graphic stockBattista Biggio, Giorgio Fumera, Paolo Russu,Luca Didaci, and Fabio RoliHabib Zaidi andMinerva BeckerThe Promise ofHybrid PET/MRITechnical advances and clinical applicationsDuring the last few decades, positron emission tomography (PET)-based molecular imaging has advancedeleg antly and steadily gained importance in the clinical and research arenas. However, the lack of structural information provided by this imaging modality motivated its correlation with structural imagingtechniques such as X-ray computed tomography (CT) or magnetic resonance imaging (MRI), which arewell established in the clinical setting. The additional capability of simultaneous acquisition of PET and MRIdata bridges the gap between molecular and morphologic diagnoses. Since diagnostic imaging methods evolvefrom the anatomical to the molecular level, the mission of multimodal and multiparametric imaging increasingly becomes more essential. Since 2010, whole-body hybrid PET/MRI has been investigated in the clinical setting for clinical diagnosis and staging, treatment response monitoring, and radiation therapy treatmentplanning of a wide range of malignancies. However, quantitative PET/MRI is still challenged by the lack ofaccurate and robust attenuation and motion compensation strategies to enable the production of artifact-freeand quantitative PET images. This article briefly summarizes the historical development of PET/MRI and givesDigital Object Identifier 10.1109/MSP.2015.2482225Date of publication: 27 April 20161053-5888/16 2016IEEEIEEE Signal Processing Magazine May 2016 67

an overview of the state of the art and recent advances in thedesign and construction of clinical systems. Progress in quantitative imaging, including MRI-guided image reconstructionand correction, and p otential clinical applications of this noveltechnology are also discussed.IntroductionPET is considered to be one of the key molecular imagingmodalities enabling noninvasive characterization and quantitative evaluation of a multiplicity of molecular and physiologic biomarkers in vivo at the cellular level in healthy anddisease states, including neurology, psychiatry, cardiology, andoncology. However, PET produces blurred and noisy imagesthat inherently lack the anatomical information required forlocalization of metabolic abnormalities. This limitation hasmotivated the combination of PET with structural imag ing modalities, such as X-ray CT and MRI. Currently, PETis capitalizing and complementing other anatomical imagingmodalities, such as CT and MRI, to address basic research andclinical questions. However, multimodality imaging requiresrobust registration of images generated by various modalities.Initially, multimodality imaging was accomplished throughthe use of software-based image registration (rigid body ordeformable) and fusion to correlate anatomical and molecular information [1]. However, the challenges and inherentlimitations of software-based image registration approachesmotivated the emergence of hardware-based approaches formultimodality imaging. The advent of combined PET/CT andPET/MRI systems, their commercial introduction, and the fastand wide acceptance of the former in the clinic have had a significant impact on patient management and clinical research.However, the latter is still an “embryonic” technology, havingthe potential to become a powerful tool and likely to play apivotal role in clinical diagnosis and research [2], [3].This article reviews the state-of-the-art developments andthe latest advances in hybrid PET/MRI instrumentation alongwith quantitative procedures developed to address the challenges of this modality. An outlook outlining potential promising developments and current and future clinical applicationsof this technology is also discussed.History of hybrid PET/MRIThe history of PET/MRI can be traced back to 1986, whenthe first attempts to perform PET imaging within strong staticmagnetic fields were initiated, motivated by the need to reducepositron range prior to annihilation through magnetic confinement of emitted positrons [4], [5]. Indeed, the static magneticfield of the MRI subsystem influences the trajectory of positrons, causing them to spiral between successive interactionswith matter, thus reducing the in-plane spatial resolution of thePET subsystem. Monte Carlo simulation studies demonstratedthat the use of a magnetic field collinear with a PET scanner’saxis improves the transaxial spatial resolution without impinging on the axial spatial resolution [5]. For instance, Wirrwaret al. [6] reported foreseen improvements in spatial resolutionfor high-energy positron-emitting tracers ranging between6818.5% (2.73 mm instead of 3.35 mm) for 68Ga and 26.8%(2.68 mm instead of 3.66 mm) for 82Rb at a field strength of7 T. Another effect, which has been characterized only veryrecently, is the degradation of the axial spatial resolution owingto the elongation of the positron range distribution along themagnetic field B0 or the so-called shine-through artifact [7].It was reported that this effect might cause severe artifacts inPET images for malignant lesions located close to air cavities,particularly when using high-energy positron-emitting radionuclides (see the section “Pitfalls and Artifacts”).Contrary to the history of PET/CT, which began with thedesign of hybrid systems suitable for clinical use, PET/MRIbegan with systems dedicated to preclinical imaging. Surprisingly, the history of hybrid PET/MRI instrumentationper se can be traced back to 1995, prior to the introduction ofPET/CT [8]. Early designs of MR-compatible PET detectormodules focused on modifying detector blocks of an existing small-animal PET scanner to avoid mutual interferenceby placing photomultiplier tubes (PMTs) at a realistic distance from the strong magnetic field of a clinical MRI unit[9]. For the sake of avoiding or reducing mutual interferencebetween imaging modalities, the coupling of detector blocks,position-sensitive PMTs, and readout electronics located outside of the magnetic field was achieved through long (4–5 m)optical fibers. The main disadvantage was, however, the nonnegligible loss of scintillation light through the long fibers,resulting in a weak signal, which negatively impacts energyand timing resolution, impairs deteriorating crystal identification, and decreases PET signal performance, reducing overallPET performance [10].Although this design concept bears inherent limitations,analogous approaches were adopted in academic settings [11].Other associated approaches based on conventional PMTbased PET detectors included split-magnet [12] and fieldcycled [13] MRI, which rely on more complex magnet designs.In the split-magnet design, an 8-cm gap in the axial directionof a 1-T magnet enables accommodation of the microPETFocus 120 small-animal PET scanner (Siemens Healthcare,Erlangen, Germany) and 1.2-m-long optical fiber bundles[12], making it possible to place the PMTs at very low fieldstrength ( 30 mT). The main advantage of this design is theneed for only minor modifications of conventional PET detectors and associated readout technologies, although the magnetand gradient coil design is more complex and costly comparedwith technologies used on current-generation MRI systems. Inthe field-cycled design, PMTs are assembled into the magnet,although PET data acquisition is barely permitted within shorttime intervals ( 2.5 seconds) when MRI polarizing and readout fields are switched off [13]. The challenges associated withthis design still need to be addressed before a viable hardwarerealization can be achieved. Moreover, the need for electromagnets instead of conventional superconducting magnetsrequires noteworthy compromises.The introduction of MR-compatible readout technologies, such as avalanche photodiodes (APDs) and siliconphotomultipliers (SiPMs), was essential to achieve this goal.IEEE Signal Processing Magazine May 2016

MRI systems were designed in anticipation of the availability of mature and economically viable simultaneous wholebody PET/MRI systems, which appeared later and becamecommercially available. Two design concepts have materialized depending on the configuration adopted for patient bedshuttling from one modality to the other. Systems belongingto the first category include the Ingenuity TF PET/MRI system (Philips Healthcare, Best, The Netherlands), in which acommon sliding/rotating bed transfers the patient from MRI toPET and vice versa [24]. The PET/CT/MR trimodality imaging system (GE Healthcare, Little Chalfont, United Kingdom)consists of commercial PET/CT and MRI systems placed inseparate but nearby rooms, and a specially designed patienttransfer tabletop, docked on both imaging systems, is usedto shuttle the patient from the PET/CT to MRI examinationrooms [25]. A similar design concept dedicated to brain imaging was pursued by Cho et al. [26] by docking a high-resolution research tomograph and 7-T MRI.The concurrent design of hybrid PET/MRI is possiblymore attractive but is technically more challenging becauseit involves addressing many difficulties to deal with spacerestrictions and to avoid interference between the twomodalities. To this end, MR-compatible photodetector technologies that are insensitive to magnetic fields and readoutelectronics producing the least amount of heat radiation haveto be used [10]. In addition, the PET detector modules shouldnot affect the operation of the MRI subsystem through electronic interference with the radio frequency (RF) and gradient coils. In essence, the operation of both modalities shouldnot be affected by their integration, and both subsystemsshould retain their full performance, similar to what can beachieved with two separate PET and MRI scanners.As mentioned in the previous section, recent developments in solid-state detectors have led to the replacementof conventional PMTs by MR-compatible position-sensitiveAPDs and SiPMs for the practical implementation of fullysimultaneous PET/MRI systems. The PET insert concept,consisting of placing the detector ring inside an MRI scanner,Avalanche photodiode-based readout technology was successfully employed on a commercial preclinical scanner [14] andvarious prototypes for small-animal [15] and breast [16] PET/MRI. Small pixelated APDs or SiPMs operated in “Geigermode” and more recent readout technologies, such as analog[17] and digital [18] SiPMs, have been investigated as possiblecandidates for PET/MRI, and their current performance issufficient for the design of combined PET/MRI systems [19],given that the bulk of MRI electronics could be significantlyreduced [20]. Convincing experimental results and in vivomouse images obtained on APD-based PET/MRI design demonstrate the capability for simultaneous PET/MRI [15]. Moreimportantly, experimental measurements confirmed that eachsubsystem performs equally well when the other is on or off,reinforcing that each modality is barely visible to the other.These technological advances motivated additional exploration of the clinical potential of PET/MRI [21].Design considerations of hybrid PET/MRI systemsContrary to sequential PET/CT, where the design concept isstraightforward and consists of putting together two separatemodalities, the design of fully integrated PET/MRI systemsis less obvious. Indeed, such development requires not onlymodifications of the PET subsystem to deal with MR computability but also significant redesign of the MRI subsystem [3],[10], [22]. Basically, two major design concepts for PET/MRIhave emerged: sequential and concurrent [23] (Figure 1). In theformer design concept, a serial arrangement of two separatescanners enables sequential data acquisition of both modalities using a single patient’s bed to transfer the patient from onemodality to the other. Conversely, the latter consists of eitheran MR-compatible PET insert that can be placed with the MRIgantry or a compact integrated system enabling truly simultaneous data acquisition.The sequential design is the more straightforward and byfar the more economical concept, requiring only minor modifications of both subsystems (e.g., shielding the PET detectors)and arranging for a common patient bed. Sequential PET/PETMRIMRIMRIPETRF CoilPETRF CoilCommon Bed(a)Single Bed(b)(c)FIGure 1. Schematic cross-sectional views of potential designs for combined PET/MRI systems: (a) a tandem design with two imagers mounted backto back (similar to that in PET/CT instrumentation) to allow sequential rather than simultaneous acquisition, (b) an insert design with the PET imager inserted between the RF coil and gradient set of the MR imager, and (c) a fully integrated design with two imagers in the same gantry. The RF coil, gradientset, PET imager, and patient bed are shown for all configurations. (Figure adapted with permission from [23].)IEEE Signal Processing Magazine May 2016 69

prototype (called BrainPET) for brain imaging manufacturedby Siemens Healthcare in collaboration with the University ofTübingen in Germany [27]. The system performance was characterized and its suitability for various clinical applicationsassessed at a number of academic institutions. Special attention was paid to the possibilities offered by high-resolutionstructural MRI, including high soft-tissue contrast sensitivityand advanced functional MRI techniques [28]. A sequentialPET/MRI system was also developed to meet the needs ofmolecular and genetic brain imaging by docking separate PETand 7-T MRI scanners together with a shared common bed forinterscanner patient translation [26].Subsequent to early groundbreaking developments, different design concepts of PET/MRI systems have materializedduring the last decade in both academic and corporate settings.Figure 2 shows photographs of current commercial clinicalwhole-body PET/MRI systems with potential design concepts.Table 1 summarizes the main characteristics of clinical PET/MRI systems developed so far.The Ingenuity TF PET/MRI system, with TOF Gemini TFPET and Achieva 3T X-series MRI systems, is one such example, allowing for sequential acquisition of aligned PET andMR images. A number of such systems were deployed worldwide, and the PET subsystem was fully characterized usingwas the first landmark, and a number of studies havedescribed different design trends, focusing particularly onthe integration of small-bore, small-animal PET scannersinside existing clinical MRI scanners. The small diametersof these devices allows them to fit into the MRI system without crowding the MRI gradients.As mentioned previously, fully integrated compact systems combining PET and MRI components in a single apparatus, such as Siemens Healthcare’s Biograph mMR and GEHealthcare’s SIGNA, constitute the most promising designconcept for PET/MRI. The exploitation of the most advancedtechnologies available for both systems is advised to achievethe best performance. For instance, using a PET scannerequipped with time-of-flight (TOF) capability is certainly abonus, as discussed in the following section. In this regard,SiPMs have many advantages compared with other solidstate photodetectors, such as APDs, because they have betterperformance characteristics, including high gain, signal-tonoise ratio (SNR), and timing resolution, enabling the implementation of TOF PET on potential PET/MRI systems.Instrumentation for c linical PET/MRIThe successful design of small-animal PET/MRI systemsspurred the development of clinical systems, with the firstPET/MRIMRIPET(a)PET/CT(b)(c)Patient Transfer Tabletop(d)MRI(e)(f)FIGure 2. (a)–(c) The Philips Healthcare whole-body Ingenuity TF PET/MRI system [in which a turntable patient-handling system facilitates patientmotion between the PET subsystem shown in (a) and the Achieva 3T X-series MRI system shown in (c) for sequential acquisition], the Siemens Healthcare Biograph mMR system, and the GE Healthcare SIGNA PET/MRI system, enabling simultaneous acquisition of PET and MRI data. (d)–(f) The GEHealthcare trimodality (PET/CT and MRI) setup using a dedicated patient transporter tabletop. [(a) and (c) used courtesy of Philips H ealthcare,(b) courtesy of Siemens Healthcare, and (d)–(f) courtesy of GE Healthcare.]70IEEE Signal Processing Magazine May 2016

Table 1. The main features of currently available clinical PET/MRI systems.SystemBiograph mMRIngenuity TFSigna PET/MRITrimodalityBrainPETBrain MGIManufacturerSiemens HealthcarePhilips HealthcareGE HealthcareGE HealthcareSiemens imultaneousSequentialSimultaneousSequentialPET TsLSO/APDsLSO-LYSO/PMTsAxial FOV (cm)25.8182515.719.225.2TOFNoYesYesYesNoNoMRIVerio 3T (modified)Achieva 3TMR750w 3.0T (modified)MR750w 3.0TTrio 3T (modified)Magnetom 7TReference[29][24][17][25][30][26]Adapted with permission from [23].the National Electrical Manufacturers Association (NEMA)NU 2-2007 standard, demonstrating that its performance wasnot compromised by the presence of the strong MR magnet[24]. Most performance parameters were comparable to thosereported for the commercial Gemini TF PET/CT system.The design concept of the concurrent BrainPET wasfurther exploited to build the Siemens Healthcare Biograph mMR whole-body PET/MRI system, which wasalso installed in a relatively large number of institutions[29]. More recently, a simultaneous PET/MRI system(SIGNA) based on MR-compatible SiPMs and enabling the implementation of TOF capability was introduced in themarket by GE Healthcare [17].Most current PET/MRI systems have been tested withina high field and proved to produce PET and MR images thatappear to be free of distortion, confirming the premise thatthe interference between the two systems is almost negligibleand that each modality is practically invisible to the other[17], [24], [27], [29], [30]. Switching clinical workflows toPET/MRI introduces a number of image registration challenges that were not of major concern with traditional PET/CT scanners. These relate to the additional artifacts withinMRI, such as bias fields, the range and number of MRIsequences, and the range of fields of view (FOVs) and orientations of the acquired images [31].During the last decade, hardware and software advanceshave enabled improved localization of the position of annihilation along the line of response. The precise measurementof the difference between the arrival times of the two annihilation photons, referred to as TOF, enables more accuratelocalization of the annihilation point. However, the annihilation point could be located only with limited precisionowing to inherent uncertainty in the detector modules andreadout electronics, causing some ambiguity in the photonarrival times. As such, the incorporation of TOF informationin the image reconstruction process enables improved SNRand tumor detectability in addition to reduction of patientscanning time and/or injected dose, all depending uponpatient size and coincidence time resolution (CTR). TheSNR improves as the CTR decreases, and this improvementbecomes more significant for overweight patients. In a clinical setting, this results in a more homogeneous image quality across different (and increasing) patient sizes and overallyields a much-improved image quality in shorter acquisitiontimes, thus providing the possibility to investigate novelacquisition protocols, such as whole-body dynamic imaging. The SNR gain when using TOF is equivalent to a nonTOF image reconstructed using higher statistics; in this way,adding TOF information to PET increases the sensitivity ofthe scanner. In addition, TOF PET scanners are less sensitive to inaccuracies in normalization and data correctionprocedures, including attenuation compensation [32]. Thefirst commercial TOF PET/MRI scanners using lutetiumoxyorthosilicate (LSO)/lutetium-yttrium oxyorthosilicate(LYSO) crystals and PMT/SiPM photodetectors have atime resolution of 400–600 picoseconds [17], [24]. APDbased hybrid PET/MRI systems, including the BrainPETand Biograph mMR scanners, are not equipped with TOFcapability owing to the poor timing resolution of APDs. ACTR of fewer than 100 picoseconds has been obtained withshort crystals of 3–5 mm [33], [34]. The interaction length of511-keV photons in LSO is 12 mm. As such, achieving a sensible detection efficiency requires 15–20-mm-long crystals.However, the CTR degrades with increasing length owing tothe reduction in the speed of light in the high refractive indexof the scintillator because the position along the length of thecrystal where the interaction of the 511-keV photon occurredis unknown. With advances in detector technology and fastelectronics, a TOF PET/MRI scanner with sub-100-picosecond CTR will likely be possible in the near future (Figure 3).The target in the long term is to attain the physical limit ofspatial resolution for clinical scanners ( 2 mm), and by definition, a target CTR of 20 picoseconds would be required toobviate the need of image reconstruction.Hybrid small-animal PET/MRI is also flourishing in bothacademic and corporate settings, with several prototypesbased on different design concepts and a number of companiesalready offering commercial solutions [15], [22]. The potential benefits of compact and integrated systems were alreadyrecognized, and it is expected that this technology will finda niche in preclinical research, which is well under way [35].Quantitative PET/MRIPET/MRI shows promise for radiotracer uptake quantification via image fusion of molecular and structural data toassist in anatomical localization of functional abnormalitiesIEEE Signal Processing Magazine May 2016 71

TOF Resolution (ps)600585 ps(Gemini, Philips)90580 ps (mCT, Siemens)450 ps (Celesteion, Toshiba)500400105600 ps(Discovery-690, GE)490 ps(Astonish TF, Philips)3002001002002005400 ps (Signa, GE)60Reconstruction Not Needed350 ps(Vereos, Philips) 200 ps(Under Development)201075 100 ps(The Future)4530Spatial Uncertainly (mm)700152-mm Resolution2015202020252030YearFIGure 3. The evolution of TOF resolution performance characteristics of current-generation and future-generation TOF PETscanners. PS: picoseconds.and delineation of regions of interest (ROIs) for quantitativeanalysis. However, there are several challenges underminingthe widespread adoption of this technology, which may, infact, represent inherent limitations. Similar to CT in PET/CT, MRI provides the structural information suitable forimplementation of attenuation compensation techniques andintroduction of a priori anatomical information into imagereconstruction, partial-volume correction, and motion correction schemes. However, contrary to PET/CT, in which CTbased attenuation correction is straightforward, MRI-guidedattenuation correction is challenging and still requires furtherdevelopment [36]. Owing to its clinical relevance and thechallenges faced, the latter issue is addressed in more detailin this article.MRI-guided attenuation correction in PET/MRIThe development of MRI-guided attenuation correction algorithms has received considerable attention during the lastdecade. This was motivated by the lack of space in PET/MRIsystems, precluding placement of external radionuclide sourceswithin the gantry. MRI-guided attenuation correction is, however, still in its infancy and remains extremely challenging forwhole-body imaging. The impact of this limitation on clinicalinterpretation of findings and patient outcome is not yet clear.MRI-guided attenuation correction is complex becauseMRI signal intensity is not correlated with electron density,thus making conversion of signal intensity to attenuation coefficients complicated (Figure 4). MRI-guided attenuation mapderivation consists of locating and mapping various biologicaltissues with different attenuation properties in the body. Thiscan be achieved by one of the three main categories of techniques: 1) MRI segmentation-based techniques, in which thebody is segmented into regions corresponding to tissues/organswith different attenuation properties, followed by assign mentof corresponding linear attenuation coefficients at 511 keV to thesegmented tissues/organs; 2) atlas-based and machine-learningµmap at 511 keVCTMRIMappingMappingFIGure 4. The conversion of CT images (Hounsfield units) to an attenuation map at 511 keV is evident (in the absence of sources of error), whereas theMR intensity level is not directly related to electronic density, which renders the conversion of MR images to attenuation maps less evident comparedwith CT.72IEEE Signal Processing Magazine May 2016

time and susceptibility to artifacts when using a large FOV,which limits their application to only brain imaging [37]. Anumber of studies have shown that ignoring bone might notbe adequate for quantification of osseous lesions with bias inestimation in tracer uptake [standardized uptake value (SUV)],varying between 5 and 15% in most cases but going up to 30%in some cases [40]–[45].The second category of approaches consists of using representative anatomical atlas registration, in which an MRItemplate is registered to a patient’s MRI, and prior knowledgeof the atlas attenuation properties, obtained by registrationto a corresponding CT template combined with a learning algorithm based on the use of support vector machines, isapplied to derive a patient-specific attenuation map [46]. Reliable deformable registration algorithms play a pivotal role inthis approach, and failure of the registration process in thecase of large deformations will produce incorrect results [47].The critical issue is the extent to which the global anatomydepicted by an atlas will predict individual and patient-specific attenuation maps. For this and a few other reasons, mosttechniques proposed so far that belong to this category weredeveloped specifically for brain imaging [48], [49]. Adaptation of these techniques for whole-body imaging applicationsrequired few modifications to be made, consisting mainlyof generating a four-class segmentation of the MR imagesto improve the registration process and optimal selection oftechniques, in which an aligned MR/CT atlas combined witha learning strategy enables the prediction of the pseudo-CTfrom an actual patient’s MR image; and 3) emission-based andtransmission-based algorithms, in which the TOF emission ortransmission data are exploited to derive the attenuation map(Figure 5) [37].Segmentation-based methods are simple to implement andusually require a single and fast MRI sequence. However, theysuffer from limited accuracy in the determination of attenuation coefficients owing to the limited number of segmentedclusters (usually three to five, including air, lungs, fat, soft tissue, and fat/nonfat mixture) and the assignment of theoretical rather than actual patient-specific attenuation coefficients.In these techniques, bones and air pockets are replaced by softtissue, and the variability of attenuation coefficients is ignored,especially in the lungs. Tissues such as bone and lung and various pathological abnormalities with varying attenuations areamong the most challenging in whole-body imaging. Withthe exception of the use of ultrashort echo time [38] and zeroecho time [39] pulse sequences, cortical bone has very lowsignal intensity on conventional MRI sequences and is difficult to distinguish from air cavities and gas in the body. Thesesequences were designed to portray tissues with low protondensity and short T2 relaxation time (e.g., cortical bone andlungs) and, as such, to separate the bone signal from soft tissue.The main drawback of these techniques is the long acquisitionMRIFour-Class MRACAtlas-Based RegistrationEmission-Based Technique (MLAA)FIGure 5. Strategies for MRI-guided attenuation map generation, including the four-class segmentation-based method, atlas-based registration and machinelearning, and MRI-guided emission-based technique (MLAA). MRAC: MRI-based attenuation correction. (Figure adapted with permission from [37].)IEEE Signal Processing Magazine May 2016 73

regions for the learning process and applying postprocessingtechniques to determine the tissue class for which sufficientinformation is available from the MRI. In a more recent contribution, Arabi and Zaidi [50] improved the robustness ofthe aforementioned technique [46] to nonsystematic registration bias and anatomical abnormalities by discarding locallygross misalignment errors from the training and pseudo-CTgeneration process through local sorting of the atlas imagesusing the local normalized cross-correlation criterion as ametric to assess the similarity to the target image prior to providing it to the training step. Despite promising preliminaryresults reported in a number of studies using more advancedapproaches [48], [50], more research is still required to makethe procedure completely automated and suitable for clinicalusage in whole-body PET/MRI.Emission-based techniques form the last category of algorithms and have gained substantial momentum during the lastdecade. They are now recognized as valuable approaches forestimation of the attenuation map in PET/MRI through thesimultaneous estimation of activity and attenuation within amaximum-likelihood (MLAA) framework [51]. However,these techniques suffer from cross-talk, depend on tracer distribution, and are susceptible to counting statistics. The use ofTOF information proved to partially mitigate the cross-talkissue and stabilize the joint estimation problem [52]. It is worthemphasizing that TOF PET is less sensitive to attenuationartifacts than conventional non-TOF PET. Recent advancesin emission-based techniques demonstrated the promise ofan MRI-guided MLAA algorithm for attenuation correctionin whole-body PET/MRI [53]. In this work, the estimation ofattenuation maps takes advantage of a constrained Gaussianmixture model and Markov random field smoothness priorimposed by MRI spatial and CT statis

The Promise of Hybrid PET/MRI Habib Zaidi and Minerva Becker Technical advances and clinical applications graphic stock. 68 IESSI NE May 2016 an overview of the state of the art and recent advances in the design and construction of clinical systems. Progress in quan-