Research On Medical Diagnosis And Treatment Task Oriented Low Dose CT Image Artifact Suppression

With the continuous development of computed tomography technology(CT),CT has been widely used in medical diagnosis and treatment tasks,playing a crucial role in the diagnosis and screening of diseases.However,the increase in X-ray radiation dose has caused public concern about the health of patients.The reduction of radiation dose will also cause a large amount of noise in the reconstructed CT image,affecting the accuracy of subsequent doctors’ diagnosis.Therefore,how to ensure that the quality of CT images can be improved while minimizing radiation dose has important research significance and application value at this stage.In recent years,deep learning has been widely applied in the field of LDCT noise reduction.Scholars have made improvements from different perspectives,improving the performance of networks,and providing important ideas for the field of LDCT noise reduction.However,most methods currently face issues such as blurred edges and loss of details in denoising results.In response to the above issues,this article mainly focuses on how to balance the generation ability of the generator and the discrimination ability of the discriminator for generating adversarial networks,as well as how to effectively separate the artifact noise in LDCT from normal organizational structure information.The main work is as follows:(1)In order to improve the separability and sensitivity of different Semantic information in LDCT images,a low-dose CT image denoising network based on dual domain GAN was proposed.Artifact noise belongs to the high-frequency information in CT images,and the direct removal effect in the image domain is not good.Therefore,we have designed a pyramid generator based on DCT transformation,which utilizes the different semantic components in the transformation domain to represent different forms.By designing the generator in different domains,we aim to achieve accurate estimation of the artifacts and noise contained in LDCT images as much as possible.Secondly,a dual domain discriminator was designed to more accurately distinguish the generated denoising results in different domains by designing discriminators in the image domain and DCT transform domain.Compared with current mainstream algorithms,this algorithm has achieved good denoising performance in both quantitative and qualitative evaluations.(2)The GAN based LDCT denoising network utilizes an interactive training method of generator and discriminator to suppress artifact noise in images.The discriminator plays a supervisory role in the denoised image results and plays a key role in ensuring the denoising performance of the GAN network.However,the discriminators of existing GAN networks directly classify the output results of the generator,and the design of discriminators only focuses on the overall quality of the generated image.This overly simplified discriminator structure design greatly limits the denoising performance of GAN networks.This article proposes a multi domain discriminator GAN for noise reduction in LDCT,which not only extracts discriminative features for the overall quality of the output image;Moreover,gradient domain discriminant features are supplemented,with a focus on the rich edge details of the tissue structure and lesion area in the image;And added ROI region discriminant features,guided by anatomical prior information,to select special structural areas that are more meaningful for downstream medical diagnosis and treatment tasks for discriminant feature extraction.Due to the extraction of discriminator features in multiple different domains,the discriminant ability of the GAN network proposed in this paper has been enhanced,improving the effectiveness of the discriminator against the generator,thereby greatly improving the noise reduction ability of the network.In addition,in conjunction with the structure of the multi discriminator GAN network,this paper also designed corresponding cost functions to conduct supervised training on the multi domain feature extraction of the discriminator,ensuring the effectiveness of the discriminator in extracting specific domain features.The experimental results show that the proposed network model has significant performance advantages compared to the latest LDCT denoising network.