稀疏Transformer联合细节还原网络的CT金属伪影去除算法  

CT Metal Artifact Reduction for Restoration Network Sparse Transformer Joint with Details

作　　者：任亚利 刘祎[1,2] 桂志国 张鹏程[1,2] REN Yali;LIU Yi;GUI Zhiguo;ZHANG Pengcheng(School of Information and Communication Engineering,North University of China,Taiyuan 030051,China;State Key Laboratory of Dynamic Testing Technology,North University of China,Taiyuan 030051,China)

机构地区：[1]中北大学信息与通信工程学院,太原030051 [2]生物医学成像与影像大数据山西重点实验室,太原030051

出　　处：《计算机测量与控制》2025年第3期197-204,212,共9页Computer Measurement &Control

基　　金：山西省基础研究计划项目(202303021211148);山西省专利转化专项计划项目(202302006)。

摘　　要：在CT成像中,若病人体内存在金属植入物时,CT重建图像中会出现严重的金属伪影,降低图像质量并影响医生诊断结果;对现有金属伪影去除算法进行了研究,分析了简单的神经网络模型去除金属伪影时存在伪影残留及组织细节模糊的问题,提出了稀疏Transformer联合细节还原网络;该网络由伪影去除网络和细节还原网络两个独立的子网络构成,伪影去除网络将标准Transformer中的自注意力替换为稀疏注意力机制,并且引入混合尺度前馈网络提取多尺度信息,以产生更好的图像去噪特征;细节还原网络同时提取全局和局部信息,在不损失图像分辨率的前提下清晰地恢复原始图像细节,然后通过加法运算将其整合到上述去除伪影的图像中;在Deeplesion数据集上验证了模型的有效性,实验结果表明,该方法在金属伪影去除效果上优于目前已有方法,在PSNR、SSIM指标上表现更优,能有效去除金属伪影,恢复出大部分精细的结构细节。In CT imaging,if there is a metal implant in the patient s body,it will appear serious metal artifact in CT reconstruction images,reducing the image quality and affecting the doctor s diagnosis.This paper studies existing metal artifact reduction methods,analyzes artifact residue and fuzzy organizational details with a simple neural network model removing metal artifacts.A joint sparse Transformer and detail restoration network is proposed,which consists of two independent sub networks:artifact removal network and detail restoration network.The artifact removal network replaces the self-attention mechanism in standard Transformer with the sparse attention mechanism,the mixed-scale feed-forward network is introduced to extract multi-scale information,achieving better image denoising features.Additionally,the detail restoration network extracts both global and local information to clearly recovers the original image details without losing image resolution,and then integrate the detail feature map into the above artifact reduction image through addition operation.The effectiveness of the model is verified on the Deeplesion dataset.Experimental results show that the proposed method outperforms the existing network models in metal artifact reduction,showing superior performance in peak signal to noise ratio(PSNR)and structural similarity index measure(SSIM)metrics,which can effectively remove metal artifacts and recover most of fine structural details.

关 键 词：CT图像 金属伪影 稀疏Transformer 注意力机制 细节还原 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]