CT Images Based Methods For Prognostic Prediction Of Esophageal Cancer Patients

Esophageal cancer is a common digestive tract tumor,and about 500,000 people worldwide die of esophageal cancer every year.Accurately predict-ing the prognosis of patients with esophageal cancer is crucial for precise clinical treatment.At this stage,with the rapid development of artificial intelligence,it has become possible to accurately predict the prognosis of patients after treatment through deep learning methods.Although there are many methods for predicting the prognosis of patients,there are few studies on the prognosis of patients with esophageal cancer.At the same time,dif-ferent stages of esophageal cancer patients take different treatment plans.Therefore,this paper proposes two different deep learning algorithms for different treatment options to predict the prognosis of patients with differ-ent stages of esophageal cancer.The main work details and innovations of this paper are as follows:(1)Accurately predicting the pathological complete response(p CR)of patients after neoadjuvant chemoradiotherapy(n CRT)is essential for clin-ical precision treatment.However,the existing methods of predicting p CR in esophageal cancer are based on single stage data.These approaches do not take into account the effect of tumor changes in longitudinal multi-stage data on predicted outcomes.In addition,traditional fusion methods may cause information redundancy.Therefore,in order to explore the relation-ship between tumor changes before and after n CRT and p CR,this paper propose a CT based multi-loss disentangled representation learning(ML-DRL)to predict p CR of esophageal cancer patients after n CRT.Specifi-cally,to discover tumor changes before and after n CRT,we first disen-tangle the latent variables of longitudinal multi-stage features into inherent and variational components.Then,we define a multi-loss function to en-sure the effectiveness and structure of disentanglement,which consists of a cross-cycle reconstruction loss,an inherent-variational loss and a super-vised classification loss.Finally,an adaptive gradient normalization algo-rithm is applied to balance the training of multiple loss terms by dynami-cally tuning the gradient magnitudes.Due to the cooperation of the multi-loss function and adaptive gradient normalization algorithm,MLDRL can effectively disentangle the inherent information and variational information of longitudinal multi-stage data and realize the effective fusion of informa-tion.The proposed method is evaluated on multi-center datasets,and the ex-perimental results show that our method not only outperforms several state-of-art methods in p CR prediction,but also achieves better performance in the prognostic analysis of multi-center datasets.(2)Accurately predicting the survival time of patients after surgical treatment is of great significance for personalized treatment.However,most of the current research on predicting survival time is based on data with accurate labels to train models,ignoring the information that is useful for predicting survival time hidden in censored data that is ubiquitous in clinical practice.Therefore,in this study,this paper propose an adaptive pseudo-label selection and correction algorithm(APLSC)based on preoperative CT images to further improve the accuracy of the model in predicting post-operative survival time of esophageal cancer patients by using censored data.Specifically,in the pseudo-label selection process,the thresholds of different categories in the pseudo-label selection process are dynamically adjusted according to the learning accuracy of different categories.Second,for the selected pseudo-labels included in training,a KL loss and an addi-tional correction loss are designed to reduce the impact of potential noise in the selected pseudo-labels on model performance.Finally,the experi-mental results show that the proposed adaptive pseudo-label selection and correction algorithm can effectively utilize the potential information in the censored data to further improve the performance of the model in predicting the postoperative survival time of patients with esophageal cancer.