An interpretable integration model improving disease-free survival prediction for gastric cancer based on CT images and clinical parameters

Preoperative prediction of disease-free survival of gastric cancer is significantly important in clinical practice. Existing studies showed the potentials of CT images in identifying predicting the disease-free survival of gastric cancer. However, no studies to date have combined deep features with radiomics features and clinical features. In this study, we proposed a model which embedded radiomics features and clinical features into deep learning model for improving the prediction performance. Our models showed a 3%-5% C-index improvement and 10% AUC improvement in predicting DFS and disease event. Interpretation analysis including T-SNE visualization and Grad-CAM visualization revealed that the model extract biologically meaning features, which are potentially useful in predicting disease trajectory and reveal tumor heterogeneity. The embedding of radiomics features and clinical features into deep learning model could guide the deep learning to learn biologically meaningful information and further improve the performance on the DFS prediction of gastric cancer. The proposed model would be extendable to related problems, at least in few-shot medical image learning. Key Points An integration model combining deep features, radiomics features and clinical parameters improved disease-free-survival prediction of gastric cancer by 3%-5% C-index. Embedding radiomics and clinical features into deep learning model through concatenation and loss design improved feature extraction ability of deep network. The model revealed disease progression trajectory and tumor heterogeneity.