基于Transformer块的混合域网络稀疏角度CT成像  

Hybrid Domain Network for Sparse View CT Imaging Based on Transformer Blocks

作　　者：张庭宇 吴凡 金潼 孙宇 刘进 亢艳芹 ZHANG Tingyu;WU Fan;JIN Tong;SUN Yu;LIU Jin;KANG Yanqin(School of Computer and Information,Anhui Polytechnic University,Anhui Wuhu 241000,China)

机构地区：[1]安徽工程大学计算机与信息学院,安徽芜湖241000

出　　处：《重庆工商大学学报（自然科学版）》2024年第5期38-48,共11页Journal of Chongqing Technology and Business University:Natural Science Edition

摘　　要：目的 针对计算机断层扫描(Computed Tomography,CT)中由于不完全扫描数据导致图像噪声伪影严重的问题,提出一种基于Transformer块的混合域网络稀疏角度CT成像算法(Hybrid Domain network for sparse view CT imaging based on Transformer,HDTransformer)。方法 算法的主要思想是借助于新型的Transformer网络,构建适用于多阶段稀疏角度CT投影数据及图像数据的处理流,以提高稀疏角度CT图像重建质量;与现有两阶段混合域处理方法相比,本方法采用图像域-投影域-图像域三阶段混合处理流程,通过多阶段信息的联合互补提高成像质量;此外,针对不同阶段数据噪声伪影特点设计不同的Transformer块,以实现差异化的处理;更进一步,算法采用可微分的解析重建和投影运算,建立投影域与图像域数据的转换,最终实现端到端的稀疏角度CT优质成像流。结果 通过Mayo数据实验验证,其视觉结果表明:处理后的不同部位CT图像噪声伪影均能够得到较好的抑制;量化结果表明:处理后的CT图像峰值信噪比和特征相似性均优于对比方法。结论 实验的定性和定量结果表明:所提算法在去除图像伪影噪声方面要优于其他算法,具有更高的质量,验证了该方法的有效性。Objective A hybrid domain network for sparse view CT imaging based on Transformer(HDTransformer)was proposed to address the serious image noise artifacts caused by incomplete scanning data in computed tomography(CT).Methods The main concept of the algorithm was to utilize a novel Transformer network to construct a processing flow suitable for multi-stage sparse view CT projection data and image data to improve the quality of sparse view CT image reconstruction.In comparison to existing two-stage hybrid domain processing methods,this approach adopted a three-stage hybrid processing flow of image domain-projection domain-image domain,enhancing imaging quality through the joint complementary information of multiple stages.Furthermore,different Transformer blocks were designed based on the characteristics of noise and artifacts in data at different stages for differentiated processing.Moreover,the algorithm adopted differentiable analytical reconstruction and projection operations to establish the conversion of data between projection domain and image domain,ultimately achieving end-to-end high-quality sparse view CT imaging flow.Results Through Mayo data experimental verification,the visual results showed that the processed CT images of different parts effectively suppressed noise artifacts.The quantization results showed that the peak signal-to-noise ratio and feature similarity of the processed CT images were better than those of the comparison method.Conclusion The qualitative and quantitative results of the experiment indicate that the proposed algorithm outperforms other algorithms in removing image artifacts and has higher quality,verifying the effectiveness of this method.

关 键 词：深度学习 图像修复 Transformer模块 混合域 

分 类 号：TN911.73[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]