Research Of Low-Dose CT Image Denoising Algorithms Based On Deep Learning

Low-dose X-ray computed tomography imaging technology is one of the important methods of modern medical detection,and it is also one of the key research contents of medical imaging.Its purpose is to ensure the accuracy of CT imaging and the reliability of medical diagnosis on the premise of reducing X-ray radiation dose.There are many low-dose CT imaging methods,either reduce the radiation intensity of X-rays or reduce the sampling during the scan.No matter which methods could reduce the harm of X-rays to the human body,but at the same time,they all damage the image quality to varying degrees,In order to reduce the radiation dose as much as possible while ensuring the quality of CT imaging,low-dose CT imaging technology has been paid more and more attention by researchers.In this thesis,the problem of low-dose CT denoising is analyzed and studied,and the solutions are put forward based on the existing problems.The main contributions are as follows:(1)In this thesis,the background of low-dose CT imaging is firstly described,and the development status of low-dose CT imaging technology is discussed in depth.The advantages and disadvantages of the current mainstream CT image denoising algorithms,such as sinogram filtration before reconstruction,iterative reconstruction algorithm,CT image postprocessing after reconstruction are analyzed,and the existing problems in the current CT denoising field are summarized and analyzed.(2)Then the convolution neural network is introduced,which lays a theoretical foundation for the follow-up depth learning-based denoising of CT images.The low-dose CT imaging theory is analyzed,and the defects of the current noise modeling scheme are discussed.The CT image noise is modeled based on the deep learning theory,and the feasibility of applying depth learning to CT image denoising is analyzed.(3)A semi-supervised learning method is proposed to solve the problem of difficulty in acquiring paired training CT data.Traditional low-dose CT denoising algorithms are mostly based on noise models,and the imaging effect is general.Nowadays,the models based on CNN have become one of the development priorities.Its powerful feature extraction capabilities provide another solution for low-dose CT denoising.However,CNN needs huge paired data support,and it is difficult to obtain paired CT images.To solve this problem,this thesis proposes a semi-supervised denoising method based on dual learning,which introduces unpaired data in the model training process and reduces the data requirements of the model,and the effect of unpaired data ratio on denoising is analyzed by experiments(4)A GAN-based algorithm is proposed to solve the imaging quality problem of low-dose CT denoising.Most of the current algorithms focus on the improvement of the image evaluation index,ignoring that the image quality is based on human visual effects to a certain extent.Simply improving the evaluation index does not mean that the image quality is getting better and better.In the process of medical diagnosis,radiologists often pay attention to some details of the image,which often determines the accuracy of the diagnosis results.Aiming at the problem of detail information loss in the imaging process,this thesis proposes a generative adversarial network-based on the attention mechanism,which greatly enhances the detail expressive force of images.Compared with existing algorithms,it fully verifies its effectiveness.