Survival Prediction of Esophageal Cancer Using 3D Residual Autoencoder and Proportional Hazards Model

Abstract

Survival prediction of esophageal cancer is a difficult task for doctors looking to make personalized cancer treatment plan. Also, explicitly-designed handcrafted features obtained from medical images need prior medical knowledge which is usually limited and incomplete, and yields unsatisfying survival prediction. To address these challenges, we propose a novel and efficient deep survival learning network for evaluating clinical outcome prior to radiotherapy.Three hundred twenty-six patients diagnosed with thoracic esophageal squamous cell carcinoma between December 2010 and September 2017 were analyzed. We used a deep learning network to extract potential features from gross tumor volumes in CT scans. For the first time, we integrate deep representations extracted from network above with dose information as feature signatures. A survival prediction model was generated to reliably predict a hazard ratio of patients with esophageal cancer. The model used Cox proportional hazards to model interactions between a patient's feature signatures and clinical outcome. The accuracy (concordance index [CI]) of this model subsequently was compared with the conventional technique.The experiments were performed using a combination of data from two input sources: deep features from CT (CI = 0.686) and dose information (CI = 0.576). Specifically, the fusion of them (CI = 0.702) significantly improved the performance of prediction comparing to using just one of them. The majority of models demonstrated improved accuracy compared with corresponding conventional models for survival predictions. Furthermore, experimental results demonstrated that the optimal model achieved concordance index of 0.702 which outperformed state-of-the-art methods using simpler and more accessible data.This study provided a novel method for accurately predicting patient hazard ratios for use in esophageal cancer treatments. Performance improved when more types of data were integrated in the input of our model. The prognosis prediction model can help clinical work to judge the survival rate. This study offered a new avenue for personalized treatment and survival prognosis of esophageal cancer based on deep learning technology.