Standard MR-based attenuation correction neglects bone attenuation and therefore leads to an underestimation of the reconstructed PET activity. In this work, we improve PET quantification of clinical non-TOF PET/MR data by joint estimation of attenuation and activity distributions from the PET emission data employing MR prior information.We have recently proposed a method to simultaneously estimate attenuation and activity distributions from non-TOF PET emission data incorporating MR prior information¹. Since the proposed algorithm is an extension of the maximum-likelihood reconstruction of attenuation and activity (MLAA)² for PET/MR, we call it MR-MLAA. A schematic overview of the algorithm, updating attenuation and activity in an alternating manner, is given in figure 1. The available MR images are used to derive initial attenuation maps and so-called attenuation masks, which consist of an air/bone (AB) and a soft tissue (ST) compartment (figure 2). The mask is parameterized by ω(r) with ω = 0 for the air/bone compartment, ω = 1 for the soft tissue compartment, and intermediate values obtained by 3D smoothing of the mask. Based on this voxel-specific information, a prior term modifying the attenuation update is defined as $$L(\boldsymbol r) = \omega(\boldsymbol r) \beta_\text{ST}L_\text{ST} + \left[1-\omega(\boldsymbol r)\right] \beta_\text{AB} L_\text{AB}.$$ Here, the terms L_ST and L_AB are modeled using pre-selected attenuation values for soft tissue and air and bone, respectively, as well as corresponding Gaussian-like probability distributions (figure 2). The parameters β_ST and β_AB represent global weighting factors defining the strength of the prior. The voxel-dependency of the prior is provided by the attenuation mask and its correspondent parameter ω(r). In previous simulation experiments, the MR prior information had been shown to efficiently reduce the cross-talk between attenuation and activity¹. We here present first results of an ongoing patient study. Three clinical non-TOF ¹⁸F-FDG PET/MR data sets of the head region were acquired with the Biograph mMR (Siemens Healthcare, Erlangen, Germany). Average injected activity was 240±7 MBq and the data was acquired 112±6 min post injection. The attenuation masks were derived from diagnostic T1-weighted MR images. All subjects underwent an additional PET/CT examination, of which only the CT data were of interest for this study. To obtain a patient-specific CT-based attenuation map, the CT image was scaled to 511 keV and co-registered with the MR. Finally, the PET rawdata acquired with the mMR was reconstructed using MR-based AC as provided by the vendor, the proposed MR-MLAA algorithm, and CT-based AC for comparison.Figure 3 presents sagittal and transversal slices of the obtained attenuation distributions for one specific patient. In contrast to MRAC, MR-MLAA is able to recover bone attenuation information while at the same time preserving air cavities like the frontal and nasal sinuses. MR-MLAA also improves attenuation estimation in regions affected by MR-induced susceptibility artifacts, e.g., surrounding dental implants. However, some misclassifications of air as bone and vice versa are observed. Neglecting bone attenuation in the MRAC attenuation maps leads to an underestimation of the activity distribution as compared to CTAC, which is particularly evident in the corresponding difference images (see figure 4). In fact, the activity evaluated in the full brain is underestimated by 11.1±1.8 % on average across all patients. The presence of bone in the MR-MLAA attenuation maps greatly reduces the activity underestimation, now reaching 3.2±1.9 % for the full brain throughout the entire study. Regional evaluation showed the activity to be underestimated by 14.0±0.8 % in the occipital lobe for MRAC, while errors compared to CTAC could be reduced to 4.0±1.8 % when using MR-MLAA.Improper segmentation of the attenuation mask as well as misclassifications of air as bone or vice versa may locally increase or decrease the MR-MLAA activity distribution. Challenges remain for thin bone structures and in case of susceptibility-induced artifacts in the MR images. However, compared to MRAC, activity underestimation evaluated in the brain is significantly reduced using the presented approach. To be applicable for whole-body PET/MR, additional tissue classes like fat need to be considered and additional challenges like a limited MR field-of-view need to be coped with.MR-MLAA shows promising potential to improve PET quantification for non-TOF PET/MR imaging. In contrast to other advanced techniques for PET/MR attenuation correction, it does not require additional TOF or transmission information, anatomical prior knowledge taken from a patient atlas, or special MR sequences like UTE. As such, it is directly applicable to current clinical PET/MR systems. MR-MLAA may also be applied retrospectively to routine clinical PET/MR data since it does not require dedicated acquisition protocols.