Research On Low-dose CT Denoising Method Based On Generative Adversarial Network

Computer tomography(CT)can quickly obtain the images of human tissue and structure.Those images enjoy a high level of spatial resolution and clarity.Currently,CT has become an important method in clinical diagnosis and played a key role in assisting the clinicians in disease diagnosis and treatment.As CT is widely applied in in hospitals,the damage caused by the resultant radiation of CT cannot be ignored since excessive radiation will harm the health of the subjects.However,reducing CT radiation will produce noises and artifacts in CT images,leading to CT images of poor quality.However,those noises and artifacts often conceal subtle but important details,which will exert a negative influence on the results of clinical diagnosis.In recent years,researchers have proposed various low-dose CT image denoising methods based on deep learning.Those methods are limited in conducting feature extraction from the network structure,which results in blurred detail information after the image denoising process.Besides,in clinical practice,it is difficult to obtain aligned data sets.And in terms of unaligned data sets,supervised methods cannot establish an effective model,while current unsupervised methods cannot effectively remove noise.To solve the problems mentioned above,this thesis proposes two low-dose CT image denoising methods.The main work of this article includes the following 2 aspects:(1)To make the blurred detail information image caused by denoising methods clear,this thesis proposes a low-dose CT image denoising method based on multi-scale feature fusion.It contains a multi-scale feature generator.For the redundant channels in the feature map,it harnesses octave convolution to realize multi-scale transformation and,thus,obtains the ability to accomplish multi-scale feature extraction.In order to generate images similar to the reality,this thesis applies the generator to the generative adversarial network so as to use the discriminator to make the generator to provide CT images of better quality.Experiment results have proved that the multi-scale feature fusion denoising method proposed in this thesis achieves better denoising results and maintains clearer detail information.(2)In clinical practice,it is difficult to obtain aligned CT data set,resulting in the ineffective training of supervised methods.The existing unsupervised methods failed to achieve effective denoising results.For unaligned data sets,this thesis proposes a low-dose CT denoising method based on adaptive feature selection.It uses the CycleGAN network as its basic framework and can denoise the low-dose CT images in an unsupervised way.The adaptive feature selection generator proposed in this thesis can automatically select the feature maps extracted from a variety of network structures,to obtain better quality CT images.It uses the PatchGAN as the discriminator and applies the perceptual loss so that the generated CT images can maintain more detail information.This method has shown good denoising performance on the clinical data sets.