Deep learning-based super-resolution dynamic contrast-enhanced radiomics model for predicting NSMP endometrial cancer

To determine the performance of a deep learning-based super-resolution (SR) dynamic contrast-enhanced (DCE) radiomic model in predicting nonspecific molecular profile (NSMP) endometrial cancer (EC).

This study included 140 surgically confirmed EC patients (76 NSMP-type EC patients and 64 non-NSMP-type EC patients) who were randomly divided into training and testing cohorts. Deep learning-based SR reconstruction techniques were applied to convert original-resolution (OR) DCE images into SR images. Radiomic features were extracted from both the SR and OR images. Logistic regression (LR), support vector machine (SVM), and multilayer perceptron (MLP) algorithms were used to develop SR-DCE and OR-DCE models, respectively. Model performance was evaluated via area under the curve (AUC) analysis and decision curve analysis (DCA) methods.

Among the three algorithms (LR, SVM, and MLP), the diagnostic effectiveness of the SR-DCE model was superior to that of the OR-DCE model (P < 0.05). In the testing set, the AUC values for the SR-DCE model were 0.841 (95% confidence interval (CI): 0.724-0.959), 0.800 (95% CI: 0.664-0.937) and 0.764 (95% CI: 0.618-0.911), whereas those for the OR-DCE model were 0.637 (95% CI: 0.464-0.810), 0.603 (95% CI: 0.422-0.785) and 0.656 (95% CI: 0.495-0.818), respectively. Specifically, for the SR-DCE model, both the LR and SVM algorithms exhibited significantly higher AUC values in the testing set than the MLP algorithm did (LR vs. MLP, P = 0.012; SVM vs. MLP, P = 0.041). The clinical utility of the three algorithms was favorable for the SR-DCE model.

Deep learning-based SR reconstruction technology can increase the diagnostic effectiveness of the DCE radiomics model for NSMP EC, and it could become a noninvasive preoperative predictive tool for NSMP EC.