Research On Artificial Intelligence-Based Diagnosis Methods For Gastric Precancerous Lesions And Early Gastric Cancer

Gastric cancer,as the third cancer with both incidence and death in our country,severely threatens people’s live and health.Early gastric cancer has a clear pathological evolution process,which gradually develops and forms from gastric precancerous lesions such as gastritis,intestinal metaplasia,and low-grade intraepithelial neoplasia.The higher the degree of progression of the lesions,the higher the risk of developing gastric cancer.Gastroscopy and gastroscopy biopsy are the gold standards for the clinical diagnosis of gastric cancer.However,the medical resources for high-quality endoscopy in our country are not evenly distributed,and the scope of the experience and diagnostic techniques of the endoscopy physicians is limited to screening early gastric cancer and precancerous lesions.Artificial intelligence(AI)technologies such as deep learning can automatically identify pixel-level features of images and infer complex microscopic imaging structures,which can assist doctors in diagnostic decision-making,alleviate the limitations of different professional levels of endoscopy doctors in our country,and be of great significance for the advancement of clinical diagnosis and treatment of early gastric cancer-related lesions.Magnifying endoscopy with narrow band imaging(ME-NBI)allows the detailed observation of the surface microstructure and morphological characteristics of microvessels of gastric mucosa,has unique advantages in diagnosing gastric mucosal lesions,and is regarded as a major diagnostic tool for gastric lesions.Under the new situation of the flourishing development of national strategies such as healthy China and artificial intelligence,this paper carried out the research of AI-assisted diagnosis system based on deep learning algorithm and ME-NBI images of gastric mucosal lesions,to provide an auxiliary reference for the clinical diagnosis of early gastric cancer and gastric precancerous lesions.This paper established a dataset of ME-NBI images of gastric mucosal lesions.Datasets are the basis of AI-assisted diagnostic system research,this paper collects ME-NBI images and clinicopathological databases from patients from the actual clinical environment.With pathological reports as the gold standard,all pathologists had rich clinical experience.The images were classified into gastritis,intestinal metaplasia,low-grade intraepithelial neoplasia,and early gastric cancer by experienced endoscopists combined with pathological diagnosis reports.All images underwent strict desensitization and screening,and 4685 ME-NBI images were finally included in the study.Accurate identification of early gastric cancer is of great significance to patients.In this paper,we designed an attention-mechanism convolutional neural network model combined with cost-sensitive learning to realize automatic recognition of early gastric cancer-related lesions.We compared it with other advanced classification methods of gastric lesions based on deep learning and published research results.Low-grade intraepithelial neoplasia is gastric cancer’s most serious precancerous lesion and the"critical lesion" closest to early gastric cancer.Therefore,accurately detecting low-grade intraepithelial neoplasia is significant for preventing early gastric cancer.In this paper,we designed an attention-mechanism feature fusion convolutional neural network model.Combined with the attention mechanism of feature fusion technology,the model’s recognition ability of subtle feature differences was enhanced,which enabled the model to maintain high performance under limited parameters and realize the high precision recognition of low-grade intraepithelial neoplasia.Compared with classical convolutional neural networks such as Xception and Inception-ResNetV2(model parameters 22 million and 55.9 million,respectively),the number of two model parameters designed in this paper has been greatly reduced(the number of model parameters is 11.4 million and 14 million,respectively).The lightweight character of the model reduces the difficulty of training and deployment and has a good prospect of practical application.The experimental results showed that,in the identification of early gastric cancer-related lesions,the identification accuracy of the lightweight attention-mechanism convolutional neural network model designed in this paper was 96.8%,93.9%,and 92.9%,respectively,for early gastric cancer,low-grade intraepithelial neoplasia,and non-neoplasm lesions,which was better than other advanced gastric disease classification methods and published research results.In addition,compared with the diagnostic results of experts,the accuracy and specificity of this model are slightly better than that of endoscope experts.and the sensitivity of this model is significantly better than that of endoscope experts.The feature fusion attention-mechanism convolutional neural network model designed in this paper achieved 94.5%accuracy,93.0%sensitivity,and 96.5%specificity in identifying low-grade intraepithelial neoplasia,achieving the most advanced classification performance for low-grade intraepithelial neoplasia.The results demonstrate the effectiveness of this method in ME-NBI image identification of early gastric cancer and precancerous lesions.Intestinal metaplasia has an obvious risk of gastric cancer,and identifying subtypes of intestinal metaplasia is helpful for developing personalized treatment,which is of great significance for the prevention of early gastric cancer.Lesions initially diagnosed as low-grade intraepithelial neoplasia by endoscopic biopsy are at risk of pathological escalation.Due to sampling error of biopsy,some patients with cancerous tissue at initial diagnosis are often misdiagnosed as lesions of lower pathological grade.Accurate identification and judgment of endoscopic images of these high-risk patients at initial diagnosis are conducive to early detection and treatment of lesions.It is of great significance to improve the prognostic benefit of patients.To screen out patients with severe intestinal metaplasia and patients with malignant lesions initially misdiagnosed as low-grade intraepithelial neoplasia,this paper proposes an intelligent screening system for severe intestinal metaplasia and an intelligent prediction system for low-grade intraepithelial neoplasia based on the designed attention mechanism feature fusion convolutional neural network.It recognizes severe intestinal metaplasia and the nature of misdiagnosed malignant lesions under ME-NBI image.The recognition accuracy of severe intestinal metaplasia and misdiagnosed malignant lesions was 95.84%and 94.73%,respectively.The experimental results showed the reliability of the system.Based on the deep learning model and system designed above,this paper designed an AI-assisted diagnosis system for precancerous lesions and early gastric cancer.Based on gastric mucosa ME-NBI images,the system can realize intelligent identification of early gastric cancer-related lesions,high-performance detection of low-grade intraepithelial neoplasia,intelligent screening of severe intestinal metaplasia,and nature prediction of misdiagnosed malignant lesions.It helps detect and personally manage patients with early gastric cancer and gastric precancerous lesions and has high clinical application value.