基于梯度引导的低剂量CT超分辨率重建算法  

Super-resolution Reconstruction for Low-dose CT Based on Guidance of Gradient

作　　者：徐颖 张道强[1] 葛荣骏 XU Ying;ZHANG Daoqiang;GE Rongjun(College of Computer Science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China;College of Instrument Science and Engineering,Southeast University,Nanjing 210096,China)

机构地区：[1]南京航空航天大学计算机科学与技术学院,南京211106 [2]东南大学仪器科学与工程学院,南京210096

出　　处：《计算机科学》2024年第8期143-151,共9页Computer Science

基　　金：江苏省自然科学基金(BK20210291);国家自然科学基金(62101249);中国博士后科学基金(2022M721611,2021TQ0149)。

摘　　要：低剂量CT(Low-dose CT,LDCT)扫描在临床实践中起着关键作用,其能有效降低放射科医生和患者的患癌风险。然而,低剂量射线的使用会给生成的CT图像引入明显的噪声,这一问题凸显了LDCT降噪重建的必要性。图像重建领域中的另一个重要任务是超分辨率(Super-resolution,SR)重建,其目标是在减少计算开销的同时实现高分辨率的CT成像。高分辨率CT图像能够更准确地捕捉复杂的解剖细节。尽管这些任务在各自领域取得了显著进展,但目前仍缺乏能够有效利用这两个任务之间固有相关性并同时处理它们的有效方法。文中将边缘信息作为两个任务之间的纽带,并利用梯度提取强相关特征。这使得LDCT降噪重建过程能够辅助超分辨率重建过程,并最终生成具有清晰边缘的结果图像。文中提出的降噪和超分辨率重建网络(NRSR-Net)包括3个组成部分:1)边缘增强框架,该框架利用梯度信息引导和提取相关特征,从而充分利用两个任务之间的相关性,使降噪任务能够辅助超分辨率任务实现更好的性能;2)梯度门控融合模块(Gradient Guided Fusion Block,GGFB),该模块增强高度相关的边缘特征并抑制无关特征,从而实现边缘区域的有效重建;3)梯度损失,该损失函数为模型引入更加丰富的梯度特征,并使网络重点还原边缘区域。一系列的实验表明,NRSR-Net在定量评估中取得了令人满意的峰值信噪比(PSNR)、结构相似性(SSIM)和学习感知图像块相似度(LPIPS),并获得了高质量的可视化结果。这些优势表明NRSR-Net在临床CT成像中具有巨大潜力。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.

关 键 词：低剂量CT 超分辨率 梯度引导 多任务 边缘 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]