Temporal and spatial blood information estimation using Bayesian ICA in dynamic cerebral positron emission tomography

Positron emission tomography (PET) is a nuclear medicine technique that provides tomographic images of the distribution of positron-emitting radiopharmaceuticals. We have previously proposed a method for estimating an input blood curve based on a standard independent component analysis using a specially designed cost function and a preprocessing scheme. While the input blood curve was successfully extracted, the voxels with a negative sign remained in the estimated blood volume image. They should be positive because of its physiological meaning. In this study, ensemble learning introduces a nonnegative constraint to correctly estimate temporal and spatial blood information from dynamic cerebral PET images. The rectified Gaussian distribution and exponential distribution were adopted as nonnegative priors on the elements of the time curves and the source images, respectively. The proposed method (extraction of the plasma time-activity curve using ensemble learning, EPEL) was applied to human brain PET studies with three kinds of radiopharmaceuticals for investigating its validity. The results implied that the EPEL-estimated blood curve was similar to the measured curve, and that the estimated blood volume correlated well with that measured clinically. We concluded that EPEL is a valid method for estimating the blood curve and the blood volume image in a noninvasive way. © 2007 Elsevier Inc. All rights reserved.