Research On Low-Dose CT Image Denoising Based On Multi-level Feature Fusion And Dual Attention Mechanism

X-ray Computer Tomography(CT)technology has the advantages of fast scanning time and clear images.Therefore,it is widely used in various fields such as medical treatment and industry.It is especially used as a common clinical examination method for follow-up diagnosis and treatment provide an intuitive reference and basis for judgment,which has greatly promoted the development of modern medicine.While the application of CT technology is more and more common,the hidden dangers of X-ray radiation in the scanning process have attracted widespread attention.Studies have shown that long-term exposure to X-ray radiation may cause serious consequences such as cancer or leukemia.Therefore,it is necessary to reduce the radiation dose during CT scanning.The most common way to reduce radiation dose in clinical practice is to reduce the X-ray tube current.However,this method will cause a lot of noise and artifacts in the reconstructed CT image,which will affect the accuracy and referenceability of the image.Misjudgments caused by image noise in diagnosis may lead to serious consequences.Therefore,how to obtain high-quality images under the premise of low-dose CT scanning has attracted the attention of many researchers.In recent years,due to the strong ability of feature representation based on deep learning,it has gradually replaced the traditional image denoising algorithm and become the mainstream research direction.The Encoder-Decoder structure can map the input image of low-dose CT to the output image of the same size,and the Generative Adversarial Networks can better learn the real data distribution.Therefore,the network framework based on the combination of Encoder-Decoder structure and Generative Adversarial Networks is widely used in the task of low-dose CT image denoising.Based on the analysis and thinking of the existing network framework,this thesis proposes the following two improved models to further improve the performance of low-dose CT image denoising.The main contributions are as follows:(1)Multi-scale Hierarchy Feature Fusion Generative Adversarial Network(MHFF-GAN).Aiming at the wide-ranging semantic gap between the low-level features and high-level features in the Encoder-Decoder structure,we propose a Multi-Scale Dilated Block(MSDB)to fuse the low-level feature maps with the high-level feature maps through this module,and retain the low-level feature space information.At the same time,the semantic information is enhanced.In addition,in view of the different degrees of semantic gaps at different levels,we adaptively assign different dilated rate to MSDB modules at different levels.The experimental results show that our method has obvious advantages over the contrast method,which proves the effectiveness of our proposed method in the task of low-dose CT image denoising.(2)Dual Attention Mechanism Generative Adversarial Network(DAM-GAN).Firstly,the preliminary feature correction is achieved by learning the dependency between the channels in the feature map of Encoder,and then the corrected feature map is merged with the feature map of Decoder.Secondly,the location attention mechanism is used to highlight important areas to further modify the feature map.By repeating the above operations at different levels,the feature map is corrected layer by layer,which promotes the network to output high-quality CT denoising images.The experimental results prove that our method has good denoising performance.