Intelligent Identification Of Intestinal Polyps By Using Capsule Endoscopy

Intestinal polyps have a high probability of worsening into malignant tumors.Therefore,an effective screening of intestinal polyps is very important in the early screening of intestinal cancer.Capsule endoscopy is a common clinical method for intestinal polyps screening.However,there are still some limitations in the existing methods for identifying intestinal polyps based on capsule endoscopy: 1)a lot of time and labor cost are contributed to label production for supervised learning;2)a large amount of original data is directly used to train the intelligent identifying model of intestinal polyps,which leads to long training time and difficult convergence of the model;3)individual differences of patients tend to cause weak generalization ability of the model.In order to improve those limitations,this thesis takes the intestinal capsule endoscopy image as the research object,uses the algorithm of image feature extraction,unsupervised clustering algorithm,target detection and segmentation algorithm as the theoretical method,to carry out the intestinal polyps intelligent identification method based on the capsule endoscopy.This thesis also incorporates the multi-center clinical data,discusses the effectiveness and superiority of the proposed intelligent identification method.Firstly,an improved 2D image entropy based on the distance discriminant ratio(2DMIE-DR)is designed to extract the information feature from the filtered endoscopic image data of capsule.Each image is extracted as a one-dimensional feature vector to reduce the data dimension.On this basis,an improved unsupervised clustering algorithm based on Mahaobanobis distance(MK-means)is proposed to perform unsupervised clustering learning on the extracted feature vectors,so as to extract the image frames of suspected included intestinal polyps,which could be used by physicians for label making.Finally,a mask region-based convolutional network based on the precise region of interest pooling(PRMRCNN)is studied and proposed to conduct a supervised learning on the pictures of suspected included intestinal polyps.This step obtains the individual intelligent identifying model of the patient so as to realize the intelligent identifying of intestinal polyps.In this thesis,experiments are designed for the proposed 2DMIE-DR,MK-Means and PRMRCNN respectively to demonstrate the effectiveness and superiority of the three algorithms.In addition,this thesis combines with multi-center clinical data to verify the effectiveness and applicability of the key models and technologies involved in intelligent identification of intestinal polyps.The experimental results show that the proposed intelligent identification method can effectively identify intestinal polyps from multi-center clinical data.Compared with the existing identification methods,the intelligent identification method proposed in this thesis improves the detection accuracy of polyps from 44.83% to 46.88%,and the segmentation accuracy of polyps from47.16% to 50.80%,significantly reduces the workload of physicians(79.46%).Therefore,the proposed method has clinical application potential to a certain extent.