Neural network-based PET image reconstruction.

In PET image analysis, conventional deconvolution alone will not give sufficient information for a precise study of a localized brain function. In the deconvolution process, which is a type of inverse problem, it is important to confine the solution space by incorporating a priori knowledge such as the tissue distribution given by MR images as well as smoothness in the blood flow distribution profile. An MR-embedded neural-network model is described to reduce the partial volume effect in the restoration of blood flow profiles from PET images.