Research And Application Of Colon Polyp Segmentation And Gastric Cancer Recognition In Endoscopy Images

Artificial intelligence technology is developing rapidly,and its application in the field of medical image analysis to assist disease diagnosis has gradually attracted attention.The use of deep learning technology to realize intelligent recognition of colon polyps and gastric cancer has important clinical significance.Aiming at the problems of colon polyp segmentation and gastric cancer detection,this thesis proposes a colon polyp segmentation method combining boundary adjustment module and multiple attention mechanisms and an improved gastric cancer detection method based on SSD(Single Shot MultiBox Detector).The main work and innovation of this thesis are as follows:1.Aiming at the problem of inaccurate segmentation of polyp boundaries due to blurred polyp boundaries under colonoscopy,a cascaded boundary adjustment algorithm is proposed,which uses the reverse attention mechanism and the hard attention mechanism as well as the boundary supervision information introduced by the boundary segmentation branch.Strengthen the relationship between the polyp area and its fuzzy boundary,so that the network pays more attention to the boundary part of the polyp,and corrects the polyp boundary step by step to obtain more accurate boundary segmentation results.In addition,a global self attention mechanism is proposed,which can capture the long-distance region-to-region relationship of the entire endoscopic image.Extensive experiments on multiple public colon polyp data sets have proved that the accuracy of polyp segmentation is superior to other segmentation methods.Through visual analysis,it is found that more accurate polyp boundaries can be segmented.2.Aiming at the problem of insufficient fusion of different semantic feature maps and the existence of semantic gap when SSD detects gastric cancer,a dualbranch feature fusion algorithm is proposed to make up for the problem of semantic gap.In addition,a recurrent feature pyramid networks was proposed to improve the expression ability of FPN(Feature Pyramid Networks)and obtain more robust features to detect gastric cancer.Finally,in view of the lack of sufficient semantic information in the feature map of SSD to detect small targets,a novel small target detection method fused with semantic supervision information is proposed to improve the detection accuracy of small lesions of gastric cancer.The experimental results on the gastric cancer dataset in this paper show that the improved SSD algorithm can improve the mAP metric by 5.9%compared with the original SSD algorithm to reach 56%,and the generalization performance of this method is verified on the natural scene dataset.3.A system that assists endoscopist in the online diagnosis of colon polyps and gastric cancer is realized.Through the system architecture design,the design of various functional modules and the design of the database,the colon polyp segmentation algorithm and the gastric cancer detection algorithm of the previous two chapters are combined to realize the auxiliary colon polyp and gastric cancer diagnosis system,which can help the endoscopist find the lesion area better and faster.