Research On Super-resolution CT Image Reconstruction Algorithm Based On Deep Learning

Compared with the traditional super-resolution reconstruction of CT images based on interpolation,reconstruction and hardware equipment,the method based on deep learning can obtain reconstructed images with good texture continuity and prominent features of the region of interest,and the cost of the algorithm is lower.Therefore,based on the deep learning image super-resolution reconstruction algorithm,this paper mainly analyzes and studies attention mechanism,residual feature extraction and image up-sampling and down-sampling techniques.The specific content is as follows:(1)Super-resolution CT image reconstruction based on UNet feature fusion(UNet SR)is proposed.Firstly,regarding channel attention(CA)and efficient channel attention(ECA),it only emphasizes the interaction of information between channels,and cannot express the limitations of the weight features of the channels themselves.This paper proposes channel learning attention(CLA)for improvement,and embeds CLA into a cascaded residual network to construct a cascaded residual channel learning attention block(CRCLAB)to mine the deep information of images.Secondly,when down-sampling feature extraction is performed for the single-scale down-sampling extraction block(SDEB)in UNet,a lot of image information will be lost due to the size of the receptive field of the convolution kernel,lead to the undesirable effects of sudden changes in the texture of the reconstructed image and blurred features.This paper proposes a Multi-scale down-sampling extraction block(MDEB)to optimize the shallow features of images by increasing the width of the network model.Then,CRCLAB and MDEB are used to extract the shallow and deep features of the image.Finally,the bilinear interpolation method is used to fuse the multi-scale features output by UNet to complete the image super-resolution reconstruction.Compared with other algorithms,the UNet SR algorithm improves the quality of the reconstructed images to a certain extent,and the parameters of the network model and the reconstruction time of the algorithm are both improved.(2)In view of the dual regression network(DRN)by introducing dual loss to constrain the solution space of the image super-resolution reconstruction mapping function,thereby improving the efficiency of image reconstruction quality.Therefore,this paper further builds on DRN and proposes residual attention aggregation dual regression network super-resolution CT reconstruction(RAADRNet).Firstly,the Spatial Attention(SA)is improved to Spatial Learning Attention(SLA),and the Spatial Feature Fusion Block(SFFB)formed by SLA is introduced into the CRCLAB network to construct residuals Attention aggregation module(Residual attention aggregation block,RAAB)for the extraction of deep image features.Secondly,in view of the fact that multi-scale down-sampling extraction of MDEB cannot strengthen the channel information and spatial features in the process of down-sampling feature extraction,this paper further introduces the attention mechanism CLA and SLA into MDEB to form a multi-feature down-sampling extraction module(Multi-feature down-Sampling extraction block,MFDEB)is used to improve the problem of information enhancement of shallow image features.Finally,the primal network fuses the shallow features with the deep features amplified by sub-pixel convolution,and uses the fused features for multi-scale super-resolution reconstruction,and the dual network is used for multi-scale down-sampling kernel estimation.Compared with related algorithms,the reconstructed images obtained by the RAADRNet algorithm have been better improved and improved in both subjective and objective evaluation indicators.(3)In view of the fact that the field of super-resolution image reconstruction mainly uses the bicubic interpolation algorithm to degrade the experimental images,the algorithm is trained and tested in a relatively ideal data set,resulting in poor adaptability of the algorithm.Therefore,this paper introduces Gaussian blur or Gaussian noise or joint Gaussian blur and Gaussian noise into the image,and obtains three low-resolution datasets containing only Gaussian blur,only Gaussian noise,and joint Gaussian blur and noise.Then,the above datasets are further put into this paper and related algorithms for super-resolution reconstruction of CT images.Compared with related algorithms,the algorithm in this paper has certain adaptability and reliability in CT image super-resolution reconstruction,Gaussian blur removal and anti-Gaussian noise.