Abstract: Low-dose CT(LDCT) scan plays a pivotal role in clinical practice,effectively decreasing cancer risks for radiologists and patients.However,the utilization of low-dose radiation introduces notable noise into the resulting CT images,highlighting the necessity of low-dose CT reconstruction.Another important task in the field of image reconstruction is super-resolution(SR),with the aim of achieving high-resolution CT imaging while minimizing computational expenses.High resolution CT images afford the capacity to capture intricate anatomical details in greater fidelity.Although significant progress has been made in their respective domains,there is still a lack of effective methodologies that can effectively harness the inherent correlation between these tasks and handle them concurrently.We employ edge information as a link between the two tasks,and utilize gradients to extract shared features from both tasks.This allows the LDCT reconstruction process to assist the SR reconstruction process and gene-rate resulting images with sharp edges.Our work consists of three components:1)Edge-enhanced framework.The framework fully exploits the correlation between the two tasks by extracting relevant features using gradient information,enabling the denoising(DN)task to assist the SR task in achieving superior performance.2)Gradient guided fusion block(GGFB),which enhances the highly correlated edge features while suppressing irrelevant features,thereby enabling effective reconstruction in edge regions.3)Gradient loss,which introduces richer gradient features into the model and guides the network to prioritize the reconstruction of edge regions.Extensive experiment demonstrates that our noise reduction and super resolution reconstruction network(NRSR-Net)achieves promising PSNR,SSIM,and LPIPS in quantitative evaluations,as well as gains high-quality readable visualizations.All of these advantages demonstrate the great potential of NRSR-Net in clinical CT imaging.

Key words: Low-dose CT, Super-resolution, Gradient guidance, Multi-task, Edge