Research On Key Problems Of Preoperative Image Diagnosis In Esophageal Cancer Based On Deep Convolution Neural Network

Esophageal cancer has become one of the eight common malignant tumors due to its high morbidity and mortality.Most patients are already at the stage of locally advanced or advanced at the time of diagnosis,missed the best period of treatment,and the prognosis is extremely poor.Factors affecting the prognosis of esophageal cancer include clinical stage,lymph node metastasis status,and pathological type.Early diagnosis and personalized treatment and reasonable control of the risk factors can effectively improve the prognosis of patients with esophageal cancer.Therefore,the early diagnosis and treatment of esophageal cancer becomes very important.In recent years,machine learning technology represented by deep learning has developed rapidly.Great success has been achieved in natural language processing,target recognition,image processing,etc.Medical imaging is a special image.The application of deep learning technology has brought new opportunities for medical image data mining.Especially in the identification,diagnosis,detection and prediction of diseases,it is more and more widely used.The emergence of new technologies has brought new opportunities for the clinical diagnosis of esophageal cancer.This article summarizes the research progress of deep learning in medical imaging in recent years,as well as the challenges and technical difficulties faced by the clinical diagnosis of preoperative staging and lymph node metastasis of esophageal cancer,and introduces deep learning technology to the preoperative related diagnosis of esophageal cancer.The main research contents and innovations are as follows:1.In view of the problem that esophageal cancer cannot be diagnosed due to the lack of key diagnosis images during the clinical diagnosis,this paper developed a generative adversarial network to generate corresponding contrast-enhanced CT images from noncontrast-enhanced CT images of patients.And designed a perceptual feature extractor in the network to measure the relationship between the generated enhanced image and the real image,while constructing a perceptual loss function to make the network training more stable,so that the generated image contains more details that meet the human visual perception information.2.In view of the low sensitivity and low accuracy of CT images for esophageal cancer staging,this paper proposes a feature extraction model based on perceptual attention mechanism of image spatial information combined with time serial semantic information.In addition to the spatial information described by conventional radiomics,a long-short-term memory network(LSTM)is introduced to extract the perceptual time series features at each time slice to obtain more information related to tumor staging from the time dimension.Aiming at the fusion of information in different scales such as spatial and time series,a perceptual attention mechanism is introduced to measure the importance of information of different dimensions to the stage.In addition,the model was validated in an independent external validation data.3.In view of the clinical problems of low accuracy,high false positive and negative rate of CT images in the diagnosis of lymph node metastasis of esophageal cancer before surgery,we propose a multi-level imaging feature comprehensive prediction model of lymph node metastasis.From the aspects of machine vision,traditional radiomics and perception,a sparse autoencoding feature fusion network is designed to process highdimensional features of different dimensions and different levels.And use statistical methods to build a lymph node metastasis prediction model,and comprehensively evaluate the model performance from the aspects of identification ability,classification ability and clinical practicality.In summary,this article has conducted research on three key scientific issues that are urgently needed to be solved clinically in esophageal cancer: image loss,preoperative accurate staging,and lymph node metastasis.The content covers the generation of medical images based on generative confrontation network,and the feature extraction and fusion of deep convolutional neural networks,multi-CT level feature information fusion,and model construction and evaluation.These studies have a positive role in promoting preoperative clinical diagnosis and treatment decisions of esophageal cancer.