Statistical physics, mixtures of distributions, and the EM algorithm
Neural Computation
Category learning through multimodality sensing
Neural Computation
Incorporation of Anatomical MR Data for Improved Dunctional Imaging with PET
IPMI '91 Proceedings of the 12th International Conference on Information Processing in Medical Imaging
Probabilistic graphical model of SPECT/MRI
MLMI'11 Proceedings of the Second international conference on Machine learning in medical imaging
| Hi-index | 0.00 |
We present a Bayesian joint mixture framework forintegrating anatomical image intensity and region segmentationinformation into emission tomographic reconstruction in medicalimaging. The joint mixture framework is particularly well suited forthis problem and allows us to integrate additional availableinformation such as anatomical region segmentation information intothe Bayesian model. Since this information is independently availableas opposed to being estimated, it acts as a good constraint on thejoint mixture model. After specifying the joint mixture model, wecombine it with the standard emission tomographic likelihood. TheBayesian posterior is a combination of this likelihood and the jointmixture prior. Since well known EM algorithms separately exist forboth the emission tomography (ET) likelihood and the joint mixtureprior, we have designed a novel EM^2 algorithm that comprises twoEM algorithms—one for the likelihood and one for the prior. Despitebeing dove-tailed in this manner, the resulting EM^2 algorithm isan alternating descent algorithm that is guaranteed to converge to alocal minimum of the negative log Bayesian posterior. Results areshown on synthetic images with bias/variance plots used to gaugeperformance. The EM^2 algorithm resulting from the joint mixtureframework has the best bias/variance performance when compared withsix other closely related algorithms that incorporate anatomicalinformation to varying degrees.