摘要
有限角度计算机断层扫描(LACT)旨在通过减少扫描角度的范围来减少辐射剂量。由于投影数据是不完备的且未考虑联合有限角度和金属伪影校正(LAMAR)任务,传统方法重建的CT图像往往存在伪影,特别是当患者携带金属植入物时,伪影将进一步加重,影响后期医疗诊断及下游任务的精度。为解决这一问题,该文利用双域知识和深度展开技术,融合Transformer的非局部特性捕获能力和卷积神经网络(CNN)的局部特征提取能力,提出了能够联合解决LAMAR和LACT任务的模型与数据双驱动双域重建网络,记为MD3Net。该文首先构建了双域优化模型,使用邻近梯度下降算法对优化模型进行求解,并将其展开成模型驱动的CT重建网络。其次,设计了任务选择(TS)模块,通过判断初始估计CT图像中有无金属以利用同一模型同时处理有金属和无金属的重建任务。在数据驱动网络中,构建了融合Transformer和CNN的双分支的迹感知投影域邻近子网络和结合通道注意力、空间注意力的图像域邻近子网络,进而提升网络表示能力。实验结果表明,与现有方法相比,所提算法在联合LACT和LAMAR任务上重建效果更好。
Objective Computed Tomography(CT)is widely used in medical diagnostics due to its non-destructive,noncontact imaging capabilities.To lower cancer risk from radiation exposure,clinical practice often limits scanning angles,referred to as Limited-Angle CT(LACT).Incomplete projection data in LACT leads to wedgeshaped artifacts in reconstructions using Filtered Back-Projection(FBP)algorithms.These artifacts worsen in the presence of metallic implants.Although LACT reconstruction without metal and full-angle CT Metal Artifact Reduction(MAR)have been extensively studied,the joint task of Limited-Angle and Metal Artifact Reduction(LAMAR)has received limited attention.This study proposes a model-and data-driven CT network that integrates a Task Selection(TS)module to apply appropriate gradient descent steps for different tasks.This enables simultaneous processing of LACT and LAMAR.The network also incorporates dual-domain information interaction during alternating iterations to reconstruct high-quality CT images.Methods First,a dual-domain reconstruction model integrating both model-and data-driven model is constructed to address the joint task of LACT reconstruction and LAMAR.The model comprises four components:an image-domain data fidelity term,a projection-domain data fidelity term,an image-domain regularization term,and a projection-domain regularization term.These terms are solved using an alternating iteration strategy.The image-and projection-domain subproblems are addressed using the proximal gradient descent algorithm,with the iterative process unrolled into a Deep Neural Network(DNN).Each stage of the deep unrolling network includes three components:a TS module,a projection-domain subnetwork,and an image-domain subnetwork.The TS module dynamically determines gradient descent step sizes for the LACT and LAMAR tasks by comparing image-domain FBP reconstruction results with predefined thresholds.The projection-domain subnetwork is shared by both tasks.Finally,the data-driven proximal network comprises the projection-domain and image-domain subnetworks.The projection-domain subnetwork includes an encoder,a dual-branch structure,and a decoder.The encoder has two stages,each consisting of a convolutional layer followed by an activation function;the decoder mirrors this architecture.A Transformer-based long-range branch incorporates non-metal trace information into a self-attention mechanism to guide correction of metal trace data using contextual information from non-metal regions.A short-range branch,composed of six residual blocks,extracts local features.The outputs of the two branches are fused using a weighted strategy before being passed to the decoder.The image-domain subnetwork is implemented as an attention-based U-Net.Channel and spatial attention mechanisms are applied before each of the four downsampling operations in the U-Net encoder.This design allows the decoder to more effectively leverage encoded information for high-quality CT image reconstruction without increasing the number of network parameters.Results and Discussions Experimental results on both LACT reconstruction and LAMAR tasks show that the proposed method outperforms existing CT reconstruction algorithms in both qualitative and quantitative evaluations.Quantitative comparisons(Table 1)indicate that the proposed method achieves higher average Peak Signal-to-Noise Ratio(PSNR),Structural Similarity Index(SSIM),and lower Root Mean Square Error(RMSE)for both tasks across three angular ranges.Specifically,average PSNR improvements for the LAMAR and LACT tasks reach 2.78 dB,2.88 dB,and 2.32 dB,respectively,compared with the best-performing baseline methods.Qualitative comparisons(Fig 4 and Fig 5)show that reconstructing CT images and projection data through alternating iterations,combined with dual-domain information interaction,enables the network to effectively suppress composite artifacts and improve the reconstruction of soft tissue regions and fine structural details.These results consistently exceed those of existing approaches.Visual assessment of reconstruction performance on the clinical dataset for the LAMAR task(Fig 6)further demonstrates the method’s effectiveness in reducing metal artifacts around implants.The reconstructed images exhibit clearer structural boundaries and improved tissue visibility,indicating strong generalization to previously unseen clinical data.Conclusions To address the combined task of LACT reconstruction and LAMAR,this study proposes a dual-domain,model-and data-driven reconstruction framework.The optimization problem is solved using an alternating iteration strategy and unfolded into a model-driven CT reconstruction network,with each subnetwork trained in a data-driven manner.In the projection-domain network,a TS module identifies the presence of metallic implants in the initial CT estimates,allowing a single model to simultaneously handle cases with and without metal.A trace-aware projection-domain proximal subnetwork,integrating Transformer and convolutional neural network architectures,is designed to capture both local and non-local contextual features for restoring metal-corrupted regions.In the image-domain network,a U-Net architecture enhanced with channel and spatial attention mechanisms is used to maximize spatial feature utilization and improve reconstruction quality.Experimental results on the AAPM and DeepLesion datasets confirm that the proposed method consistently outperforms existing algorithms under various limited-angle conditions and in the presence of metal artifacts.Further evaluation on the SpineWeb dataset demonstrates the network’s generalization capability across clinical scenarios.
作者
石保顺
程诗展
姜轲
傅昭然
SHI Baoshun;CHENG Shizhan;JIANG Ke;FU Zhaoran(School of Information Science and Engineering,Yanshan University,Qinhuangdao 066004,China;Department of Stomatology,Qinhuangdao First Hospital,Qinhuangdao 066001,China)
出处
《电子与信息学报》
北大核心
2025年第5期1569-1581,共13页
Journal of Electronics & Information Technology
基金
国家自然科学基金(62371414)
河北省自然科学基金(F2023203043)
河北省重点实验室资助课题(202250701010046)。
关键词
有限角度CT
金属伪影校正
模型与数据双驱动
双域网络
Limited-Angle Computed Tomography(LACT)
Metal Artifact Reduction(MAR)
Model and data dual-driven
Dual-domain network
