Research On CT Metal Artifact Reduction Algorithms Based On Deep Learning

Computed tomography(CT),a non-invasive imaging technology,has significant research and application meaning in industrial detection,medical diagnosis and other fields.In medical CT imaging,the reconstructed image will show a large area of dark bands or radial black and white stripes when metal objects appear in the detection field of view,that is,metal artifacts.It can destroy the tomographic structure of the reconstructed images and reduce the image resolution.To solve the problems of remaining artifacts and inexact reconstruction of tissue structure,we focused on several classical metal artifact correction methods,and proposeed improved algorithms based on deep learning.The main research works include:Firstly,three classic metal artifact correction algorithms were researched and implemented,such as linear interpolation(LI),beam harding correction(BHC),normalization metal artifact reduction(NMAR)algorithms.And pros and cons of the algorithms were analyzed and summarized.Simulation experiments were conducted to test the correction effect of the above algorithms on metal implants with simple structure,and validated the effectiveness of the methods.Secondly,the metal artifact reduction method based on residual encoder-decoder network(RED-CNN-MAR)was proposed to solve the problem that the traditional single metal artifact correction methods could not achieve the desired correction effects.First,the trained RED-CNN network was used to preliminarily repair the metal artifact image.The three channels inputs of RED-CNN were the original image,BHC image and LI image,aiming to integrate the effective information of different images.Then,to further suppress metal artifacts,the tissue processing technology based on threshold segmentation was applied to the RED-CNN output image to obtain a excellent prior image.Finally,the image without metal artifacts can be obtained by using the filtered back projection algorithm.Experiment results demonstrate that the RED-CNN-MAR can efficiently eliminate metal artifacts of water equivalent tissue and restore original structure details.Thirdly,the metal artifact correction method based on improved generative adversarial network(Pix2Pix-MAR)was proposed to solve the problems of image details missing and incomplete elimination of image artifacts after simple residual network.First,the Pix2 Pix network,used RED-CNN as the generator,added the metal artifact-free images as the reference condition to its discriminator,and generated the corrected images under conditional training.Then,to improve the accuracy of discrimination,Patch GAN network based on local image block judgment was applied to discriminate the authenticity of the generated image and the real image.Experiment results demonstrate that the Pix2Pix-MAR can improve the image resolution,remove the remaining metal artifacts at bone-water junction and restore the fine tissue details.