Research On Anatomical Structure Identification And Focus Region Segmentation Based On Gastrointestinal Endoscopic Images

Magnetic-controlled capsule endoscopy,as a non-invasive gastric examination instrument,generates a large number of images during the examination process.However,the low resolution and unclear anatomical structures of the images make it difficult for doctors to locate the location of the lesions during the diagnostic process,which brings difficulties to secondary diagnosis and treatment.Existing intelligent segmentation methods for gastrointestinal lesions are mostly developed for a single organ and a single type of lesion.A complete gastrointestinal lesion region segmentation system requires the configuration of multiple different segmentation algorithms or models,greatly increasing the computational complexity of the system and the efficiency of auxiliary diagnosis.Therefore,this paper uses deep learning(DL)as a technique to study the algorithm for identifying anatomical locations in magnetic-controlled capsule endoscopic images based on DL technology.Then,based on multi-type gastrointestinal disease endoscopic images,a DL-based algorithm for segmenting multiple types of gastrointestinal diseases throughout the digestive tract is studied.The main research contents of this paper are as follows:(1)A method for classifying the anatomical structure of stomach magnetic controlled capsule endoscope images based on an improved Pyramid Vision Transformer(PVT)network was developed.Based on a pyramid visual transformer network,this method uses a residual structure to add shallow features to deep features,enhancing the network’s feature extraction ability.Next,a YOLOv5 target detection network based on the channel attention module was developed.After the classification and recognition of the upper stomach,lower stomach,and small intestine,the gastric angle and pylorus in the lower stomach were located.This study uses 127014 images from the image enhanced dataset to train and verify the improved PVT network,and the classification accuracy reaches 94.86%,which is superior to other classification networks.Using 9169 image datasets containing gastric horn or pylorus,the YOLOv5 target detection network with channel attention module was trained to achieve an average category accuracy of target detection at a threshold of 50% mAP@0.5 It reaches 0.921,which is superior to other target detection networks.This study will contribute to improving the accuracy of secondary diagnosis and treatment of gastric diseases based on magnetic control capsules.(2)A Transformer based semantic segmentation method for multiple types of lesions in the entire digestive tract was developed.This method uses the PVT model as the backbone network,and adds an improved routing spatial pyramid pooling(ASPP)module and a non local attention module to the original decoder of Polyp-PVT.By enhancing the fusion of shallow and deep features,it enhances the ability to extract global features and adapt to the segmentation of multiple organs and multiple types of lesions.A total of 11149 endoscopic images including early gastric cancer,early esophageal cancer,gastrointestinal polyps,and gastrointestinal ulcers were used for training and validation.The average similarity coefficient Dice reached 87.8%,which is superior to other segmentation networks.This study will contribute to improving the efficiency and accuracy of the total digestive assistance diagnostic system.The designed algorithm for the identification of anatomical structures in magnetic-controlled capsule endoscopy images and the algorithm for lesion segmentation in full gastrointestinal endoscopic images in this article both exhibit excellent performance.They can provide reliable and objective basis for doctors and have certain clinical application value.