusion coefficient (ADC) maps, and late contrast-enhanced T1-weighted imaging (T1CE). Following data dimensionality reduction and feature selection, a comprehensive model integrating clinical data, MRI-based radiomics, and a deep learning (CRDL model), as well as a clinical model, an MRI-based radiomics model (R model), and an MRI-based deep learning (DL) model, were constructed on the training cohort via the support vector machine (SVM) classifier. The predictive performances of these models were evaluated with the area under the curve (AUC) and were compared with the Delong test in the training cohort. The optimal model was validated both in the internal and external validation cohorts. RESULTS: The AUCs of the clinical, R, DL, and CRDL models were 0.748, 0.810, 0.823, and 0.924 in the training cohort, respectively. The DeLong test showed that the predictive performance of the CRDL model was significantly superior to that of the other three models, and the difference remained statistically significant after Bonferroni correction. The AUC of the CRDL model was higher in the training cohort than in the internal (AUC =0.873) and external (AUC =0.831) validation cohorts, but the differences were not statistically significant (P>0.05), suggesting good generalizability across different datasets. CONCLUSIONS: The comprehensive model (CRDL model) could preoperatively predict the LVSI status in endometrial cancer with high performance, which may be a promising imaging biomarker for preoperative risk stratification and support individualized treatment decision-making.