Research On Methods For Metal Artifacts Removal In CT Image Based On Convolutional Neural Network

In computed tomography(CT)images,if a patient has implanted a metal implant in the mandible(mainly the teeth in this paper),including dental implants,orthopedic implants or other metal implants,because the attenuation coefficient of the internal tissue of the human body is far lower than the embedded metal objects,X-ray will be penetrated into the metal inserts severe attenuation.This leads to some differences between the reconstructed image and the real object,not only blurring the important diagnosis information of the surrounding tissue,but also making the metal artifacts in the CT image.Because of the existence of metal inserts,metal artifacts have been affecting the effect of CT.Therefore,it is not only of clinical significance,but also of great importance for medical diagnosis and analysis to study the correction algorithm of metal artifacts in mandible CT imagesThere are many kinds of traditional metal artifact correction methods,but they have their own limitations in clinical application.For example,most of the interpolation-based methods need image segmentation.For CT images with serious artifacts,it is difficult to segment the image.For example,the iterative based method,the calculation cost is very high.In recent years,in the field of computer vision,the development of deep learning is very rapid,in which convolutional neural network is more and more widely used,which also brings new research direction to scholars in the field of metal artifact correctionIn order to reduce the metal artifacts in the mandible region of CT images,a new algorithm based on deep convolution neural network(CNN)is proposed.The main work is as followsFirstly,a database of mandible CT images including metal artifacts and without metal artifacts was established.Before CNN training,the deformable image registration method is implemented to preprocess the image.Therefore,paired data with and without metal artifacts can be obtained from the datasetSecondly,we build a simple 17-layer CNN architecture to learn metal artifacts,and use GPU to accelerate the speed of training data,improve the learning efficiency of the network.At the same time,the experimental results show that the method has a good ability of metal artifacts correction,PSNR and SSIM values also show significant improvementFinally,based on the convolutional neural network,the concept of residual learning is introduced,and a convolutional neural network model based on residual learning is proposed.Based on the idea of residual learning,the improved neural network model can not only learn the artifact image independently,but also reduce the influence of the artifact part on the whole image in the learning process.The experimental results show that in most cases,the better PSNR value can be obtained from the CT image of mandible corrected by convolution neural network based on residual learning.