Research On Artificial Intelligence In Sites Identification During Esophagogastroduodenoscopy And In The Automatic Detection Of Precancerous Lesions And Superficial Esophageal Cancer

Introduction:In recent years,artificial intelligence has achieved rapid development,and its image identification technology has made remarkable progress in various fields.The technique utilizes deep convolutional neural network(DCNN)for self-learning,and extracts key features without human instruction by changing the internal parameters of the algorithm and neural network.So far,it is already used in the medical field to assist and improve the diagnostic level of physicians,and preliminary results have been achieved especially in endoscopic-assisted diagnosis.Gastroscopy is the main method for diagnosing upper gastrointestinal diseases,and it not only requires physicians to have appropriate skills,but also requires sufficient knowledge to correctly identify lesions.The trained DCNN auxiliary diagnosis system can not only assist endoscopists in identifying lesions,may also improve the level of diagnosis and reduce missed lesions.The critical first step in applying DCCN to diagnose upper gastrointestinal diseases is to correctly identify the anatomical sites.Therefore,in this study,we first used DCNN to build an artificial intelligence-assisted upper gastrointestinal tract identification model to correctly identify the anatomical parts of the upper gastrointestinal tract as a preliminary step in computer-aided diagnosis of upper gastrointestinal lesions.In addition,the auxiliary sites identification model is also expected to become an objective evaluation model for quality control and standardized procedure of endoscopy,so as to monitor the blind spots of upper gastrointestinal endoscopy,standardize endoscopy procedures,and improve procedure quality.Materials and Methods:We divided the upper gastrointestinal endoscopy area into 30 sites based on anatomy,and collected 500-1000 images for each site.A total of 21,310 upper gastrointestinal endoscopy images were collected from our hospital from January 2019 to June 2021,of which 19,191 images were used to construct sites identification model,and 2,119 images were used to test the model.The research compared the performance of the ResNet model and the RESENet model constructed by the two DCCN networks Inception-ResNetV2(ResNetV2)and Inception-ResNetV2-Squeeze-Excitation Networks(RESENet)on the sites identification.Results:The average accuracy,sensitivity,specificity,positive predictive value and negative predictive value of the ResNetV2 model and RESENet model in sites identification were 97.60%vs 99.34%,75.58%vs 99.57%,98.75%vs 99.66%,63.44%vs 90.20%,98.76%vs 99.66%,respectively.In these evaluation indicators,the performance of the RESENet model was better than that of the ResNetV2 model,and the difference is significant(p<0.05).Conclusions:Compared with the traditional ResNetV2 model,the accuracy,sensitivity,specificity,positive predictive value and negative predictive value of the artificial intelligence-assisted esophagogastroduodenoscopy sites identification model constructed by RESENNet have been significantly improved.It has great potential in the ability to classify and identify anatomical sites.It can ensure the integrity of the esophagogastroduodenoscopy procedures and is expected to become an important assistant for standardizing esophagogastroduodenoscopy and improving quality of the procedures,as well as a significant tool for quality supervision and control of esophagogastroduodenoscopy.Introduction:Early diagnosis of esophageal cancer is crucial to reducing mortality,and endoscopic screening is an important means to improve the early diagnosis rate of esophageal cancer.Due to the lack of specific symptoms of early esophageal cancer and precancerous lesions,and the atypical morphology of white light imaging under endoscopy,the missed diagnosis rate is high.The use of narrow band inaging(NBI)and chromoendoscopy requires sufficient experience and technology,and endoscopists who lack relevant experience or technology need a simpler,more effective and intelligent method to assist in the diagnosis of early esophageal cancer.For the past years,the application of AI in endoscopy has shown great potential,especially in the detection and identification of gastrointestinal lesions.In this study,an artificial intelligence deep learning method was used to construct an auxiliary esophageal precancerous lesion and superficial esophageal squamous cell carcinoma lesion identification model based on the Yolov51 model,aiming to reduce missed diagnosis and improve the diagnostic level of esophageal precancerous lesions and superficial esophageal squamous cell carcinoma.Materials and Methods:We collected 13009 esophagus pictures of white light imaging(WLI),NBI and lugol chromoendoscopy(LCE)from 1126 patients in our hospital from June 2019 to July 2021,including low-grade intraepithelial neoplasia,high-grade intraepithelial neoplasia,and esophageal squamous cell carcinoma limited to the mucosal layer,Benign esophageal disease(reflux esophagitis,fungal esophagitis,esophagogastric mucosal ectopic,pathologically confirmed esophagitis with positive iodine staining)and normal esophagus.By computer random function method,these images were divided into training set detabase(11547 images of 1025 patients)and validation set detabase(1462 images of 101 patients).The YOLOv51 model was trained and constructed with the training set datebase and was tested with the validation set detabse.At the same time,two senior and two junior endoscopists made a diagnosis on the validation set database.We compared the performance of the YOLOv51 model and endoscopists.Results:In the diagnosis of the validation set database,the accuracy,sensitivity,specificity,PPV and NPV of the YOLOv51 model in the diagnosis of superficial esophageal cancer and precancerous lesions on WLI,NBI and LCE were 96.9%,87.9%,98.3%88.8%,98.1%;98.6%,89.3%,99.5%,94.4%,98.2%;93.0%,77.5%,98.0%,92.6%,93.1%,respectively.The diagnostic accuracy of YOLOv51 on NBI was significantly higher than that in WLI(P<0.05),and the accuracy in LCE was lower than that of WLI(P<0.05).In WLI,NBI and LCE,the accuracy of YOLOv51 model is comparable to the diagnosis accuracy of the 2 senior endoscopists(96.9%,98.8%,94.3%and 97.5%,99.6%,91.9%,respectively,P>0.05),but was significantly higher than that of the 2 junior endoscopists(84.7%,92.9%,81.6%and 88.3%,91.9%,81.2%,respectively,P<0.05).Conclusions:The Yolov51 network model constructed in this study has high accuracy in the detection of superficial esophageal cancer and precancerous lesions in WLI,NBI and LCE,which can assist junior endoscopists to improve their diagnostic level and reduce missed diagnosis.The Yolov51 model constructed in this study has high accuracy in the diagnosis of early esophageal cancer and precancerous lesions by endoscopic WLI,NBI and LCE,which is comparable to that of senior endoscopists and better than that of junior endoscopists.This model can assist junior endoscopists to improve their diagnostic level and reduce missed diagnosis.