Med Phys (2015) submitted
IV. A Impact of time of flight on erroneous attenuation correction
With the advent of clinical TOF-PET systems, several studies reported that TOF can improve PET image quality in terms of signal-to-noise ratio, lesion detectability, convergence rate [175, 176] and tolerance to inconsistencies between emission and correction data (including attenuation) compared to non-TOF reconstructions [177, 178].
With TOF capability, the detection time differences of the coincident annihilation photons are measured, with a temporal uncertainty governed by the timing resolution of the PET detectors, and exploited during image reconstruction. The image voxels are locally updated based on the spatial uncertainty associated with the TOF resolution of the scanner and therefore the intra-voxel dependencies and therefore error propagations are reduced. Recently, Ahn et al.  developed analytical methods for the evaluation of attenuation error propagation in TOF and non-TOF PET image reconstructions. They demonstrated that the propagated errors in TOF PET are proportional to TOF timing resolution, while in the case of non-TOF PET, they are proportional to the patient size.
Wollenweber et al.  evaluated the effect of excluding a surface coil from the MRAC maps on PET quantification using TOF and non-TOF reconstructions. They found that with TOF information, the SUV errors due to neglecting the coil attenuation are slightly reduced from –8.2% to –7.3%. Davison et al.  evaluated the impact of TOF on reduction of PET quantification errors induced by metal artifact voids in MRAC maps. In this work, artificial signal voids were simulated in different regions of the attenuation maps of 7 patients undergoing sequential PET/CT and simultaneous TOF PET/MRI using the Signa system (GE Healthcare, Waukesha, WI) with nominal TOF time resolution of <400 ps  scans. It was found that the TOF capability significantly reduces the artifacts. The percentage error reduction with TOF ranged from 21% to 60% for medium-sized artifacts simulated in the maxilla and the sternum, respectively. For artifacts located in the chest, sternum, and pelvic regions, the most significant error reduction achieved was at least 40%. Using the same PET/MRI system referenced above, the influence of TOF on improving the image quality and diagnostic interpretation of PET images in the presence of attenuation artifacts was evaluated using 25 clinical studies . It was concluded that PET/MR reconstructions benefit from TOF information in terms of recovery of pathologies missed on non-TOF PET images and artifact reduction especially around metallic implants.
Mehranian and Zaidi  also studied the impact of TOF PET image reconstruction on the reduction of quantification errors induced by the standard 4-class MRAC and the presence of metal and respiratory-phase mismatch related artifacts. In this work, 27 whole-body FDG PET/CT datasets acquired on the Siemens mCT scanner (with nominal TOF resolution of 580 ps) were analyzed. The results showed that non-TOF MRAC method resulted in an average error of –3.4% and –21.8% in the lungs and bones, respectively, whereas the TOF reconstructions reduced the errors to –2.9% and –15.3%, respectively. Simulation results also showed that as TOF resolution improves, the visual artifacts and quantification errors are substantially reduced (Figure 9). It was concluded that MRI-guided attenuation correction should be less of a concern on future TOF PET/MR scanners with improved timing resolution.
133 IV. B Advances in ultrashort and zero echo time MRI
Ultra-short echo time MRI sequences have been developed to depict tissues with low proton density and short T2 relaxation time, such as the lung and cortical bone. The transverse relaxation of hydrogen protons is mainly governed by their bonding status and chemical environment. In compactly structured tissues, the spin-spin interactions are high leading to fast dephasing and T2 shortening, while in liquid water such as cerebrospinal fluid, the free and bound-water molecules are sparsely distributed leading to less spin-spin interactions and longer T2. The MR signal of short T2 tissues decay quickly even during RF excitation. The UTE pulse sequence has therefore been developed to start signal acquisition as quickly as possible. It differs from conventional sequences in three aspects :
(i) Free induction decay (FID) signal sampling: In conventional spin- and gradient-echo sequences, the FID signal is refocused by a 180º pulse and its echo is sampled at time TE or spoiled and recalled at time TE by a bi-polar gradient pulse, respectively. While in UTE, the FID signal is directly sampled after RF excitation with only a delay from hardware transmit/receive switching time.
(ii) Short RF pulse: The signal of short T2 components can decay during the application of long RF pulses, therefore in UTE short or half RF excitation pulses are employed. This, however, requires high bandwidth pulses which results in small flip angles for the same RF power.
(iii) Radial center-out sampling of k-space: In UTE sequences, the k-space is sampled using radial trajectories starting from the center of k-space moving outward by synchronized application of gradient coils.
Therefore, in contrast to conventional Cartesian trajectories, there is no need for applying a rewinder-type gradient to return to the beginning of each k-space line.
However, the acquisition of high-resolution UTE images is time-consuming, typically around 6 mins in 3T brain imaging . The acquisition time can be reduced by acquiring data with a coarser resolution, but this would lead to segmentation and learning errors for the task of attenuation correction. Recent attempts for accelerating UTE data acquisition focus on k-space under-sampling in the context of compressed sensing (CS) and parallel MR imaging (pMRI), or the combination of both. Du et al.  proposed to combine UTE with a highly undersampled interleaved multi-echo variable-TE acquisition scheme. The progressively increasing TEs of this technique allowed for spectroscopic imaging of short T2 tissues ranging from 8 μs to 1.5 ms in clinically acceptable scan times. Li et al.  described a 3D CS UTE sequence with hybrid-radial encoding strategy and demonstrated the feasibility of their technique for achieving an acceleration factor of 10. Hu et al.  studied the acceleration of UTE scans through k-space sampling optimization. In this work, the angular sampling rate of 3D radial k-space trajectories was reduced from 100% to 25% for a series of UTE sequences with TEs in the range 0.1-2.3 ms. The resulting scan times on the Philips 3T Achieva MRI system were in the range 172-43 sec.
They demonstrated that high-quality bone-enhanced images can be generated using the UTE sequence with k-space undersampling as low as 25% (acceleration factor of 4) while preserving bone-air contrast at the cost of a minimal increase of noise level. Although radial undersampling violates the Nyquist criterion, spherical symmetry of this sampling pattern helps to preserve the spatial resolution. However, higher undersampling rate can give rise to aliasing artifacts that degrade image quality.
Aitken et al.  proposed to combine CS and sensitivity encoding (SENSE) pMRI for accelerating dual-echo UTE. To investigate undersampling limits, the full radially sampled MRI data of a volunteer (acquired for about 7.5 min on a 3T MRI scanner) was retrospectively sampled by factors of 2-16. Fully sampled images were reconstructed by gridding, while the undersampled echo images were reconstructed by model-based iterative reconstruction incorporating coils’ sensitivity and an L1-sparsity prior. The results showed that there is a good agreement between the fully sampled and undersampled maps with undersampling factors of up to 8 (scan time of 53 sec.). For higher acceleration, the contrast between bone and soft tissue was deteriorated leading to bone-air misclassification during the generation of MRAC maps. MR images of 5 volunteers were prospectively undersampled by an acceleration factor of 8. Similarly, the results were in good agreement with fully sampled images. The sampling pattern in radial UTE sequences; however, does not meet sampling requirements of pMRI techniques, such as SENSE. Therefore, advanced non-Cartesian image reconstruction algorithms are required at the expense of increased reconstruction time. Johansson et al.  studied two non-Cartesian parallel image reconstruction algorithms for the reconstruction of undersampled radial UTE and GRE data of 23 head datasets.
The k-spaces were retrospectively undersampled by factors of 3-30 and reconstructed using conventional re-gridding and non-Cartesian iterative algorithms. The authors reported that for acceleration of up to 5, the acceptable pseudo-CT images can be obtained by the reconstruction methods. The authors concluded that non-
134 Figure 10. High-resolution zero TE image of the head in linear (Top) and inverse logarithmic (Bottom) scale.
Reprinted with permission from Ref .
Cartesian parallel reconstruction methods slightly improved image quality at the expense of increased computational time. As such, gridding reconstruction is sufficient for moderate k-space sub-sampling by which the scan time scan be decreased by a factor of 3 (from 6 to 2 min).
In UTE imaging, the suppression of long T2 components is necessary to separate the bone signal from soft tissue, which is performed by echo subtraction, saturation prepulses, multiple sequences, …etc . As mentioned earlier, the most widely used approach is echo subtraction using a dual-echo UTE. Recently, there has been promising advances in zero time echo (ZTE) sequences for imaging of short T2 structures without the need for time-consuming long T2 suppression methods . Unlike UTE and most MR pulse sequences, in ZTE the frequency-encoding (readout) gradients are set before spin excitation and not ramped down between pulse repetitions. As a result, signal readout starts instantaneously upon excitation leading to a nominal TE of zero.
The gradients are in fact slightly ramped between repetitions, which in turn minimizes eddy currents and the artifacts thereof. In ZTE, RF excitation pulse with very short duration and thus high bandwidth is employed which leads to small flip angles and therefore native proton-density (PD) weighted image contrasts.
Wiesinger et al.  investigated a PD-weighted ZTE sequence for visualization and segmentation of the skull. They developed and optimized a rotating ultra-fast imaging sequence (RUFIS) of type ZT pulse sequence for efficient bone signal enhancement and flat PD response of soft-tissues. An inverse logarithmic image scaling was used to highlight bone and differentiate it from surrounding soft-tissue and air. The flat PD response of soft tissue is in fact important to avoid T1 saturation that would result in the misclassification of long T1 soft-tissues (i.e., CSF and eyes) as bone or air. In contrast to most UTE sequences which mostly explore T2 relaxation time differences, the presented PD-weighted ZTE takes advantage of PD differences. Therefore, it eliminates the application of long T2 suppression methods. The authors studied standard and high-resolution protocols with acquisition times of about 3 and 6 min, respectively. Figure 10 shows a representative high-resolution ZTE and log-scaled images of the head.
Delso et al.  evaluated the feasibility of PD-weighted ZTE imaging for skull segmentation of 15 clinical studies acquired on a trimodality PET/CT-MR system. Quantitative evaluation based on the Jaccard distance between ZTE and corresponding CT bone masks showed improved performance of ZTE over dual-echo UTE by achieving overlap distances over the entire head of 38-63% compared to 47-79%. It was found that contrary to UTE, the presence of tendons on bone masks obtained with ZTE was minor. However, they reported remaining mis-classifications at air/tissue interfaces, i.e. nasal cartilage and inner ear as in UTE.
Lung tissues are characterized by low PD and fast decaying signal due to local magnetic inhomogeneities induced at air-tissue interfaces. Recently, Johnson et al.  demonstrated the feasibility of free-breathing 3D
135 radial UTE whole lung imaging by optimizing this sequence for improved SNR and reduced T2*-induced blurring in conjunction with respiratory gating motion rejection. Radial undersampling and 8-channel pMRI were also used to reduce scan time to 5.5 min. Gibiino et al.  studied a free-breathing 3D RUFIS zero-TE approach for visualization of lung parenchyma and vessels, including prospective and retrospective motion correction. They also demonstrated that high-quality images of lung parenchyma free from blurring and eddy-current artifacts can be obtained using ZTE in less than 6 min.