PET/MR is a novel hybrid imaging modality which promises to influence diagnostic imaging in the near and long-term future. Despite dedicated and focused research efforts, PET/MR attenuation correction (AC) as required for quantitative PET is still considered challenging; especially when compared to PET/CT. A particular challenge is the correct characterization of bone; i.e. the tissue with the highest PET attenuation but which appears invisible in standard MRI. Recently, zero TE was demonstrated for fast and reliable MR bone depiction [1]. Here we extend that work and demonstrate zero TE for PET/MR attenuation correction.N=10 oncology patients were scanned using a 3-tesla, time-of-flight GE Signa PET/MR scanner (GE Healthcare, Waukesha, WI). Zero TE images were acquired using proton density (PD) weighted parameter settings (FA=1deg, BW=±62.5kHz, FOV=26.4cm, res=2.4mm, 48400 spokes acquired in 41sec) and a GEM head array coil for signal reception. All zero TE images were bias correction using N4ITK [2] and normalized relative to soft tissue. Additionally, outside patient structures (e.g. plastic components from the coil housing and the table) were removed using connected component analysis. For each patient, available CT datasets were aligned to the PET/MR by rigidly registering the whole-body CT to zero TE data followed by diffeomorphic non-rigid registration [3,4]. The zero TE images were converted into pseudo CTs (ZT) using either: 1) image segmentation based on local, pixel-wise thresholding [ZT_SEG(AIR)=-1000HU for zero TE < 0.2 and ZT_SEG(SOFT-TISSUE)=40HU for zero TE > 0.8] and linear scaling [ZT_SEG(BONE)=-2350*(ZTE-1)+40) for 0.2 < zero TE < 0.8] (cf. Fig. 1), or 2) machine learning based on a random forests regression model with each voxel in the pseudo CT represented by a feature vector derived from the corresponding voxel and its surrounding local neighborhood in the zero TE image. The corresponding pseudo CTs are referred to as ZT_SEG and ZT_ML, respectively. PET images were reconstructed using a time-of-flight, ordered subset expectation maximization (TOF-OSEM) algorithm (2 iterations, 24 subsets). For PET/MR attenuation correction, existing atlas-based AC (AT-AC) [4] and gold standard CT-AC were compared relative to ZT_SEG-AC and ZT_ML-AC.The left subplot of Fig. 1 compares zero TE in linear (left), and negative linear scale (middle) vs. CT (right). The inverted zero TE appears similar to the CT by depicting soft-tissue dark and bone grayish against a bright background corresponding to air. The scatter plot, depicted on the right, justifies the linear scaling for ZT_SEG(BONE) as described above. Figure 2 illustrates a close correspondence between the true CT (left) and the two zero TE derived pseudo CTs; i.e. ZT_SEG (middle) and ZT_ML (right). By taking into account neighborhood information and local non-linearity, the Machine Learning based approach better captures tissue / air interfaces otherwise affected by partial volume effects. Figure 3 illustrates the same comparison after rescaling the CTs towards 511keV PET attenuation maps [1/mm] and pasting hardware templates to account for attenuation of auxiliary MR equipment (i.e. coil and table). The energy rescaling and the down-sampling to PET resolution minimize / smooth differences. Figure 4 shows the same comparison for the reconstructed PET images. The error maps are calculated relative to gold-standard PET_CT-AC (left) i.e. (PET_X-PET_CT-AC)/max(PET_CT-AC) and demonstrate significantly improved PET quantitation using ZT-derived pseudo CTs. Aforementioned advantages of the Machine Learning based approach also translate into more accurate PET images using ZT_ML-AC; especially around the trachea, the sinuses and implants. The flat PDw soft-tissue contrast and the efficient capture of short-lived bone signals renders zero TE well suited for pseudo CT conversion as required for PET/MR-AC. This becomes especially apparent when inverting the color scale (cf. Fig. 1). In comparison to a pixel-based pseudo CT conversion (using thresholding and linear scaling), the Machine Learning based framework is more flexible in the sense that it allows a local non-parametric mapping by including local neighborhood and high-dimensional information. This becomes apparent around air cavities often affected by partial volume effects. Figure 5 exemplarily demonstrates improved robustness of ZT_ML for implants (less pronounced and more localized metal artifact) and its applicability for other anatomies such as the pelvis. In contrast to previous Machine Learning attempts requiring multi-contrast (PD, T1, T2, …) input data, the presented method achieves the same with only a single channel input information (i.e. PDw Zero TE) acquired in a scan of only 41sec. From a PET quantitation perspective, the performed N=10 patient study indicates significant improved PET quantitation (relative to gold standard CT-AC) for ZT-AC in the head when compared to existing alternatives.