| Introduction Gastric cancer is the fifth most common cancer in the world and the fourth leading cause of cancer-related death.The incidence and mortality of gastric cancer in China are also in the forefront.Because the early symptoms of gastric cancer are not obvious,it often progresses to an advanced stage at the time of diagnosis,and the mortality rate of advanced gastric cancer is high and the 5-year survival rate is low.Therefore,gastric cancer screening is very important,and effective intervention treatment after early detection can effectively prevent the progression of gastric cancer to the late stage and significantly prolong the 5-year survival rate.Endoscopy is an important means of clinical diagnosis and treatment of gastric cancer,advanced endoscopic imaging is helpful for lesion detection and characterization,can visualize the gastric mucosa and blood vessels in detail,with higher detection rate and more accurate mucosal lesion characteristics.Microendoscopic imaging enables microscopic cell imaging at the microscopic level,enabling optical biopsy,helping to guide endoscopic biopsy and reducing sampling errors generated by random biopsies.Endoscopic imaging requires high-quality images for observation and diagnosis,which is susceptible to interference from many factors and blurry images.On the other hand,the diagnosis of imaging results is highly dependent on the experience of endoscopists,which is prone to significant differences.The rapid development of artificial intelligence has provided new opportunities for medical image processing,recognition and auxiliary diagnosis.Materials and Methods In this study,based on the multispectral fluorescence endomicroscopy(MFE)device built by our group,we collected clinical gastric tissue biopsy and surgical specimens,imaged them with MFE,labeled them by histopathology,constructed relevant MFE image datasets,and formulated diagnostic models for early gastric cancer and precancerous lesions.In this study,different image processing methods,including spatial domain processing method,frequency domain processing method and deep learning-based image denoising method,are used to demesh the image by optical fiber demeshing,and the image quality is improved by the image enhancement algorithm.At the same time,ResNet and ViT classification algorithms were used to classify normal gastric mucosa,low-grade intraepithelial neoplasia,high-grade intraepithelial neoplasia,and gastric cancer,and verify their performance.On this basis,the intelligent diagnosis system is constructed and evaluated and studied,and the diagnostic accuracy and diagnostic efficiency of the intelligent diagnosis system and doctors are compared.Result In this study,more than 1000 cases of normal mucosa of the digestive tract and biopsy tissues or surgical specimens of different lesions were collected endoscopic images and pathological results,and they were imaged by the MFE equipment constructed in the early stage,and an MFE image database with a data volume of 500 G was successfully established.It included 600 images of 38 cases of normal gastric mucosa,620 images of25 cases of low-grade intraepithelial neoplasia,630 images of 25 cases of high-grade intraepithelial neoplasia,and 580 images of 25 cases of gastric cancer.Combined with the MFE-based diagnostic criteria for normal gastric mucosa,gastric mucosal intestinal metaplasia,low-grade and high-grade intraepithelial tumor and gastric cancer based on MFE,image features were annotated,and the training set database required for artificial intelligence diagnosis was established.At the same time,this study completes optical fiber demeshing through different image processing methods,among which the traditional spatial processing method and notch filtering for MFE image design can achieve better demeshing effect than the deep learning algorithm.In the study of image quality improvement,the gamma value is 40,and the contrast adjustment is 100 to better improve the contrast of the image with grid and the image after grid removal,and the image imaging quality is the best.On the other hand,the ResNet classification algorithm and ViT classification algorithm were used to distinguish between normal gastric mucosa,low-grade intraepithelial neoplasia of gastric mucosa,high-grade intraepithelial neoplasia of gastric mucosa,and gastric cancer,and the classification accuracy was about 74-80% and 81-85%,respectively.From the perspective of execution efficiency,the execution speed of the ResNet classification method is higher than that of the ViT algorithm,and under the experimental conditions of NVIDIA 3090 GPU,the efficiency of the two algorithms is 32 fps and 14 fps,respectively.Based on this study,an intelligent diagnostic system was further constructed to realize intelligent auxiliary diagnosis,which achieved 89.3%,74.3%,78.2% and 91.3% of the diagnostic accuracy of gastric mucosal intestinal metaplasia,low-grade intraepithelial neoplasia,high-grade intraepithelial neoplasia and gastric cancer,respectively,with an average diagnosis time of 1.3 minutes.This result is compared to the accuracy and efficiency of the doctor’s diagnosis,which is lower than that of the doctor,but more efficient in diagnosis.Conclusion This study confirms that the image processing optimization combined with MFE equipment and the construction of artificial intelligence diagnosis system can assist in the diagnosis of early gastric cancer,which can identify precancerous lesions at different stages,and has clinical application potential. |
PDF Full Text Request