3.2 Prediction of CR
A combination of 10 clinicopathological parameters and 34 CT radiomic features performed best in predicting CR [balanced accuracies: 79.73% and 80.79%; AUC (95% confidence interval (CI)): 0.88 (0.82-0.93) and 0.87 (0.77-0.96) in the training and validation sets]. The CT radiomic model outperformed clinicopathological model in predicting CR (Table 3). Figure 2a shows the ROC curves of three models for classifying the CR and non-CR groups. Additionally, only using the tumor volume to predict CR achieved balanced accuracies of 64.92% and 72.87% in the training and validation sets.
The CT radiomic model outperformed MRI radiomic model for predicting CR (balanced accuracy: 74.51% and 52.46% in the validation set), and the addition of MRI radiomic features decreased the performance of the clinical and radiomic model (Table S2 and Figure S1).