Multimodal reconstruction is used to eliminate the need for the present a posteriori registration approaches for positron emission tomography (PET) and magnetic resonance imaging (MRI) brain images, reducing the overall computational effort. The problem of reconstructing a PET brain image is formulated as a constrained maximization problem. The knowledge of the MRI anatomical boundary for each brain slice during the PET reconstruction process yields an MRI-registered PET functional image. Preliminary results indicate: a) a registration accuracy of 0.1 PET pixels between the geometric centroids of the transmission and MRI binary images and b) an accuracy from 0.5 to 1.3 PET pixels between the PET and MRI brain surfaces at the principal axes intersection. To accelerate the multimodal reconstruction algorithm it is formulated as an extension of the PET multigrid reconstruction algorithm of Ranganath et al. (1988). The functional information contained in the averaged PET image of 40 registered normal brains is used as a starting nominal to further accelerate convergence. Preliminary results illustrate the feasibility of the method.