Preoperative prediction of aggressive endometrial cancer using multiparametric MRI-based deep transfer learning...

Accurate preoperative prediction of endometrial cancer (EC) aggressiveness is critical for individualized treatment planning. This study proposes a method for integrating multimodal data to improve the prediction of aggressiveness in EC.

A total of 207 patients with pathologically confirmed EC were retrospectively enrolled. The patients were randomized (7:3) into a training cohort (n=144) and a test cohort (n=63). All patients underwent preoperative MRI including T2-weighted imaging, diffusion-weighted imaging, apparent diffusion coefficient mapping, and contrast-enhanced T1-weighted imaging (CE-T1WI). Deep learning (DL) models using ResNet50, ResNet101, DenseNet121 were employed to extract deep transfer learning (DTL) features. Three decision-level fusion strategies (mean, maximum, and minimum) were applied to integrate the multi-sequence model outputs, from which the optimal DTL model was selected. Subsequently, a combined clinical-DTL model was constructed by incorporating independent clinical predictors identified through univariate and multivariate logistic regression analyses. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), clinical utility by decision curve analysis, and goodness-of-fit through calibration curves.

The mean fusion model integrating features from T2WI, ADC, and CE-T1WI (excluding DWI due to suboptimal performance) yielded the best predictive efficacy, with an AUC of 0.963 [95% confidence interval (CI): 0.933-0.992] in the training cohort and 0.925 (95% CI: 0.859-0.990) in the test cohort. The combined clinical-DTL model further achieved AUCs of 0.972 (95%CI: 0.948-0.997) and 0.950 (95%CI: 0.891-1.000) in the training and test cohorts, respectively. Decision curve analysis and calibration analyses confirmed its clinical utility and good model fit.

The proposed DTL model based on multiparametric MRI demonstrates strong performance in preoperatively predicting aggressive EC. The integration of clinical features further enhances model performance, offering a non-invasive tool to support personalized treatment strategies.