While this conclusion is encouraging, current methods have several limitations. First, most require long computing times, not suitable for a clinical setting. Second, they present the fundamental drawback of not incorporating information about attenuation at the typical energies of PET annihilation photons of 511 keV. Atlas-based methods use transmission data obtained from CT photons with energies between 30 and 140 keV, much lower than 511 keV from positron annihilation (Kinahan 1998, Kinahan 2003). These data are scaled to obtain the attenuation at 511 keV, but these scalings are based on piece-wise linear functions that cannot account for non-linear effects, such as those occurring when CT photons are fully absorbed at metal dental implants (Fuin 2017). Third, computational complexity for atlas-based methods scale linearly with the number of subjects in the atlas and therefore atlases are limited to small number of similar subjects. Fourth, methods based on MR segmentation are limited by the MR insensitivity to the electronic density.We propose to use an existing dataset with more than 3000 scanned subjects, each with coregistered transmission PET, emission PET, and MR scans, to train an artificial neural network to predict the attenuation correction given only the subject MR scan. Scans in the dataset were performed as part of past and current grants in our Harvard Aging Brain group at the Massachusetts General Hospital and the Harvard Medical School.The fundamental goal of this project is improve the quality of brain PET/MR imaging by performing an accurate attenuation correction that could also be used in a clinical setting.  A secondary goal is to provide a gold-standard, which we currently lack, that might be useful to develop other methods.The fundamental hypothesis of this project is that an artificial neural network can learn how photons in PET attenuate in the brain using pairs of MR and transmission scans.  This hypothesis has two basis. First, several mathematical theorems guarantee that artificial neural networks are universal approximators (LeCun 2015).  This means that artificial neural networks can approximate to arbitrary precision, i.e. learn, any function.  The second basis of our hypothesis is that pairs of MR and transmission scans are suitable examples to learn this task.Our approach addresses some of the limitations of current methods. First, once the neural network is trained, the prediction of the attenuation from a new MR scan is computationally very fast, requiring at most seconds, and could be easily incorporated in a clinical setting. Second, information about the actual physical process causing the attenuation (Compton scattering at 511 keV) is encoded in the network. Third, the network could be trained with little additional data on specific populations absent from our initial dataset and that are challenging for current methods, such as subjects with solid lesions or pediatric population (Gatidis 2017).We think that our approach might offer improvements critical when accurate PET quantification is needed, for example, when evaluating disease progression in longitudinal studies.