基于U-Net的两阶段SPECT骨显像降噪方法研究  

Research on denoising method of two-stage SPECT bone imaging based on U-Net

作　　者：余泓 罗仁泽[1] 陈春梦 郭亮 罗任权 YU Hong;LUO Renze;CHEN Chunmeng;GUO Liang;LUO Renquan(School of Electrical Engineering and Information,Southwest Petroleum University,Chengdu,Sichun 610500,China;Department of Nuclear Medicine,The No.2 People's Hospital of Yibin,Yibin,Sichun 644000,China;College of Geoscience and Technology,Southwest Petroleum University,Chengdu,Sichun 610500,China)

机构地区：[1]西南石油大学电气信息学院,四川成都610500 [2]宜宾市第二人民医院核医学科,四川宜宾644000 [3]西南石油大学地球科学与技术学院,四川成都610500

出　　处：《光电子．激光》2023年第7期771-784,共14页Journal of Optoelectronics·Laser

基　　金：四川省科技计划项目(2019CXRC0027)资助项目。

摘　　要：进行单光子发射计算机断层成像(single-photon emission computed tomography,SPECT)骨显像检查时,为减少给病人带来的辐射伤害,医师常会减轻辐射剂量,导致骨显像信噪比、分辨率较低,严重影响诊断以及病灶自动检测效果。为提升骨显像质量,提出了一种基于U-Net的两阶段SPECT骨显像降噪方法。首先,设计了一种U-Net噪声估计网络来快速估计每张骨显像的噪声水平,为主干降噪网络提供噪声先验知识。其次,主干降噪网络同样以U-Net为基础框架,同时结合多尺度特征融合、通道-空间注意力机制结构来增强网络的噪声特征提取能力,预测出噪声图。最后,通过残差学习得到降噪骨显像。同时,为解决使用均方误差(mean square error,MSE)损失函数的重建图像过于平滑的问题,设计了一种复合损失函数,保留骨显像的原有细节信息。实验中,向训练集中的骨显像施加不同噪声水平进行数据扩充,并且采用迁移策略解决模型过拟合问题。结果表明,与目前主流算法相比,所提出的降噪方法能够有效降低骨显像噪声,并且保留病灶细节特征。此外,通过盲降噪能够改善原骨显像质量、提升病灶自动分割效果。When performing single-photon emission computed tomography(SPECT)bone imaging examination,physicians often reduce radiation damage by reducing radiation dose,resulting in low signal-to-noise ratio and resolution of bone imaging,which seriously affects the diagnosis and automatic detection of lesions.In order to improve the quality of bone imaging,a two-stage SPECT bone imaging noise reduction method based on U-Net is proposed.Firstly,a U-Net noise estimation network is designed to quickly estimate the noise level of each bone image,providing noise prior knowledge for the backbone noise reduction network.Secondly,the backbone noise reduction network also uses U-Net as the basic framework,and combines multi-scale feature fusion and channel-spatial attention mechanism structure to enhance the noise feature extraction ability of the network and predict the noise map.Finally,denoised bone imaging is obtained through residual learning.At the same time,in order to solve the problem that the reconstructed image using the mean square error(MSE)loss function is too smooth,a composite loss function is designed to retain the original detailed information of bone imaging.In the experiments,different noise levels are applied to the bone images in the training set for data augmentation,and a transfer strategy is used to solve the problem of model overfitting.The results show that,compared with the current mainstream algorithms,the proposed noise reduction method can effectively reduce the noise of bone imaging and preserve the detailed features of the lesions.In addition,blind noise reduction can improve the imaging quality of the original bone imaging and improve the automatic segmentation effect of lesions.

关 键 词：SPECT骨显像 深度学习 噪声水平 注意力机制 残差学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]