Analysis Of Laparoscopic And Capsule Endoscopic Video Images Based On Deep Learning

Gastrointestinal endoscopy is the most important tool for the diagnosis and treatment of digestive diseases.However,the clinical diagnosis of digestive diseases suffers from large manual readings,strong subjectivity,and there is difficulty in ensuring the accuracy of diagnosis.In recent years,with the development of deep learning,there are more and more works that combine artificial intelligence with endoscopy,which not only helps digestive endoscope diagnosis and treatment get more accurate and objective results,but also greatly reduces the burden on doctors.In this thesis,the research goal is to explore the fundamental applications of deep learning in the digestive endoscope diagnosis and treatment.There are three sub-tasks studied in detail:laparoscopic surgical instrument segmentation,laparoscopic surgical video prediction and localization of digestive organs in capsule endoscopy.The main work of this research includes the following aspects:1.A U-shaped fully convolutional neural network is proposed for the task of laparoscopic surgical instrument segmentation.In this model,parallel convolutional branches are introduced during downsampling,and the temporal attention mechanism is embedded during upsampling.The correlation between adjacent video images provides prior information of the surgical instruments for input images,which improves the performance of this segmentation model.2.To solve the problem of laparoscopic surgery video prediction,a convolutional neural network incorporating optical flow estimation,the adaptive kernel,and the vector method is designed to predict the next frame of the input laparoscopic surgery frames.A U-shaped network performing kernel function prediction and an auxiliary loss branch are used to get higher quality for predicted images.3.For the localization of digestive organs in capsule endoscopy,a model combining the binary search algorithm and a CNN classification network is proposed.With the premise that video images can be mapped into a sorted array according to their labels,the model using the binary search for image selection and the CNN model for classifying selected images performs organ localization quickly and accurately.Methodological comparative experiments for the proposed models are conducted in this thesis,and the experimental results show the effectiveness and robustness of these models.These automatic analysis algorithms proposed in this thesis provide references for applications of deep learning in the field of digestive endoscopy,which have certain clinical value.