摘要
从ML-EM重建算法入手,分析了贝叶斯模型的一些关键点.针对采用传统方法求解MAP问题的局限性,提出一种用于正电子成像的贝叶斯神经网络(BNN)重建算法.为了保留边缘信息,引入了二进制的保边缘变量,并应用共轭神经网络求解.模拟的重建结果表明,应用这种算法可以得到比ML-EM算法更好的重建图像.
Some key aspects in the Bayesian Maximum Likelihood- Expectation Maximization Method (ML- EM)for positron emission tomography(PET)imaging were investigated.In order to overcome the limitation of traditional solutions to estimate Maximum a Posteriori(MAP),a Bayesian neural network(BNN) algorithm was proposed for PET imaging.In addition to real- valued source intensities,binary variables were introduced to protect the information of the edges.These two different kinds of variables can be obtained by a coupled gradient network composed of two interacting recurrent networks corresponding to the two kinds of variables respectively.Compared with ML- EM reconstruction,the BNN results showed higher quality.
出处
《浙江大学学报（工学版）》
EI
CAS
CSCD
北大核心
2003年第5期543-546,569,共5页
Journal of Zhejiang University：Engineering Science
