摘要
锥束X射线磷光断层成像(cone-beam X-ray luminescence computed tomography,CB-XLCT)作为一种新兴的混合成像技术可同时获得解剖结构信息和功能代谢信息,在疾病的早期诊断和靶向治疗上具有广泛发展前景.然而,由于锥束XLCT成像过程中的高度不适定性,重建结果的空间分辨率较低.针对该问题,提出了一种基于分组注意力残差网络的锥束XLCT定量磷光分布重建方法.该方法采用注意力机制优化光子密度特征通道的权重系数,并结合近似域约束损失函数,从而改善残差网络的特征表达过程,提高图像重建精度和模型稳健性.结果表明:所提出的方法在仿真实验中可清晰分辨并重建出靶心距离为7 mm的双目标体,位置误差(LE)约为0.37 mm,戴斯相似性指数(Dice)达到了84%.在仿体实验中,LE约为0.48 mm,Dice达到了79%.因此,分组注意力残差网络方法能够有效在定位精度、空间分辨率、稳健性等多方面提高图像重建质量.
Cone-beam X-ray luminescence computed tomography(CB-XLCT)is an emerging hybrid imaging modality that can simultaneously obtain both anatomical structure information and functional metabolism information,providing good prospects for the early diagnosis and targeted treatment of diseases.However,CB-XLCT generally suffers from high ill-posedness during image reconstruction,severely deteriorating the spatial resolution.To solve this problem,a quantitative phosphorescence distribution reconstruction method for CB-XLCT based on a group attention residual network is proposed.In this method,the attention module is used to optimize the weight of each photon density feature channel,which is combined with the approximate domain constraint loss function to improve the feature expression process and enhance the reconstruction accuracy and model robustness.Results show that the proposed method can clearly distinguish and reconstruct two targets with the target center distance of 7 mm,yielding approximately 0.37 mm location error(LE)and 84%Dice similarity index in the simulation experiment.In the phantom experiment,the LE is approximately 0.48 mm,and the Dice reaches 79%.Therefore,the group attention residual network method can effectively improve the quality of image reconstruction in LE,spatial resolution,and robustness.
作者
周仲兴
郭司琪
贾梦宇
张林
吴越
Zhou Zhongxing;Guo Siqi;Jia Mengyu;Zhang Lin;Wu Yue(School of Precision Instruments and Optoelectronics Engineering,Tianjin 300072,China;Tianjin Key Laboratory of Biomedical Detection Technology and Instrument,Tianjin 300072,China)
出处
《天津大学学报（自然科学与工程技术版）》
EI
CAS
CSCD
北大核心
2022年第10期1082-1092,共11页
Journal of Tianjin University：Science and Technology
基金
国家自然科学基金资助项目(81971656,62175183)
天津市自然科学基金资助项目(19JCYBJC28600).
关键词
锥束XLCT
深度学习
分组注意力
图像重建
cone-beam X-ray luminescence computed tomography(CB-XLCT)
deep learning
group attention
image reconstruction
作者简介
周仲兴(1979—),男,博士,副教授;通信作者:周仲兴,zhouzhongxing@tju.edu.cn.
