Figure 1.

Workflow of the prediction model. First, radiomic features were extracted from CT images using multimodal radiomics platform. To remove the centric effect, we applied ComBat realignment on the radiomic features. Second, the dimension feature was reduced with feature selection to avoid the overfitting effect. The remained features after feature selection were applied to build a model based on DeepSurv. Finally, the assessment of model performance relied on group-wise survival curves, C-index, and time-dependent ROC curves.