Research And Implementation Of Real-time Location Algorithm Of Dynamic Colonoscopy Polyp Image

Colorectal cancer is one of the most frequent cancers in China,which is closely related to colorectal polyps.Colorectal polyp is a common intestinal disease.Colonoscopy is the most effective and safe method.However,due to its high intensity of work,the results of physicians' visual inspection are very dependent on their own experience,and the detection rate of polyps is also affected by factors such as doctors' fatigue,so there is a certain degree of polyp missed detection.In view of the shortcomings of colonoscopy,especially with the emergence of deep learning algorithm,the computer-aided diagnosis of colonoscopy is becoming a reality.This method can detect and prompt doctors to pay attention to the polyps that may be neglected in real time,so as to improve the detection rate.Therefore,this project takes the colonoscopy polyp as the research object,aiming to develop a set of portable colonoscopy polyp real-time positioning system based on deep learning.The main work of this paper is as follows:(1)This paper studies the performance of three popular target detection algorithms based on convolutional neural network,Faster RCNN,SSD,yolov3 in colonoscopy polyp image,and evaluates the accuracy,specificity and sensitivity through experiments,and analyzes the performance differences of the three networks in colonoscopy polyp image detection.The experimental results show that SSD has the fastest detection speed,but the performance of small target detection is poor.(2)In this paper,the idea of SSD multi-scale prediction is adopted,and a high-precision I?SSD detection algorithm is designed.On the one hand,a variety of inception modules are used instead of SSD basic features to extract the traditional convolution layer in the network structure,so as to enhance the depth feature extraction ability of the network;on the other hand,adjust the scale of anchor frame in SSD and optimize the number of anchor generation according to the shape characteristics of polyp.The experimental results show that the improved algorithm improves the detection accuracy significantly,and the detection speed is 31 fps,which can better meet the needs of real-time detection.(3)A real-time positioning system of colonoscopy polyp image is designed based on I?SSD,which is realized by Jetson TX2 embedded platform.The input signal is obtained from the Olympus CV-290 electronic endoscopy detection equipment in the hospital,connected to the Jetson TX2 development board through the video format conversion card,transplanted the trained I?SSD network model to the Jetson TX2 embedded platform,and realized the portable real-time detection and tracking display of enteroscopy polyp image through the system development and debugging.After experimental test,the developed image positioning system of colonoscopy polyp based on I?SSD model can detect the image of colonoscopy polyp in real time,and the accuracy of the system reaches 96%.It realizes the accurate positioning of colonoscopy polyp,which can provide reference for further diagnosis of doctors.