Research On Endoscopic Diagnosis Of Chronic Atrophic Gastritis Based On Deep Learning

With the advent of the era of big data,artificial intelligence technology represented by deep learning has been widely applied in the medical field and achieved significant results in imaging and pathology.It has been successfully applied in the intelligent assisted diagnosis of gastrointestinal benign lesions,precancerous lesions,early gastric cancer and other diseases,which can effectively improve the detection rate of diseases and reduce misdiagnosis and missed diagnosis.In clinical medical diagnosis,gastroscopy is a necessary means for medical personnel to detect patients’ stomach diseases.Hospitals have accumulated a large number of gastroscopy images,which provides abundant data for computer-aided diagnosis of stomach diseases.Chronic atrophic gastritis is a common stomach disease,a precancerous lesion of gastric cancer,and has a certain correlation with the occurrence of gastric cancer.Using intelligent diagnosis of chronic atrophic gastritis,early drug treatment or other methods can be carried out to effectively prevent the formation of early gastric cancer.Therefore,timely diagnosis and treatment of chronic atrophic gastritis is particularly important.In the intelligent identification of chronic atrophic gastritis,atrophy lesions are subtle and detailed features are particularly important.However,previous methods mostly directly downsampled high-resolution images,losing texture and detail information in high-resolution images,while lacking recognition and classification of atrophic sites,and the diagnosis of the severity of atrophy,which hinders the opportunity for early intervention and timely treatment.In response to the above issues,this thesis conducts research on the identification and classification of chronic atrophic gastritis,and proposes two intelligent diagnosis methods for chronic atrophic gastritis.The main innovative work is as follows:(1)Current intelligent diagnosis methods for chronic atrophic gastritis endoscopic images usually downsampled high-resolution images to lower resolutions(generally less than 512 × 512 pixels)for intelligent diagnosis.However,this approach leads to the loss of most texture details in the images,resulting in a significant decrease in the recognition rate of chronic atrophic gastritis lesions and even the inability to recognize them.To address this issue,we designed an intelligent recognition method for chronic atrophic gastritis based on high-resolution endoscopic images.First,a target detection method is used to preliminarily locate the suspected lesion area in the chronic atrophic gastritis image,which is then mapped to the high-resolution image for cropping,obtaining a highresolution image of the suspected lesion area and filtering out obvious non-atrophic areas.Second,the high-resolution images of different sizes are divided into several patch blocks,and the lesion details in the high-resolution patch blocks are learned to achieve discriminative classification for individual patch blocks and combinations of multiple patch blocks.(2)Currently,there is limited research on the level classification of chronic atrophic gastritis,and there is a lack of research methods that combine atrophy recognition and location classification to complete the diagnosis,resulting in patients with different degrees of atrophy unable to receive more targeted treatment.To address this issue,a deep learning-based method for classifying the level of chronic atrophic gastritis was designed.The two-stage method first uses the previously proposed intelligent recognition method for atrophic gastritis as the initial lesion recognition network to complete the recognition of atrophic lesions.Then,a parallel dual-branch architecture combining convolution and Transformer is used as the location recognition network to learn the morphology and contour features of different parts of the stomach,determine the location of the atrophic image,and finally obtain the level classification results based on the parts and scope involved in the atrophy.(3)A chronic atrophic gastritis intelligent recognition system has been designed.Users can directly run the packaged executable file on their local computer to complete the recognition of atrophic lesions and grade classification.This localized approach avoids the need for users to rely on an internet connection,making it more convenient and faster to use.The system provides diagnostic reference data for doctors,improving their diagnostic efficiency.(4)A precisely annotated endoscopic image dataset for gastric atrophy was created.Currently,there are few publicly available endoscopic image datasets for the stomach,and this study obtained the training dataset primarily through self-collection.We collected1,272 gastric images including atrophic and normal cases from the hospital.Additionally,we collected 2,196 gastric images of different locations,including the cardia,gastric fundus,pylorus,angular notch,greater curvature of the corpus,and greater curvature of the antrum,according to the order of endoscopic examination.