Realization Of Automatic Human Body Parts Recognition Based On Digital Radiography Scenes

Digital Radiography(DR)is an indispensable imaging technology in the field of medical imaging.In medical examination,it sends out X-ray photons of designated area to the examined part of the patient through the ray source.The rays are attenuated to different degrees through different tissue structures of the human body,and then received by the flat panel detector.The interior of the human body is visualized according to the different degrees of absorption of the rays.DR is widely used because of its low dose,high quality and fast imaging speed.Due to the inaccurate manual operations in traditional DR shooting and the lack of research on automated DR shooting,this thesis is devoted to the research of automatic recognition of human body parts under different protocols in DR scenarios.It aims to study the detection of human parts and keypoints through deep learning.Under different shooting protocols,the machine can automatically and accurately locate the area that needs to be exposed.The ultimate goal is to reduce the radiation,optimize the traditional DR shooting process,and intelligent and automate medical examinations.In this paper,a convolutional neural network is used to build a human basic part detection model and a human keypoint detection model,and a set of result post-processing solutions and a complete detection workflow are designed.Firstly,for the human body part data of different shooting protocols in DR scenarios,the construction method of datasets is proposed.This method uses the outdoor human keypoint dataset and deep learning model,it completes the construction of the human basic parts dataset Hparts Full and the human keypoint dataset Hpoints Full with high quality under a small amount of manual supervision.Secondly,we propose a human body model which combines skeleton and human body outline.It uses the convolutional neural network to build a human part detection model based on Yolov5 and a human keypoint detection model based on Alphapose.At the same time,the improvement of the adaptive anchor frame,Mosaic data enhancement and feature fusion pyramid module effectively improves the detection result of human body parts.Finally,a static key frame detection algorithm based on optical flow method and a postprocessing calculation scheme are designed.The precise positioning of human body parts in different shooting protocols in DR scenarios can be achieved by combining the results of human basic parts and human keypoints.Relevant experimental results show that the whole workflow can achieve the goal of intelligent DR,and it is expected to be further validated and deployed in a clinical setting.