Research On Survival Of Prognosis Model In Esophageal Cancer Based On Radiomics And Deep Learning

Esophageal cancer is one of the most lethal cancers in the world,with a high death rate,and it occurs frequently in our country.There are often no obvious symptoms in the early stage and the prognosis is relatively poor.By making effective prognostic survival predictions,physicians can better assess the patient’s condition and make appropriate treatment measures,thus improving patients’ survival rate.In this paper,the patients with esophageal cancer were taken as research object.The medical information of CT images was mined deeply,and the prognosis and survival prediction model of esophageal cancer patients was constructed by using radiomics and deep learning technology.The main research contents are as follows:(1)In view of the problem that traditional prognostic indicators of esophageal cancer cannot fully reflect the internal heterogeneity of tumors,this paper presents a survival prediction method of esophageal cancer based on radiomics in order to obtain more accurate prognosis evaluation result.Through the radiomics technology,the comprehensive,non-invasive and repeatable examination of the tumor can be realized,so that the heterogeneity of the lesion area of the patients with esophageal cancer can be evaluated more accurately,and the heterogeneity can be taken as a new research thought for predicting the prognosis of the patients with esophageal cancer.Firstly,CT images of esophageal cancer patients are obtained from the image archiving and interaction system(PACS)of the scientific research cooperative hospital;then,imaging doctors use 3D-Slicer software to manually segment to obtain region of interest(ROI);and quantitative CT image group features of esophageal cancer patients are extracted by using the segmented ROI.By applying correlation analysis and LASSO algorithm,the dimension reduction processing is carried out on the characteristics to construct an image histologic prognosis label,the radiomics prognosis label is combined with clinical factors closely related to the prognosis of the esophageal cancer,the model is fitted by Cox risk proportion regression,and visualized using a nomogram.By calculating C-index,AUC and other evaluation indexes,it was verified that the prediction model could predict the prognosis of patients with esophageal cancer relatively accurately.Kaplan-Meier curve,calibration curve and clinical decision curve were used to evaluate the prediction performance of the model for prognosis of esophageal cancer.(2)In view of the lack of comprehensiveness and rationality of data feature mining in traditional radiomics,which leads to unsatisfactory effect of feature extraction.In this paper,a survival prediction model of esophageal cancer is proposed by combining deep learning and radiomics.Deep learning can extract more objective,quantitative and deeper tumor features from medical images,which are more abstract,more accurate,more reliable and more effective than traditional radiomics features.Unlike traditional imaging genomics features,clinical goal-oriented deep learning features can be learned automatically from the data.Deep learning features contain"real world & quot;textures that are extracted from a pre-trained sensenet-169 network by a migration learning strategy.Then,based on the mutual information(MI)and the survival state of the features,the features are sorted by using the minimum redundancy maximum correlation(m RMR)scheme,the first 50 features are reserved,the optimal features are selected by LASSO regression algorithm,and the deep learning image group(DLR)label is constructed by using the selected features.The results showed that the performance of deep learning characteristics in predicting survival time of patients with esophageal cancer was better than that of traditional imaging group features.The same method is used to construct the traditional radiomics label.The deep learning imaging genomics nomogram is constructed by combining deep learning radiomics label,imaging genomics label and clinical factors through Cox risk proportional regression algorithm.The results showed that the predictive model had high accuracy in predicting the 3-year survival of patients with esophageal cancer and was significantly better than the prognosis prediction model constructed using deep learning radiomics tags and radiomics tags alone.(3)In view of the small amount of standardized data of esophageal cancer images and the difficulty in acquiring labeling information,this paper starts with the thinking mode of small sample learning,gives full play to the advantages of small sample learning in small data processing,and studies how to design the model on the small sample data and fully mine the information of a small amount of standard data.In this paper,a survival prediction model of esophageal cancer based on graph neural network is proposed by combining the idea of few-shot learning and graph neural network.The metric-based meta-learning strategy is introduced to construct a graph neural network model of feature fusion by using the features of radiomics,depth learning and clinical features extracted from CT images.According to the relationship between samples,further excavate the information contained in the samples to enhance the characteristics of the samples.Different task experiences are accumulated through the meta-learning device to guide the updating of graph nodes.By updating the multi-layer graph neural network,the category information contained in a small number of samples is highlighted,and the information contained in the graph neural network is fully considered,so that the three-year overall survival rate of the esophageal cancer can be accurately predicted under the condition of few training data.