A neural network for fusing the MR information into PET images to improve spatial resolution

We propose a neural network architecture to fuse the anatomical information given by an MR image, into a PET image to reconstruct a reasonable activity distribution in the brain. In the network, convolutional parameters and the anatomical brain structure are expressed in pre-wired weights. When an observed PET image is given to the comparison side of the network, the activity profile of the activity layer is iteratively adjusted to constitute a reasonable model for the positron generating profile, using a modified network inversion technique.<
>