Research On Low Dose CT Image Processing With Controllable Network Learning

Computed Tomography(CT)is a medical imaging technique that uses X-rays to produce cross-sectional images of the body’s internal structures.However,the accumulated radiation exposure from CT imaging may increase the risk of cancer in patients.To reduce this risk,low-dose CT(LDCT)scans are used,but they often result in noisy and blurry images,which can affect the accuracy of diagnosis.Researchers and radiologists have therefore developed LDCT image processing methods to improve image quality and reduce radiation exposure for patients.This thesis focuses on post-processing deep learning methods,using convolutional neural networks to model and fit the complex problem of low-dose noise and artifact prediction,with the aim of providing effective assistance to clinical diagnosis.The thesis provides a detailed discussion of the theoretical knowledge related to LDCT processing,introduces efficient image denoising methods,and analyzes and summarizes several classic LDCT processing algorithms.The performance of these algorithms is evaluated through experimental simulations,and the key issues to be considered in constructing the LDCT processing algorithm model are summarized.The main contributions and research innovations of this thesis are as follows.(1)The levels of artifacts and noise in LDCT images vary greatly depending on the specific body part being imaged,and traditional neural network methods cannot effectively handle the different levels of artifacts and noise present in LDCT images obtained at the same dose.To address this issue,this study proposes a controllable conditional residual network based on the basic residual network architecture.The basic residual network has significant advantages in LDCT image processing,but its identity mapping form is limited.By adding control weight values to the conventional residual connections,a controllable residual module is designed to adapt to features with different intensities and forms.By combining the input and output layers with the controllable residual module,a controllable conditional residual network is established for LDCT image processing to estimate high-quality CT images.This part of the study aims to initially suppress artifacts and noise present in LDCT images at different levels,while preserving the original image details as much as possible.The obtained network model can be applied to the processing of LDCT images with complex degradation processes.(2)In order to obtain more detailed texture and complete organizational structure information,on the basis of conventional U-shaped network(UNet),the residual jump connection will be added.The UNet composed of encoder and decoder will be improved to achieve residual multi-scale feature learning,and the extraction and representation of small-scale deep features and large-scale shallow features will be established.In the design of conditional network,combining the characteristics of noise and artifacts,a better conditional network will be used to accurately estimate the feature control weight,and correct the use of different scale feature maps in UNet to achieve the adjustable function of features,so as to obtain better noise and artifact removal performance,maximize the retention of details such as texture organization,and improve the efficiency of algorithm operation.The results are analyzed and summarized to adjust the hyperparametric optimization network model and enhance the ability of the model to maintain details to achieve the expected research objectives.(3)In view of the problem that convolution neural network can not deal with the detail texture of CT images well,and there are still some differences between it and the real NDCT images,this thesis introduces a network based on adjustable network in the GAN framework,which can improve the complete imaging from LDCT projection to LDCT.In the generator network,a combined network composed of basic subnetwork and conditional subnetwork is used to improve the quality of LDCT image based on.In the discriminator,Modified Vgg-128 structure is used to improve the generalization ability of the generator and can well suppress the noise artifacts in various parts of the LDCT image.In addition,in order to improve the denoising ability and retain image details as much as possible,the constructed denoising network constructs a composite loss function based on three loss measures: counter loss,MSSIM loss and gradient penalty.Compared with other most advanced algorithms,this method has superior denoising effect both qualitatively and quantitatively.Finally,the challenges that the LDCT image processing algorithms still need to face at the present stage are presented,and the future research trends of LDCT image processing are discussed in view of these problems.