PET-MRI: a review of challenges and solutions in the development of integrated multimodality imaging

The main challenge in combining PET and MRI has been the development of compact MRI-compatible PET detector technology. Detectors for the 511 keV gamma rays emitted following positron decay are the key component of all PET scanners (Lewellen et al 2008, Peng and Levin 2010). The vast majority of PET systems constructed to date have been based on some configuration of an inorganic scintillation crystal array optically coupled to a small number, typically four, of photomultiplier tubes (PMT). The PMTs detect the light that is emitted when a 511 keV gamma ray interacts with the scintillator and enable the position, energy and time of the interaction to be determined. Unfortunately, the performance of standard PMTs is severely degraded in even a weak magnetic field of several mT (Pichler et al 2006). The main challenge in developing a PET scanner that can operate within or in close proximity to an MRI scanner, where the field within the magnet bore is typically between 0.5 and 10 T, has therefore been either to develop a PET detector that can operate within such a high field, or to find other ways to circumvent this problem. Whilst many different solutions have been implemented for small animal imaging (with less need for good energy and timing resolution), the development of viable human PET-MRI systems capable of simultaneous PET and MRI acquisition has been associated with the emergence of robust and reliable MRI compatible solid state photodetectors such as avalanche photodiodes (APDs) (Renker 2007) and more recently silicon photomultipliers (SiPMs) (Britvitch et al 2007, Roncali and Cherry 2011). These devices have many desirable properties that make them of great interest as alternatives to photomultiplier tubes in many applications (including non-MRI-compatible PET), but as they are essentially insensitive to large magnetic fields, they were required for developing integrated simultaneous PET-MRI with high performance. It is important to note that there are important differences between small animal and clinical imaging regarding the MRI field. Most human PET-MRI scanners have been developed around 3 T MRI magnets. However for small animal systems there is a wide variety of MRI systems depending on the type of studies required, from low field 0.3 T systems up to 11 T, with 7 T being one of the most used field strengths. The lower field systems are more oriented towards anatomical imaging while the stronger fields are used for more challenging imaging tasks.

The majority of small animal and human PET-MRI research projects are aimed at a purpose-designed MRI-compatible PET scanner to be located inside the bore of the MRI magnet, allowing simultaneous PET and MRI acquisition. The choice between sequential and simultaneous configurations is linked to the uncertainties over clinical applications, but for human systems the trend appears to be towards simultaneous acquisition. In this case the limited space for the PET detectors inside the MRI bore adds another important challenge for the detector technology. Besides the potential for simultaneous PET and (f)MRI imaging there are also two practical benefits for a fully integrated clinical system compared to a sequential system (Delso and Ziegler 2009). The first factor is determined by the typical longer acquisition times in MRI compared to CT: a CT scan will last about 15 s–1 min while typically several MRI sequences are acquired in one imaging session and the total acquisition time is about 20–40 min. The most recent evolutions in PET technology (3D, Time-of-Flight and longer axial FOV) (Muehllehner and Karp 2006) have led to a strong increase in effective system sensitivity and reduced the total acquisition time for brain imaging to 3–15 min and for whole body scans within the range 10–20 min. In a sequential system the PET and MRI need to be at a significant distance from each other and therefore the total acquisition time will be the sum of the PET acquisition time and MRI acquisition time. Compared to PET-CT this leads to a very long acquisition time and patient throughput will be limited. In a simultaneous system the acquisition time will be mostly determined by the slowest component, i.e. the MRI acquisition and throughput will be comparable with current MRI scanners. A second factor is the size of the system, MRI systems are already quite large and the addition of a PET in sequential mode requires a room larger than most current imaging rooms. Both factors (acquisition time and required installation space) are less important in small animal imaging. The first commercial sequential PET-MRI systems have recently become available for small animal imaging (Nagy et al 2013). To date there are not yet any commercial simultaneous small animal systems.

Besides the complexity of the integration of hardware, PET-MRI has another major technical challenge compared to PET-CT. Quantitative PET imaging requires the linear attenuation coefficients (at 511 keV) of the object to perform attenuation and scatter correction (Turkington 2000). In a PET-CT system the CT image is not only used for anatomical information but is also easily transformed to linear attenuation coefficients. In a PET-MRI system the derivation of attenuation correction maps from either MRI images, emission data or transmission imaging is complex (Martinez-Möller and Nekolla 2012, Keereman et al 2013). Although several methods have been developed the problem is not yet completely solved and remains an active area of research.

In addition to the many technical obstacles and the attenuation correction problems, another reason why it has taken so long for human PET-MRI systems to develop as quickly as PET-CT is the current absence of definitive clinical applications. Early pre-clinical prototype systems were developed with the aim of acquiring complementary functional and anatomical information concurrently to create a powerful research tool. PET provides highly specific data on molecular pathways, and MRI provides images of anatomy with excellent soft tissue contrast combined with an ever expanding range of functional and molecular measurements. Whilst there is no shortage of potential clinical research applications in humans and there are clear benefits in terms of dose reduction, the benefits of PET-MRI in routine clinical patient management are not yet so clear. The most promising areas are studies where MRI is already the preferred anatomical imaging modality (eg. in the brain, breast and abdomen) where MRI's superior soft tissue contrast make it preferable to CT), and an improved throughput and patient experience is made possible by acquiring PET and MRI images in the same scanning session. Another important limiting factor for integration in clinical routine is the large budget for acquiring a system and the high maintenance costs. Now that human systems are available it will be possible to see exactly where the utility of PET-MRI in patient management and clinical research lies (Pichler et al 2010, Jadvar and Colletti 2014).

In this review the technical problems associated with combining PET and MRI into a single system, and the approaches that have been used to address these, are described, along with outlines of some representative complete imaging systems (see table in Disselhorst et al (2014) for a complete overview of all developed systems). Most of the developments described are PET-related, reflecting the majority of work that has been carried out to date, i.e. MRI-compatible PET scanners operating within an essentially unmodified MRI scanner. Only recently with the advent of commercial human systems have developers started to modify aspects of the MRI in order to achieve close integration and minimise interactions between the PET scanner and the MRI gradient and RF coils within the magnet bore. Now it has been established that the basic compatibility issues are solvable, attention is turning to the operational and quantitative aspects of (in particular, human) PET-MRI systems. Reliable and accurate methods for attenuation correction in PET-MRI are essential and must have comparable accuracy to CT-based methods. Whilst PET data acquisition is essentially passive and limited to the acquisition of data from a single radiopharmaceutical at a time, an MRI imaging session typically involves acquisition of images acquired with many different pulse sequences resulting in many datasets all providing very different information.

Currently available human PET systems for clinical and research imaging all follow a very similar overall configuration comprising an annulus of scintillation detectors (as shown in figure 1) that surround the patient in order to detect pairs of 511 keV gammas in coincidence (Muehllehner and Karp 2006). The inner diameter of the scintillator ring is typically in the range of 85–90 cm, in the axial direction the detectors extend 15–25 cm and with a patient aperture of typically 70–80 cm diameter. The scintillation detectors that make up the ring are known as 'block detectors' each comprising a segmented block of inorganic scintillator coupled to an array of, usually four, photomultiplier tubes (PMTs). Processing of the signals from the PMTs allows the position, energy and time of a gamma ray interaction in scintillator to be determined. The scintillator material most commonly used is Lutetium Oxyorthosilicate (LSO or LYSO) which is chosen because of its high effective Z and density, coupled with very good light output and timing properties, resulting in high spatial resolution, sensitivity and temporal resolution. Detectors have a thickness in the range of 1.5–3 cm to have sufficient stopping power at 511 kev. The most recent systems now have detectors with good energy resolution (11–12 percent) for limiting the detection of scattered coincidences originating in the patient. There is also the trend towards excellent time resolution. This allows the accurate measurement of the time difference between the arrival of both photons. With this information the position of the annihilation can be localised along the line joining the two detection positions. The uncertainty in position is proportional to the uncertainty in measuring the difference in gamma arrival times. A typical 'time-of flight' resolution of 500 ps results in a positional uncertainty of 7.5 cm, and this additional information is used to improve the quality of the reconstructed image (Karp et al 2008). Depending on the length of the object a typical PET acquisition requires one or more (step and shoot acquisition) bed positions. The majority of studies are for oncological investigations and will acquire an image of the whole torso. The most recent scanners have high sensitivity enabling the acquisition of a whole body in under 20 min. For brain imaging one bed position will cover the whole object, which also enables dynamic imaging with time frames below one min.

Zoom In Zoom Out Reset image size

The majority of small animal PET scanners also have a cylindrical geometry, but there is a larger variety in the dimensions. The systems typically have a diameter in the range of 10–20 cm enabling the acquisition of mice and rats. The axial length typically has a range from 3–12 cm. Compared to human systems, thinner scintillators are used (0.5–1.5 cm). Excellent energy and timing resolution are less important for small animal imaging as the objects are small, resulting in limited scatter and no benefit of time-of-flight PET.

A whole body MRI scanner (Carpenter and Williams 1999, Blamire 2008) has the following key components: the main magnet, the gradient coils and RF coils for sending and receiving the signal. The magnet will have a field strength ranging from to 0.2 T–7 T. Nearly all current clinical systems use a superconducting magnet of 1.5 T–3 T. The most important criterion for the magnet is the need to maintain a good homogeneity in the static field B₀ of typically below 4 ppm over a spherical region of 50 cm diameter which effectively defines the field of view. The image SNR nominally increases linearly with B₀, with higher field strengths also providing improved spectroscopic resolution. The diameter of the magnet bore is usually 60–70 cm, and in general good homogeneity becomes harder to achieve as the diameter of the magnet increases. The x, y, z gradient coils are mounted concentrically within the magnet bore. The gradient coils allow the strength of the B₀ field to be varied linearly as a function of x, y and z, so providing spatial localisation of the MRI signal and determining the image spatial resolution and geometrical accuracy. Gradient performance is a key factor for fast dynamic imaging. This is characterised by the maximum gradient amplitude they can achieve (field gradient) as well as by the rate of change of the gradient (slew rate). The high power dissipated by high performance gradients requires air- or water cooling.

Installed concentrically within the gradient set is the whole body transmit coil (as shown in figure 1) that provides uniform RF excitation at the Larmor frequency over the whole field view. Whilst large volume transmit coils can also be used to receive MRI signals, the SNR of the detected signal is considerably increased by the use of small application-specific local receive coils (surface coils) that can be placed very close to the volume of interest (eg spine, neck, knee, wrist, shoulder, breast etc). Receive coils can also be used in the form of phased arrays, that allow imaging of large volumes whilst retaining the high SNR.

Most small animal MRI scanners have a similar setup (Carpenter and Williams 1999, Blamire 2008). The inner bore of the magnet typically ranges from 16–30 cm. Inside this bore the gradient coil set is inserted which reduces the inner bore to a range of 9–20 cm. Inside this volume RF coils are positioned for imaging specific parts of the animal. Typical field strengths are 4.7 T, 7 T, 9.4 T. The most advanced systems are based on fields of 11.7 and 15.2 T. Recently new small animal systems with lower field strengths (range of 0.5 T–3 T) have become available. These systems do not require cryogen cooling and have smaller stray fields, eliminating the need for dedicated MRI rooms with quench pipes.

Whereas most PET imaging sessions acquire essentially a single image, with minimal scope for optimising acquisition parameters, a clinical MRI session usually involves many examinations with different contrast mechanisms (sequences) performed sequentially to examine different aspects of the disease. The parameters of each investigation can be adjusted to optimise the complex trade-offs between spatial resolution, SNR, field of view, speed and contrast, and imaging sessions are usually longer than for PET with 30–45 min being typical.

The combination of MRI and PET scanners is highly challenging (Delso and Ziegler 2009). First of all there is a challenging physical constraint. Both whole body PET and MRI systems have an inner bore diameter of 60–80 cm and an outer bore above one meter. The integration of both modalities in one system capable of simultaneously imaging the same object therefore requires the reduction in size of one of the modalities. The most obvious choice is the reduction of the thickness of the current PET detectors. Once the detector is compact, it can be integrated inside a big bore MRI (inner diameter above 70 cm) without reducing the FOV below the size required for whole body imaging. For small animal systems similar limitations of physical constraint are present (Judenhofer et al 2008): inner clear bore MRI diameters range from 9–30 cm. Although the small animal MRI systems with the largest diameter are a factor 2–3 × more expensive than the ones with a small bore, system integration efforts for small animal imaging were focussed on the large bore systems due to the significant thickness of PET detectors.

There are three main effects (see table 1) that prevent the technology currently used in standalone PET systems being able to function within or near to an operating MRI scanner. The high static magnetic field, quickly changing gradient fields, and radiofrequency signals from the MRI scanner prevent the normal operation of photomultiplier tubes and may induce interference in the front-end electronics of PET detectors.

• (a)  
  Main magnetic fieldPhotomultiplier tubes such as those used in the standard PET block detector design will not function in even very weak static magnetic fields. The magnetic field perturbs the paths of electrons moving from the photocathode down the dynode chain to the anode resulting in a loss of gain. Attempts can be made to shield PMTs using steel or mu-metal but such measures area only effective against relatively weak fields. There are designs for field-insensitive PMTs and some position-sensitive PMT (PSPMT) configurations are less sensitive to magnetics fields, but none of them will tolerate the fields of several Tesla encountered in an MRI system. The solution to this problem has therefore been either to remove PMTs from the high magnetic field region or to replace PMTs with new light detectors that have better tolerance to magnetic fields.
• (b)  
  Gradient fieldsGradient fields are switched rapidly at a frequencies of the order 1 kHz and so, due to the greater skin depth at lower frequencies, are more difficult to shield than the higher frequency RF (120 MHz @ 3 T). These rapidly switching magnetic fields can induce eddy current loops in any conductive components introduced into the magnet bore, including PET circuitry. In addition to signal interference, these can lead to heating and mechanical vibration. Because of the very confined space in the magnet bore, temperature sensitive solid state photodetectors detectors and electronics are likely to be placed very close to the gradient set, and mechanical vibration places much greater emphasis on the robustness of all aspects of detectors and electronics than would normally be necessary. The solution is to redesign the PET readout system and minimise the vibration effects.
• (c)  
  RF interferenceAny electronics situated within the magnet bore may be susceptible to RF interference generated by the MRI transmit coil. This effect is responsible for the drop in PET count rate that is observed in many MRI-compatible PET systems during MRI acquisition. RF shielding is, however, more effective at this higher frequency range and so in principle PET detectors and electronics within the magnet bore can be enclosed in a conducting shield to reduce RF interference.

In an avalanche photodiode the operating voltage is set below the device breakdown voltage so the charge signal obtained is proportional to the original number of electron–hole pairs produced and to the energy deposited in the scintillator by the incident gamma. The gain increases with the applied voltage, however at very high gains there is a deterioration in SNR and the optimum operating range corresponds to a gain of only about 50–150, with voltages in the range of 100–1000 V. APDs can have a quantum efficiency (QE) as high as 80 percent (at 420 nm—the peak emission wavelength for LSO) however an APD/LSO combination can typically achieve a temporal resolution only of a few ns (compared with <600 ps FWHM for a PMT/LSO combination) which is inadequate for time-of-flight PET. A major challenge of APDs is that the gain is very sensitive to changes in temperature (∼ >3%/deg C) and voltage (∼ >10% V⁻¹), so these must be very carefully controlled. To read out the very small signals produced, low noise readout electronics is required and this must be placed as close to the detector as possible. Short optical fibres have also been used in combination with position sensitive APDs (just outside the FOV) to position the electronics at a distant position. An example of an APD detector with short optical fibres in shown in figure 5 and one with direct coupling of APD to scintillator in figure 6.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

For these reasons APDs have not achieved widespread use in conventional PET systems, nevertheless, as there is more experience in the use of these devices than for SiPMs, they are used in several current generation small animal MRI-compatible PET scanners and in the simultaneous whole body PET-MRI scanner, the Siemens mMR, which is described below.

If the operating voltage is increased beyond a threshold level, then the creation of charge carriers in the detector results in a Geiger discharge. This signal is very large but is no longer proportional to the energy deposited by the incident gamma. An SiPM (Roncali and Cherry 2011) consists of APDs operated in Geiger mode, however the proportionality with energy is regained by dividing the active area of the detector into an array of very small GAPD elements or 'cells' each of dimensions typically 20 × 20–60 × 60 um². Each cell acts independently so, provided the number of incident optical/UV photons is much less than the total number of cells (so that the probability of one cell to be hit by two photons will be small), then the summed output from all the cells is proportional to the energy deposited in the scintillator. Good energy linearity is important as photodetector outputs need to be compared in crystal identification schemes.

SiPMs exhibit gains comparable to that of a PMT for very modest bias voltages of 30–100 V, so the need for very low noise front end electronics is removed. The high SNR and speed of the Geiger breakdown result in excellent timing properties with a temporal resolution of ∼100 ps having being reported for room temperature measurements on single LaBr3:Ce(5%) crystals individually coupled to SiPMs (Schaart et al 2010). Because of a thinner depletion layer, and because the sensitive area is <100 percent, the photon detection efficiency (PDE, which is more relevant than the QE for SiPMs) is usually lower than can be obtained with APDs, i.e. typically <40% with LSO.

SiPMs can be fabricated into arrays of increasing dimensions and with very high gain, excellent temporal resolution and low bias voltage and promise to address the various shortcomings of APDs. These properties of course make SiPMs very attractive not only for PET-MRI but also for conventional PET, though as for APDs there is a strong dependance of gain on temperature and especially bias voltage in addition to various trade-offs required to maintain acceptable levels of dark counts. A recent development is the digital SIPM (Degenhardt et al 2009) where the signals from individual microcells are summed digitally (i.e. a binary '1' for every cell that discharges) using electronics integrated on the sensor chip. This results in improved performance (though the PDE may be lower) and lower temperature sensitivity. The sensor itself performs some of the functions (e.g. time to digital conversion and trigger logic) that would otherwise have to be performed by an additional readout ASIC (application specific integrated circuit). An example of an analog SiPM detector and module in shown in figure 7 and one based on digital silicon photomultipliers in figure 8.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Experiments using scintillators, light guides and PMTs to demonstrate both the reduction in positron range caused by a strong magnetic field and the feasibility of a simultaneous PET-MRI scanner were reported in 1995 by Christensen et al (1995). Similar experiments had been presented by the same group previously using photodiodes (Hammer et al 1994), however it was thought that, due to the lack of any conducting or magnetic components in the field of view, the lightguide approach would be preferable for the construction of a combined PET-MRI scanner such as had been previously proposed by Hammer (1990). Around the same time, Buchanan et al (1996) presented a (non-imaging) scintillator/light guide combination that was able to perform ¹⁸F-FDG uptake measurements simultaneously with ³¹P spectroscopy of the Langendorf perfused rat heart in a 9.4 T spectroscopy magnet. Scintillation light from a NaI crystal, situated adjacent to the rat heart within the bore of the magnet, was transferred via a long light guide to a magnetic field insensitive PMT situated in a field of ∼50 mT just outside the magnet. The device enabled changes in the PET and MRIS measurements, resulting from interventions such as hypoxia or ischaemia, to be directly correlated with each other in the same experiment. The first version of the microPET small animal PET-only scanner (Cherry et al 1997) utilised short 15 cm optical fibres to transfer scintillation light from LSO crystal arrays to position sensitive photomultiplier tubes, and it was soon realised that by extending the length of the optical fibres, the microPET detector configuration could easily be adapted to work in a magnetic field. Several prototype systems were constructed with long (∼3 m) 2 mm diameter optical fibres linking a single detector ring (5–12 cm in diameter) of LSO crystals to 3 or more PS-PMTs in a low field region (<10 mT) outside the magnet (Shao 1997, Mackewn et al 2010). The total absence of magnetic or conducting components in the field of view resulted in excellent MRI compatibility in a range of MRI scanners (0.2 T–9.4 T), however scintillation light was attenuated by the optical fibres (typically the amount of light was reduced by a factor 5–10) and combined with the single ring configuration this resulted in poor timing resolution, energy resolution and sensitivity.

The optical fibre approach was further developed by Raylman et al (2006) who used optical fibre bundles with a 90 degrees bend in order to accomodate an axially extended PET field of view within a conventional MRI scanner, however more recent approaches using PMT/fibre combinations have also considered other magnet configurations. Yamamoto et al (2010) have constructed an integrated PMT/fibre system where again an axially extended PET field-of-view is obtained by using a 90 degree angled light guide to channel light from the crystal arrays into 75 cm optical fibres. The PET annulus is located in the field of view of a 0.3 T permanent magnet with the fibres passing out of the magnet through a hole in the magnet yoke enabling the PMTs to be located in a 0.3 mT magnetically shielded region immdiately behind the yoke. The slanted light guides still result in ∼90% light loss, and the coarse crystal segmention results in a transaxial PET spatial resolution of 2.9 mm FWHM, however despite the limited capablities of the 0.3 T MRI, this very simple and practical configuration allows excellent access to the animal and negligible PET-MRI interference.

A straightforward way to capitalise on well characterised existing PET and MRI technology is to adopt a sequential PET-MRI configuration. The Mediso nanoScan® PET-MRI imaging system (Nagy et al 2013) comprises a small animal PET scanner (LYSO arrays read out via relatively field insensitive 256-channel PS-PMTs) mounted in line with a compact 1 T permanent magnet MRI and with a common animal holder. Whilst the relatively low field MRI does not have all the capabilities of more highly specified MRI systems, a straightforward replacement of sequential PET-CT with sequential PET-MRI has many advantages for small animal imaging, including dramatically improved soft tissue contrast and reduced radiation dose.

Another approach is to modify the MRI system itself in such a way that PMTs only have to operate in an acceptably low field. In the Cambridge split-magnet system (Hawkes et al 2008, Buonincontri et al 2013), a 1 T superconducting magnet is split in the axial direction leaving an 80 mm gap into which a modified microPET Focus 120 small animal PET scanner (Laforest et al 2007) can be inserted. Optical fibre bundles transfer scintillation light in the radial direction from the scintillator arrays in the magnet bore to PMTs situated in a low field region just outside the magnet—to achieve this, the fibre bundles are extended from 10 cm in the standalone PET scanner design to 110 cm long. In order to create a sufficiently low field for the PMTs the magnet is self-shielded either side of the gap, and mild steel shields are also inserted. With further mu-metal shielding, the PMTs experience a field of just 1 mT and overall there is only a small reduction in PET performance compared with the standalone system. As it is not desirable to place attenuating objects (other than a relatively low-attenuation RF coil) between the object being imaged and the PET detectors, the gradient set is also split (Poole et al 2009). Split magnet configurations like this have previously been used for interventional human MRI systems (Schenck et al 1995), however the configuration does inevitably compromise the gradient performance and the homogeneity of the main magnetic field to a small extent.

Field-cycled MRI systems use combinations of resistive electromagnets with the capability of varying field strengths dynamically. They can produce images of comparable quality to clinical superconducting systems, and allow new types of MRI contrast. Because the magnets can be rapidly switched on and off (i.e. in just 30 ms) combined PET-MRI operation is possible by performing PET and MRI alternately with very little interference between the modalities. In the prototype system described by Bindseil et al (2011), standard PET block detectors are placed in a gap in the 0.3 T polarizing magnet in a similar way to the Cambridge split magnet arrangement, however only minimal mu-metal shielding of the detectors is implemented and no optical fibres are needed to distance the detectors from the high magnetic field (in this system the gradient set is not split). Interleaved PET and MRI acquisitions have been demonstrated with a repetition time (TR) for a combined PET/MRI sequence of 2752 ms with PET acquisition occuring for ∼1/3 of the time.

The split magnet and field-cycling approaches continue to be developed, however they have not been widely adopted, and remain the only examples of the use of non-standard magnet configurations for PET-MRI.

MRI-compatible small animal PET scanners using APDs are exemplified by those first developed at UC Davis (Catana et al 2006) and in Tuebingen (Judenhofer et al 2007, Judenhofer et al 2008). The Tuebingen system is shown in figure 9. The different approaches taken illustrate some of the issues encountered when detectors and electronic components are placed within the magnet bore. The PET imaging annulus is similar for both systems, consisting of a ring of LSO arrays that fit within the 11.8 icm nside diameter gradient coils a of 7 T Bruker small animal MRI system, but the readout and shielding arrangements are implemented differently.

Zoom In Zoom Out Reset image size

The UC Davis scanner detects the scintillation light via 10 cm long optical fibres connected to 14 × 14 mm² position-sensitive APDs which are cooled by nitrogen gas to −5 °C and operate with a bias voltage of −1.66 kV. The fibres allow the APDs, preamplifiers and other readout electronics to be placed outside the RF coil and the linear regions of the gradients in the axial direction (but still within the high field of the magnet bore). This arrangement requires bending of the optical fibres so the axial PET field of view is limited to 12 mm. To exclude all conducting and magnetic components from the field of view, the APDs and electronics are enclosed in concentric cylindrical RF shields, made from a glass micro-fibre reinforced PTFE composite, however there is no shielding in the central MRI imaging volume where only scintillator, fibres and supporting structures are present. The use of position sensitive APDs in the magnet bore introduces some small distortions to the detector flood map and there is some light loss due to the bent fibres, however overall there is no significant degradation in PET or MRI performance when using standard MRI pulse sequences. In the Tuebingen (Judenhofer et al 2007) system the optical fibres are eliminated and each crystal array is coupled directly to a 3 × 3 APD array (each element is 5 × 5 mm²) via a thin light guide, so the APDs and electronics (with non-magnetic components) are positioned within the MRI imaging volume close to the RF and gradient coils. Double-sided printed circuit board material coated with 10 µm thick copper provides an enclosure for the whole detector module—the copper layer was only 10 µm thick to avoid MRI image artefacts caused by eddy currents induced in the shielding. The system is operated at room temperature with a bias voltage of −380 V, and a detailed assessment of the various sources of interference (Wehrl et al 2011) showed that simultaneous PET-MRI data could be acquired for a wide range of demanding sequences including fMRI and MRIS, noting just a small reduction in MRI image SNR and a small drift in mean signal intensity (attributed to non-stable temperature control).

Maramraju et al (2011) describe a system with a similar overall configuration to the Tuebingen one. This system is using an APD with direct one-to-one coupling, instead of a block detector. Again, fairly small interference effects were seen, including a reduction in homogeneity of the 9.4 T magnet which was correctable by standard shimming coil adjustments. Sensitivity of the PET to RF excitation was initially addressed by switching off the PET electronics during RF excitation, resulting in deadtime of 6%–28% depending on the MRI sequence. A detailed investigation of shielding options (Maramraju et al 2011) demonstrated that the RF interference, and reduction of the MRI SNR due to eddy currents induced in the shielding, could be reduced to acceptable levels by using a thin segmented copper shield around the RF coil to separate it from the PET electronics, although gating of PET signals was still required for sequences with high RF power.

The systems above have demonstrated that, for most standard MRI sequences, acceptable PET and MRI performance can be obtained by placing PET detector components within the MRI field of view, provided careful attention is paid to the selection of components and shielding, and the majority of subsequent simultaneous systems have adopted this configuration. Nevertheless, the PET performance of these prototypes, particularly in terms of sensitivity which is typically <1%, lags behind that of state-of-the-art PET-only small animal systems with sensitivities of typically 4–10% . The most recent systems from these and other groups therefore feature much longer axial fields of view of 5–7 cm for whole body mouse imaging with improved sensitivity.

The European Hyperimage collaboration has constructed a fully functioning SiPM-based small animal MRI-compatible PET scanner (a picture of the system is shown in figure 10) to operate within a 3 T clinical MRI system (Schulz et al 2009) which allows a large transaxial field of view and provides extra capabilities for translational studies. With a ring diameter of 20 cm and axial field of view of 9 cm the system is designed to provide in a PET performance comparable with standalone systems, though the sensitivity of the first realisation of the system was limited to 0.6% as only 1/3 of the axial field of view was populated with detectors.

Zoom In Zoom Out Reset image size

Each detector comprises a 22 by 22 array of 1.3 × 1.3 × 10 mm³ LYSO crystals coupled by a thin light guide to an 8 by 8 array of 4 × 4 mm² SiPMs. Rather than take the approach of reducing the number of readout channels or removing electronic components from the field of view, the problems associated with a large number of readout channels from the SiPM array has been addressed head on and, in order to minimize potential crosstalk, all analogue signals coming from the SiPM elements are digitized immediately adjacent to the sensor, within the field of view, using a custom designed ASIC which outputs digital energy, timing, and channel information. The detector units, comprising the scintillator, SiPM array and ASIC, are enclosed in RF shielded module and liquid cooled to 22 °C. A custom designed transmit/receive coil is integrated into the gantry—the 16 cm diameter birdcage design is low density with all components moved outside the PET field of view to minimise attenuation artefacts.

Given the many potential sources of interference inherent in this approach, great attention was paid to all aspects of MRI compatibility including the selection and production of non-ferromagentic components, differential signalling for all analogue signals, circuit layout to avoid eddy currents and heating effects and minimising RF from power supplies. These considerations are described in detail by Wehner et al (2014). A small increase in the MRI noise floor was measured when the PET system was operated, however overall only very minimal effects on B₀ and gradient homogeneity or RF related effects could be measured, and likewise no effects of RF or gradient switching on PET were measured. This is attributed primarily to the direct digitization of the SiPM-signals coupled with detailed attention to all potential compatibility issues. Although time-of-flight capability is not currently relevant for pre-clinical PET, individual detector modules demonstrated (Schulz et al 2009) a temporal resolution of 520 ps in a 3 T magnet indicating that this approach is also likely to be suitable for a TOF human MRI-compatible PET scanner.

A second version of the Hyperimage system has recently been constructed using digital silicon photomultiplier (dSiPM) technology (Weissler et al 2012). In addition to the improved spatial and temporal resolution of the detectors, the digital SiPMs are less temperature dependent, have a lower power consumption with respect to the previously used analog SiPM/ASIC combination, and as a separate ASIC to digitise the sensor outputs is not required, and the detector stack can be more compact, all of which are advantageous for MRI integration. The detector modules incorporate liquid cooling and are enclosed in carbon fibre housings which provide effective RF shielding whilst having negligible interaction with the gradients. The sensor is operated at a lower temperature than the analog in the first Hyperimage system (for details check the reference Wehner et al (2014)). Data transfer to and from the PET modules is all performed using specially engineered optical transceivers. The system also includes many means for synchronising data acquisition between PET and MRI, for example MRI sequence data can be inserted directly in to the PET datastream. PET performance degradation was examined under extreme MRI conditions (i.e. intense RF and fast gradient switching). Apart from some minor unresolved issues (sensitivity of the PET electronics to the z-gradients, and an increase in the MRI noise floor during PET acquisition) interactions between the two systems are negligible (Wehner et al 2014). MRI-PET Interference of the digital SiPM based system has recently been investigated in detail by Weissler et al (2014).

Several other groups have constructed small animal systems based on SiPM arrays. The basic detectors configurations are on the whole similar, however different approaches have been taken in particular to transferring signals out of the magnet and shielding arrangements. Hong et al (2012) place short optical fibre bundles between the crystal arrays and sensors. This arrangement results in gaps between the shielded sensor modules thus maximising RF transmission to the subject from the MRI body coil, which can then be used in conjunction with a separate receive-only coil. Kang et al (2011) transfer SiPM charge output signals to a preamplifier located remotely using 300 cm flexible flat cables. PET electronics can then be positioned outside of the 5 Gauss line (1.5 meters away from the magnet isocenter of the MRI. Only non-magnetic PET components, scintillation crystals and SiPM arrays without preamplifiers or subsequent electrical circuits need be placed inside the MRI bore, and no electromagnetic shielding was used to protect the PET components. This group has used the same concept to construct a human brain system(see below)

The major image degrading factors in PET are the attenuation and scatter from the object inside the FOV (Kinahan et al 1998, Turkington 2000). The reconstruction of quantitative PET images requires accurate correction for attenuation of 511 kev gamma rays, which is a large effect. The half value layer thickness for a 511 keV gamma ray in tissue is about 7 cm. To detect a coincidence both photons should not be attenuated and therefore the fraction of gamma ray pairs attenuated from the inner parts of the body can be higher than 90 percent. An accurate correction method is therefore essential both to avoid image distortions and artefacts and to permit accurate regional quantification both for routine clinical studies and for quantitative dynamic studies. If a map of linear attenuation coefficients at 511 keV in the object is available then it is straightforward to perform an attenuation correction. Such a map can be obtained rapidly (in a couple of minutes for a whole body) using the CT scanner component in PET-CT (Kinahan et al 2003), however none of the integrated PET-MRI systems constructed to date incorporate a CT scanner or rotating transmission source, so this information must be somehow derived (Martinez-Möller and Nekolla 2012) from the MRI image and/or any other available sources of information on attenuation like additional sources or PET emission data.

The effect of attenuation and the requirements for the accuracy of the attenuation map depends strongly on the object size. For small animal imaging the effect of attenuation is much smaller and methods like contour based attenuation correction derived from MRI will have sufficient accuracy for most studies (Keereman et al 2012). In this part we therefore only focus on the problem of attenuation correction for human brain and whole body imaging. Three different groups of attenuation correction techniques are currently being investigated for PET-MRI: MRI based, emission based and transmission based (Bezrukov et al 2013, Keereman et al 2013). Within each of these groups several methods with differences in acquisition and data postprocessing are under investigation as illustrated in figure 14 for the majority of studies CT based attenuation correction is used as a reference for comparison. Before going into more detail for these different approaches, it is important to determine the requirements of the attenuation correction method.

Zoom In Zoom Out Reset image size

The main requirement for the implementation of an attenuation correction method in clinical practice is the robustness of the method, i.e. the guarantee that no error that could lead to an incorrect diagnosis will be present in the attenuation map. To enable this, it is required that all attenuating objects (also the ones on the path between the emitting object and detectors) inside the FOV are included in the attenuation measurement. A specific problem for PET-MRI is the attenuation caused by MRI coils and other hardware within the field of view typically positioned close to the object of interest. Such objects contribute to the total attenuation, thus reducing the PET signal, and can result in artefacts if not accurately corrected for.

Any attenuation correction method should also have a reasonable acquisition time: ideally the attenuation map is acquired simultaneously with the PET data (to avoid misalignment) and it should only lead to a minimal increase of the total acquisition time and very limited additional dose to the patient.

Several of the attenuation correction techniques involve a segmentation step. A detailed simulation study by Keereman et al (2011) has determined the number of tissue classes that needs to be segmented to have acceptable errors in the emission reconstruction. Air, lung, soft tissue, spongeous bone and cortical bone should be present, while adipose tissue is not required. There is a large variation in the attenuation coefficients in the lung for different patients (Keereman et al 2008). As this is a large volume in the body, the patient specific linear attenuation coefficients should be obtained to avoid large quantification errors. (Martinez-MÖeller et al 2009, Steinberg et al 2010 and Eiber et al 2011)

Specific requirements will also depend on the area of interest. For head scans detection of lung tissue is not needed but the relative amount of skull to soft tissue will be relatively high. Truncation of arms is not relevant in this case. For torso scans accurate detection of the lungs is required and truncation of the FOV can have a significant effect.

There are several major challenges in the development of an attenuation correction method based on MRI data.

• (a)  
  MRI image intensity values reflect proton density and tissue relaxation properties (Stanisz et al 2005) so there is no direct relationship between the signal measured in MRI and the linear attenuation coefficient at 511 keV. MRI based methods are therefore based on segmentation of the object into different tissue classes which are assigned a predefined linear attenuation coefficient.
• (b)  
  Standard MRI sequences give a very low signal in bones (which is the most attenuating tissue in the body) and lung because of the low proton density and the very short T₂ of these tissues (Stanisz et al 2005). Also air has a very low signal in these sequences. Therefore these different tissues can not be differentiated based on the MRI signal (from standard sequences), whilst their respective gamma attenuation coefficients are very different.
• (c)  
  The field-of-view of the MRI system is much smaller than the field-of-view in PET. This can lead to truncation of arms in patients.
• (d)  
  There is a wide variety in the size, shape and composition of the attenuating objects that can be present in the FOV of the PET-MRI system and these can cause significant attenuation (Delso et al 2010). Therefore it is difficult to make prior assumptions about the object. The advantages and disadvantages of each method are summarized in table 3.

The only potential advantage of MRI based attenuation correction compared to PET-CT is the simultaneous acquisition. In a PET-CT scanner there is a mismatch in acquisition time between the PET emission data (typically acquired as an average image over many respiratory cycles) and CT (which provides a snapshot at one phase of the respiratory cycle). This introduces artefacts at the boundary between the lungs and diaphragm (Kinahan et al 2003). Artefacts may also result from the use of CT contrast agents and in the presence of metallic implants and prosthesis, and PET-CT registration may not be perfect. In a simultaneous system the acquisition of PET and MRI data can be acquired during the same breathing phase, which can minimise these artefacts.

Segmentation based MRI-AC.

Template and atlas based MRI-AC.

Evaluation of MRI-AC on patient data.

An alternative for MRI based attenuation correction is to derive the attenuation information from the data collected by the PET scanner. As the acquisition time in PET is shorter than the typical acquisition time of the MRI part, this method is interesting for patient throughput. Several methods have been developed that rely on emission and/or transmission data. The main difficulties of transmission and emission based methods are clearly different from MRI based methods.

Transmission based attenuation correction.

Emission based attenuation correction.

A wide variety of methods have been developed and tested for solving the attenuation correction problem in PET-MRI. Significant progress has been made but none of the standalone methods is currently able to achieve the same accuracy as CT based attenuation correction in PET-CT. MRI based methods seem to fail for body scans in a relative large number of patient studies due to artefacts and errors in segmentation while transmission based methods deliver attenuation maps of lower quality than CT. Potentially improved methods can be developed in the future by combining sources of information. Salomon et al (2011) used the emission based technique to determine the attenuation coefficients from different anatomical regions. These regions can be derived from MRI images but also have the limitations of data truncation and segmentation errors. A recent paper (Panin et al 2013) already combined the emission data with a simultaneously acquired external transmission source. The updates of the emission and attenuation image are alternated and it was shown that use of an external source improves the reconstruction of attenuation coefficients from emission data.

Several methods have been proposed to correct for the partial volume effect in PET (Rousset et al 2007). PET reconstruction can be modified and improved with prior information derived from MRI (Müller-Gärtner 1992). The collection of co-registered MRI and PET data is easier in a simultaneous system and especially for brain imaging resolution enhancement can result in nice high resolution PET datasets (Vunckx et al 2012). A recent overview of the different methods to improve PET image reconstruction using MRI information was given in Bai et al (2013). Compared to PET-CT image reconstruction however one also needs to account for the effect of the magnetic field on positron annihilation distribution (Kraus et al 2012). The positron distribution will be highly non-isotropic due to the main magnetic field. This can introduce effects at the boundaries of tissues with low and high density (ag. at edges of lungs).

Due to the extended time required to acquire PET images, images of many parts of the body are subject to the effects of patient motion. This may be due to periodic motion, as in the case of the lungs or heart, or non periodic motion as in the case of peristalis of the GI tract, or bulk patient movements during the duration of the PET acquisition. The effects of blurring due to patient motion during the PET acquisition fall into two broad categories—firstly, for small regions of activity blurring of the PET image leads to a reduction in apparent activity due to an effect similar to the partial volume effect, thus reducing quantitative accuracy. Secondly, the reduction in apparent activity for small regions of uptake may result in sufficient reduction in contrast that such regions can not be detected in a noisy background, resulting in reduced sensitivity for detection of small lesions.

External devices (belts, triggering, optical cameras) have been used to derive input for motion correction (Rahmim et al 2007), but motion of the internal organs can only be derived with limited accuracy. Attempts to correct motion using PET data alone is only successfull in regions with high enough uptake to yield sufficient motion information. Motion detection for tiny structures, e.g. arteriosclerotic plaques or small tumors, can prove difficult.

PET-MRI offers the possibility of acquiring high quality anatomical images continuously throughout the PET acquisition at a high dynamic rate, exactly spatially and temporally correlated to the PET emission data. In principle a similar task can be attempted with CT however there are severe limitations of radiation dose, and the acquisition is not simultaneous. The development of MRI based motion correction (Ouyang et al 2013) involves several steps. Firstly, dedicated MRI acquisition protocols are required for sufficiently rapid dynamic imaging sampling of the field of view. In a next step motion fields need to be extracted from the measured data and finally one needs to apply these to the PET data in such a way that a single motion-free PET image can be obtained. The motion fields can be applied either to a series of reconstructed PET images ('reconstruct-transform-average'), or can be applied as part of the reconstruction process itself (Polycarpou et al 2012). Whilst evidence of the effectiveness of MRI-based motion PET correction is only just starting to emerge, it has been suggested that this may be essential in order realise the higher resolution promised in the next generation of PET systems (Polycarpou et al 2014).