| BackgroundThe incidence of adenocarcinoma of the esophagogastric junction(AEGJ)had increased in recent years.Since the signs and symptoms of most patients tend to be latent and nonspecific,tumors were often diagnosed at advanced stage.Esophagogastroduodenoscopy is the standard method in the diagnosis of early AEGJ,but the characteristics of early cancerous foci are usually not obvious and can be easily confused with noncancerous lesions such as gastritis.The endoscopist’s inability to concentrate because of fatigue or emotional factors may contribute to the occurrence of missed diagnoses.Artificial intelligence(AI)uses computers to simulate certain thought processes and forms of intelligent behavior.In the field of gastrointestinal endoscopy,the computer-aided detection system based on deep learning artificial was used to hint suspected lesions and reduce missed diagnosis.So far,the research on AI in detecting early AEGJ was limited,and based on still images.This study aimed to develop an AI model for detecting early AEGJ by analyzing white light endoscopic images.MethodsThis study reviewed the endoscopic system database of Qilu Hospital of Shandong University,and collected 1608 images of AEGJ from 409 patients and 1656 non-cancerous images from 504 patients for model training and validation from January 2014 to June 2020.You Only Look Once Version 5X(YOLOv5x)was used as a model to build a computer aided detection system.Endoscopic data,gender,age,lesion type,and histopathology were prospectively collected from 200 patients from July 2020 to June 2021 according to inclusion and exclusion criteria.The image test set included 301 images of early cancer from 89 patients and 304 images of non-cancer from 111 patients.Endoscopic videos of 29 patients were recorded as a video test set.Five endoscopy experts and three junior endoscopists independently judged the image test set.The interobserver agreement was compared by Kappa analysis.The performance of the AI model was compared with junior endoscopists and the assistant efficiency of the AI model for novices was evaluated.ResultsThe sensitivity,specificity,accuracy,positive predictive value and negative predictive value of the AI model in the image test set were 0.924,0.852,0.888,0.861 and 0.918,respectively.At the patient level,the sensitivity,specificity and accuracy of the model were 0.993,0.739 and 0.825.The kappa values between the model and experts were high,which are 0.722,0.737,0.749,0.761 and 0.805 respectively.Al model had higher sensitivity,specificity and accuracy than novices,and could significantly improve the sensitivity of novices(0.949 vs.0.807,p<0.01).In the video test,the AI model correctly identified 93.1%of early cancers(27/29).ConclusionsIn this study,an AI model for detecting early AEGJ using white light endoscopy was designed and developed.Static images and dynamic videos were collected prospectively to evaluate the detection performance of the model.The AI model had high diagnostic consistency with endoscopy experts,and could help novices effectively identify early cancer lesions.The model could accurately find the early AEGJ lesions,which could reduce the missed lesions in the process of endoscopy due to the lack of experience and fatigue of endoscopists.Multi-center,large-sample randomized controlled studies may be needed in the future to further explore the value of AI model in practical clinical work. |
PDF Full Text Request