Research On Noise Suppression Method Of Cone Beam Reconstruction

Owing to shorter scanning time and less radiation injury,low-dose CT(LDCT)is widely used in clinical diagnosis and micro-CT imaging.The reduction in X-ray radiation dose leads to a decrease in numbers of X-ray photons to detector,which may degrade the image quality both in signal-to-noise ratio and resolution.Many conventional methods have been proposed to reduce the image degradation caused by low-dose noise.Due to the complexity of model and algorithm,it is hard to make a trade-off between image quality and denoising performance.Deep learning method is able to learn the inherent relevance between samples,which provides a new tool for nonlinear complex modeling.In this paper,the noise suppression method of micro-CT imaging has been studied based on deep learning techniques.The main contents are organized as follows:Firstly,the sources of noise in low-dose CT are introduced and the distributions of noise are analyzed.The expression of noise variance in CT image is derived referring to FBP algorithm,and a conclusion has been drawn that noise in reconstruction domain is non-uniform and unstationary.Based on the self-developed micro-CT device,an approach for establishing the denoising label dataset is proposed by step-by-step CT scanning.Specifically,each projection image is acquired multi-times at each angle and then the label image is achieved by averaging them.In this way,the pixel-wised accuracy can be secured in a pair of low-dose image and label image.On account of the dataset prepared above,a projection domain denoising method based on C-GAN(Conditional Generative Adversarial Nets)is proposed.In generative network,residual learning is used to map the noisy projection to the corresponding noise distritution.Dilated convolution and skip connection are applied to generate better performance.In discriminant network,the denoised image is coupled with the noisy projection as a condition,and are imported into the Patch-GAN discriminator to identify the authenticity of the denoised image.In addition,variance loss is added to the traditional GAN loss to make the noise map more reliable.The results of simulation and real data demonstrate that the proposed method can reduce low-dose noise effectively.Furthermore,by stacking the projection images under the adjacent angles,a 3D-convolution denoising network is constructed.Since the mutual features among adjacent angles is exploited,more high-frequency structure is preserved and less blurring occurs.To prevent missing image information from projection processing,an image domain denoising network called APCNet is proposed.Low-dose image and its edge features extracted by the sobel operator are concatenated into the network so as to avoid the degradation of image structures.Inspired by Inception module and ACNet,the convolution layer is replaced by three different kernels: N×N,N×1,1×N,to acquire multi-scale information and capture more detailed structures.In addition,learning from conventional iterative denoising algorithms,a fidelity term related to the noisy images is added to loss function and a smooth coefficient is utilized to control the degree of denoising.From the results of simulation and real data,it can be seen that the network precision is improved and APCNet presents superior performance than other denosing networks.Considering the tridimensional nature of CT imaging,a 3D reconstruction domain denoising network(3D-RDNet)is finally proposed by stacking CT images of adjacent layers.Due to the hardware limitations,the structure of 3D-RDNet is simplified.Although the denosing performance is inferior to APCNet,it is proved that 3D denoising network is effective,which holds potential in the future research.