[en] Magnetic resonance imaging (MRI) provides high-resolution anatomical information. In this study, we investigated the usefulness of an MRI-constrained method for image reconstruction of positron emission tomography (PET). We started the reconstruction process with a conventional Bayesian reconstruction (BAY) method. Typically after 7 iterations, we calculated the apparent radioactivity ratio between gray matter and other tissues such as white matter and cerebrospinal fluid from both the PET image and tissue classification probability images generated from segmented MR images, and then increased this ratio to enhance the functional boundary gradient. We multiplied the apparent ratio by 2.0. Subsequently, the cross-correlation between anatomical and enhanced functional boundary maps was incorporated into BAY using the weighted line site method. This algorithm was tested with simulated PET projection data generated from an MRI-based three-dimensional brain phantom, including the effects of resolution, attenuation, scatter and Poisson noise. The reconstructed images were evaluated in terms of root mean square distance (RMSD) from the true image. This method improved RMSD by 39.6%, 33.5%, and 33.0% on average, compared to the filtered backprojection method, the maximum-likelihood expectation-maximization method and BAY method, respectively. Our preliminary results suggest that incorporation of both anatomical and tissue classification information into the image reconstruction process is useful for improving the qualitative and quantitative accuracy of PET images. (author)