Segmentation Network Of Lesion Region In Gastrointestinal Endoscopy Image Based On Deep Learning

Endoscopy is the main method for the diagnosing gastrointestinal diseases.Chronic gastrointestinal diseases such as gastric ulcer and intestinal polyps have the possibility of canceration.Therefore,it is of great clinical significance to use deep learning algorithm assisting doctors in accurate and rapid diagnosis of gastric ulcer and intestinal polyps which improve the diagnosis efficiency and relieve the pressure of uneven distribution of medical resources.The combination of medical image and computer knowledge has theoretical research value to improve the efficiency and accuracy of diagnosis.This research mainly focuses on the segmentation algorithm of gastric ulcer image based on invasive optical fiber endoscope and intestinal polyp image based on wireless capsule endoscope.The traditional digital image processing method is used to desensitize the dataset and enhance the image quality.The attention module based on hybrid domain is combined with the traditional encoder-decoder segmentation structure to solve the basic problem of the subject.On this basis,the model is improved by the way of multi-scale feature fusion,attention mechanism,pooling method.The model can enhance the ability of global information extraction,improve the abstract modeling ability of long-distance feature dependence,and solve the specific problem of fuzzy boundary of gastric ulcer and intestinal polyp lesions.The main contents of this study are summarized as follows(1)Cooperate with a hospital in Shanghai to complete the production of gastric ulcer endoscopy image dataset.After the data desensitization progressing,we identify the focus area of the endoscopic image under the guidance of professional doctors,and explore the image quality enhancement method of the data set.(2)Based on the encoder-decoder structure,this paper solves the problem of lesion region segmentation of gastric ulcer examination image based on invasive optical fiber endoscope and intestinal polyp examination image based on wireless capsule endoscope.And we explore the impact of multi-scale information fusion module and residual network.(3)Based on the development sequence of time domain,space domain and mixed domain,the development of attention mechanism in visual field is introduced.The attention module based on the similarity calculation of feature vectors is combined with the basic network to explore the dependence of long-distance features of the original image.We explore the influence of similarity calculation method,the number of attention modules and the location of attention module on the model feature modeling ability.The visualization is used to explore how the attention module improves the performance of the model in convolutional neural network,which solves the problem of difficult segmentation of lesion edge region in endoscopic images of gastric ulcer and intestinal polyps.(4)An image segmentation algorithm model is involved and the improved attention mechanism and deformation pooling is proposed.The multi-scale feature fusion method and new training strategy is added,and the long strip pooling is combined with the attention module based on the hybrid domain to explore the ability of the model to capture long-distance spatial dependence.To sum up,the improved encoder-decoder segmentation model proposed in this experiment has achieved excellent performance in the segmentation tasks of gastric ulcer and intestinal polyps.It can more fully understand the global information of the image,provide help for doctors’diagnosis and treatment,reduce the work pressure of medical workers,and has a broad landing scene and rich theoretical research significance.