CT Image Super Resolution Reconstruction Based On Deep Neural Network

Computed Tomography（CT）utilizes the transmission information of X-ray through the object to obtain the internal structure information of the object noninvasively by image reconstruction algorithm.It is widely used in the fields of medical examination and industrial detection.Resolution is an important index to evaluate the quality of CT images.High-resolution CT images are helpful to improve the accuracy of diagnosis and detection.However,in the process of CT imaging,the resolution of CT images is restricted by imaging system hardware and software factors such as the focus size of X-ray source,pixel size of detector,geometric optical amplification,imaging field of view,and reconstruction algorithm.However,improving the resolution through hardware equipment will bring a significant increase in manufacturing cost.Therefore,the research of CT image super resolution reconstruction technology has important theoretical significance and practical application value.This paper focuses on the super-resolution reconstruction of CT images and,facing different scene requirements,and carries out researches on the super-resolution post-processing technology based on image domain and the image super-resolution reconstruction algorithm based on projection iteration for single-energy CT,and researches on the super-resolution method based on image domain for dual-energy CT.The main research achievements are as follows:1.A CT Image super-resolution reconstruction method based on the guidance of deep gradient information is proposed.Aiming at the problems of structural distortion and detail blurring in the reconstruction results of the current super-resolution network using pixel-level loss function,this paper proposes a gradient-guided CT super-resolution network（G-CTSR）.G-CTSR adopts generative adversarial network framework,and its generator consists of super-resolution branch and gradient branch.Super-resolution branching is used to learn the mapping from low resolution CT images to high resolution CT images.Combined with gradient branching guidance,it enhances the learning of edge structure features from low resolution images to high resolution images.The loss function of G-CTSR network is composed of image space loss function and gradient space loss function,which further assists the network to generate more detailed textures.The experimental results show that the super-resolution result of G-CTSR has advantages in retaining high frequency information and restoring details compared with the structure-preserving super-resolution network,which can effectively reduce the structural distortion in the reconstruction results and improve the visual effect of the super-resolution reconstructed CT images.2.An iterative reconstruction algorithm of CT image based on plug and play framework is proposed.The traditional regularization super-resolution reconstruction model lacks the description of deeper image feature information,resulting in too smooth results and insufficient high-frequency information.This paper proposes a CT image super-resolution iterative reconstruction algorithm based on the plug and play framework,combining the ability of deep networks to depict image features and the fidelity performance of traditional mathematical models.In this method,data fidelity items are constructed by projecting data,combined with image gradient L₀-norm,and super resolution regularization items based on deep learning are introduced by plug and play framework.The reconstructed model is transformed into three sub-problems by alternating direction method of multipliers for iterative solution.The experimental results show that the experimental results of the proposed algorithm are more superior in both quantitative evaluation and qualitative comparison.It can effectively suppress noise and recover detailed texture information,improving the resolution of CT images.3.A super resolution image generation method for dual energy CT based on dual-channel information fusion is proposed.Dual-energy CT scans objects by X-ray under two different energies,and makes use of the difference in attenuation coefficients of substances to obtain more abundant material information.Focusing on the problems of low signal-to-noise ratio of low-energy CT images,low contrast of high-energy CT images,and serious noise interference in subsequent material decomposition results,Based on the characteristics of structural consistency of high-low energy CT images,this paper proposes a super-resolution adversarial network for dual-energy CT（De-SRGAN）based on channel information fusion.De-SRGAN network extracts the features of high-low energy CT images through two encoders,and the two decoders learn the features of CT images in different energy channels in an interactive way,aiming to integrate the high signal-to-noise ratio information of high-energy CT images with the high contrast information of low-energy CT images,so as to better realize the super-resolution mapping of dual-energy CT images.The experimental results show that compared with the super-resolution network without CT image information fusion,the ultra-resolution results of De-SRGAN have clearer texture,more advantages in detail information restoration,and can obtain higher signal-to-noise ratio of material decomposition results.