Research On Nano-CT Image Noise Reduction Based On Deep Learning

Computed Tomography(CT)is a novel imaging technique that enables nondestructive observation of the internal structure of samples at nanoscale resolution.The laboratory nano-CT based on ultrafine focus ray source is limited by the power of ray source,and the X-ray intensity is 1 ～ 2 orders of magnitude lower than that of common cone-beam CT.Therefore,at the same exposure time,nano-CT has problems such as increased noise and decreased contrast,which seriously affect the imaging quality.It is of great practical significance to study the noise reduction technology of nano-CT images for improving the imaging quality and work efficiency of nano-CT.Aiming at the problem of nano-CT imaging quality decline in laboratory,this paper studies the three aspects of noise reduction in projection domain,non-matching noise reduction in image domain and three-dimensional noise reduction in image domain from two perspectives: projection domain and image domain.The main research results are as follows:(1)Aiming at the problem that the contrast of projection image details of nano-CT is weak and the existing noise reduction methods have limited ability to retain projection details,a noise reduction method of nano-CT projection domain based on Inception-U-net is proposed.In this paper,Inception module is introduced in the encoder part of U-net network,and multiple convolution kernel parallel operations are used to obtain multi-scale image features,which increases the depth and width of the network and improves the feature transformation ability of the network.In addition,structural similarity loss and L1 loss are introduced to prevent the structural information from being over-smoothed.Simulation and experimental results show that the peak signal-to-noise ratio is improved by 8% on average compared with Residual Encoder-Decoder Convolutional Neural Network(RED-CNN).(2)Aiming at the problem that the large-scale matching data of nano-CT is difficult to obtain and the image domain noise reduction method is difficult to migrate and apply in the field of nano-CT,a non-matching nano-CT image domain noise reduction method based on decoupling characterization network is proposed.In this paper,based on the joint loss function of decoupling network,the image quality of nano-CT is improved by learning the data distribution of different forms of noise-free images generated by cycles.The experimental results show that the network has good image noise reduction ability for non-matching datasets,and the peak signal-to-noise ratio is increased by about 5% on average compared with the Cycle-Consistent Generative Adversarial Networks(Cycle GAN),which are cyclically consistent with non-matching methods.(3)In the process of three-dimensional overall noise reduction of actual samples,a single cross-sectional treatment causes the loss of vector and coronal information and the unsatisfactory overall noise reduction effect.A nano-CT image domain noise reduction method based on 3D Residual Self-Attention U-net(3D RSAU-net)network is proposed.In this paper,by introducing a three-dimensional self-attention module in the network,we make full use of the large regional similarity within and between slices to overcome the problems of limited convolution operation acceptance domain and low efficiency of modeling structural information,and in addition,use multiple residual connections to reduce the number of network parameters.The experimental results show that compared with Domain Progressive 3D Residual Convolution Network(DPResnet),the peak signal-to-noise ratio of this method is increased by 5% on average,and the running time is increased by 50%.