Research On Low-dose Spectral CT Image Reconstruction Method Based On Deep Learning

Computed Tomography(CT)can obtain tomographic images of human internal tissues in a non-invasive way,which has become a research hotspot in the field of medical imaging in recent years.Spectral CT reconstructs images using the difference in photons attenuation about the detected object,which can improve the material resolution of traditional CT and provide accurate material separation and quantified information.Recently,the material decomposition method commonly used in multienergy projection data processing is seriously affected by noise,resulting in low signalto-noise ratio of reconstructed images under low dose conditions.In this thesis,aiming at the low-dose spectral CT imaging,the monochromatic reconstruction of spectral CT images was performed on the basis of removing low-dose noise by the advantages of deep learning technology in image processing.In the aspect of low-dose image denoising,a model based on a deep generative adversarial network was proposed to utilize the redundant characteristics of spectral CT images in this thesis.By using the multi-energy CT images as input data,the energy correlation and spatial correlation were combined with the generative process of the network to preserve the image texture information while removing noise.At the same time,residual learning was introduced to multiplex features from different levels to avoid the loss of detail caused by the deep network layers,then the low-dose images were restored to the spectral CT images that are equivalent to the normal dose images.In addition,in order to solve the problem of network training,a Wasserstein distance was added to make the convergence faster and better.Based on the denoised images,a hybrid perceptual loss function was proposed to map the correlation between polychromatic images and monochromatic images,which can measure the diffierences from pixel consistency,texture details,CT value distribution and perceptual characteristics.In this way,the noise amplification problem caused by the traditional basic material decomposition method can be solved.The proposed algorithm can generate high-quality monochromatic images without beam hardening and metal artifacts,and fit accurate CT values to make judgments on the measured material density.The experimental results showed that the low-dose denoising algorithm improved the PSNR value of the input images by about 5d B,the SSIM value by about 0.18,and the FSIM value by about 0.056.The final reconstructed monochromatic images at40 ke V,55 ke V,70 ke V,and 100 ke V performed better than other methods in quantitative and qualitative evaluations of CT accuracy,beam hardening artifacts,and metal artifacts.The proposed method can improve the quality of reconstructed images while reducing the low dose noise in spectral CT.