Research On CT Reconstruction Based On Deep Learning

CT(Computed Tomography)is a medical imaging technique.In recent years,CT has been widely used in various clinical diagnoses.Due to the potential health risks as-sociated with X-ray radiation,the current research aims to reduce X-ray radiation while achieving high-quality CT imaging.However,most of the existing research on low-dose CT image super-resolution reconstruction has not fully explored the potential connections and interactions between denoising and super-resolution tasks.Existing low-dose CT im-age reconstruction methods can effectively improve image quality,but they still need to be improved in reconstructing high-frequency detail information in low-dose CT images.This paper focuses on the research of super-resolution reconstruction of noisy CT images(low-dose CT or sparse-view CT images)and the dual-domain low-dose CT reconstruc-tion technology,and the main research work is as follows:(1)For the super-resolution reconstruction of low-dose CT images,considering the interdependence between low-dose reconstruction and super-resolution reconstruction,we propose a denoising-guided CT image super-resolution reconstruction framework,using the reconstructed denoising information to guide super-resolution reconstruction.The framework performs denoising and super-resolution tasks in parallel and uses a filter gate module to filter the information from the denoising branch,highlighting the significant features that are favorable for super-resolution reconstruction.And the filter gate module can assist and guide the reconstruction of the super-resolution branch.(2)A dual-channel joint learning framework is proposed to achieve super-resolution reconstruction of sparse-view CT images.The framework can effectively establish the connection between the two tasks and explore the deep connections between them.And achieve a balance between feature interaction performance and model computation by alternately using channel interaction modules and spatial interaction modules at differ-ent stages.Additionally,a wavelet fusion module is introduced to further recover high-frequency information such as the detail texture of the image by fusing information be-tween different high-frequency wavelet sub-bands.(3)Aiming at the shortcomings of insufficient information in existing single-domain models and the limitations of cascading dual-domain models such as cumulative errors,a low-dose CT reconstruction technology based on dual-domain is proposed,which fully utilizes the dual-domain information through an end-to-end dual-domain joint learning strategy to reconstruct high-quality full-dose CT images from low-dose CT scan data.Additionally,an edge detection module based on learnable edge operators is proposed to extract rich edge information,further assisting and guiding the reconstruction of edge detail features in the image.Furthermore,structure-aware loss and gradient loss are intro-duced to facilitate the reconstruction of clear texture details.