X-Ray Scattering Correction Based On Deep Learning

Digital X-ray photography(Digitial Radiography,DR)is a conventional X-ray photography technology formed by the combination of computer digital image processing technology and X-ray radiation technology.DR has been recognized by medical personnel for its advantages such as small radiation dose,high sensitivity,high image resolution,and fast processing speed,and has been widely used in the field of medical imaging.However,in the process of DR photography imaging,it will inevitably produce some scattered rays that affect the image imaging quality.Therefore,effectively correcting the scattered rays in the DR image will improve the imaging quality of the DR image and reduce the patient’s radiation dose.In this thesis,we use convolutional neural networks to carry out research on DR image scatter correction,aiming to reduce the influence of scatter signals on the DR image,improve the quality of the DR image after scatter correction,and at the same time reduce the potential harm of radiation to the scanned person.This paper is divided into the following three parts:In this thesis,the Monte Carlo(MC)algorithm is first used to study the simulation of DR images and corresponding scatter images.MC simulation is an algorithm based on statistical probability and simulating a real scene according to a real environment model.First,down-sampling the collected human chest CT data;then removing non-human tissue and dividing human tissue according to an algorithm based on region growth and threshold segmentation;then the ray source and detector are modeled according to the environment of the real DR shooting,and a large number of DR images containing scattering and the corresponding scattering distribution are simulated.This simulation method can use the human chest CT to simulate a large number of scattering DR images and their corresponding scattering distribution maps when the real DR images are lacking,which can be used as a data set for deep learning training and a test set for verification.Convolutional neural networks can directly take image data as input,without the need for manual preprocessing of images and additional feature extraction and other complex operations,can automatically learn the characteristics of images at various levels through convolution and pooling operations,which is more in line with our the process of understanding images.Therefore,the convolutional neural network is used to process a large number of training sets containing scattered DR images and corresponding scattering distributions.After continuous adjusting and participating in optimization,the obtained model is used for post-processing of images.The main idea of the algorithm is to design a two-dimensional convolutional neural network to perform end-to-end image-to-image regression modeling of DR images containing scattering and corresponding scattering distribution images,and use 3×3 convolution kernels to extract and map features in the convolution layer,and then use the ReLu activation function to increase the nonlinear mapping capability of the model,and adjust the parameters during the model training phase to explore the factors that affect the performance of the model to obtain the optimal solution of the model.The training results show that using a convolutional neural network structure,selecting appropriate network depth,width,and convolution kernels,and designing a reasonable loss function will help improve the network’s ability to fit and generalize.The method provided in this paper achieves a better scatter correction effect at the level of simulated data,and can better preserve the human tissue structure in the image.Finally,the C++ programming language is used to implement the software-level algorithms for CT data preprocessing and then apply for software copyright.Patented and applied for the process of scatter correction processing involved in this paper.