Establishment And Clinical Research Of Artificial Intelligence Diagnosis System For Metastatic Lymph Nodes Of Adenocarcinoma At The Gastroesophageal Junction Based On Convolutional Neural Network

Objective: The global incidence of adenocarcinoma of Esophagogastric Junction(AEG)is increasing year by year,and lymph node metastasis will significantly reduce the survival rate of patients with this disease.The route of AEG lymph node metastasis is complex,and accurate prediction of the area of lymph node metastasis before surgery is of great significance for the selection of the diagnosis and treatment plan of the disease.Contrast-enhanced CT is a preoperative auxiliary examination widely used in AEG patients.Convolutional Neural Network(CNN)has a good application performance in the field of image recognition.In this study,based on Faster Region-based Convolutional Neural Network(Faster RCNN)and enhanced CT images,an artificial intelligence(AI)diagnosis system for AEG lymph node metastasis was established.The purpose is to explore the intelligence of preoperative prediction of metastatic lymph node regions in patients with AEG,and to assist clinicians in diagnosis and treatment.Methods: First,our researchers retrospectively collected contrast-enhanced CT images of the chest and abdomen of 248 patients with AEG who underwent radical surgery in Qingdao University Affiliated Hospital from December 2015 to December 2019.We integrated the reading results of radiologists,gastrointestinal surgeons,and thoracic surgeons,used Labelimg software to mark metastatic lymph node regions,and established a CT image database of AEG patients.Then we randomly divided all the enrolled patients into training group and test group according to the ratio of 3:1,and used the CT images of the training group to train and fine-tune the pre-trained Faster RCNN to establish an AI diagnosis system.The basic structure of the target detection network of the system includes four parts: convolution layer,RPN layer,ROI Pooling layer and classifier.We further tested the AI diagnostic system by using the enhanced CT images of patients in the test group,output the detection frame of suspected lymph node metastasis in the images of the test group,and calculated the probability of lymph node metastasis in the detection frame.In this way,the test data of the test group of the AI diagnostic system is obtained.Finally,we document the cumulative time-consuming of the diagnostic process by the multidisciplinary clinician and the AI diagnostic system,respectively.The average time-consuming of diagnosing a series of CT images of a single patient and diagnosing a single CT image by the two methods were calculated,and then the two were compared.At the same time,the performance evaluation index of the AI diagnosis system is calculated by predicting probability and Intersection Over Union(IOU).The accuracy,specificity,sensitivity,positive predictive value,negative predictive value,Receiver Operating Characteristic Curve(ROC)and Area Under Curve(AUC)and other indicators were used to evaluate the diagnostic level of the AI diagnostic system for lymph node metastasis.Results: The average time taken by multidisciplinary clinicians to diagnose a single patient’s series of CT images was about 351.37 s,and the average time taken to diagnose a single image was about 4.78 s.A total of 354 CT images are diagnosed with lymph node metastasis.The average time taken by the AI diagnostic system to diagnose a series of CT images of a single patient is 10.72 s,and the average time of diagnosing a single image is about 0.15 s.A total of 348 CT images are diagnosed with lymph node metastasis.The AUC value of the diagnostic results of the AI system was 0.912,the accuracy was 0.870,the sensitivity was 0.858,the specificity was 0.883,the positive predictive value was0.892,and the negative predictive value was 0.847.A total of 301 images were obtained with the same diagnosis results for the location and number of metastatic lymph nodes.Limitations: The purpose of this study was to help physicians read CT images preoperatively to guide preoperative diagnosis and treatment of patients with AEG.Because this study was retrospective,a one-to-one correspondence between preoperative CT images and postoperative histopathological examinations of lymph nodes was not possible.Therefore,the diagnostic criteria learned by the AI diagnostic system are CT signs,not pathological criteria,and its accuracy can only be infinitely close to the comprehensive level of physicians,but cannot reach the level of pathological diagnosis.Conclusion: The accuracy of the AI diagnostic system in identifying metastatic lymph nodes in AEG-enhanced CT images is high,close to the diagnostic level of clinicians,and the recognition speed is faster than that of clinicians.Therefore,the AI diagnostic system we established can reduce the work pressure of clinicians to a certain extent.And based on its high prediction level,the AI diagnostic system has the potential to assist clinical diagnosis and treatment.