Principles of PET/MR Imaging

PET Detectors

Since the beginning of PET development, PMTs have been the devices of choice for detecting scintillation light. However, their strong sensitivity to magnetic fields hinders their use inside an MR scanner. Even the earth’s magnetic field (∼50 μT) affects PMTs, but low fields such as this can easily be shielded by covering the outside glass. However, the static magnetic field used in MR is several orders of magnitude stronger, making shielding a viable option only at a certain distance from the magnet. Figure 1 illustrates the effect of increasing field strengths on the operation of a PMT-based PET detector.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 1.

Effect of increasing magnetic field strength on crystal maps of PMT-based PET detector at 1 mT (A), 3 mT (B), and 5 mT (C). (Reprinted with permission of (79).)

In some hybrid systems, only the scintillators have been placed inside the MR imager. The light is guided to PMTs through optical fibers; the PMTs have been placed outside the magnetic fringe field, at field strengths of less than 500 μT. Simultaneous imaging with this configuration is possible, but approximately 90% of the scintillation light is lost (13). With the loss of scintillation light, the energy resolution is drastically degraded.

To avoid the use of light fibers and their accompanying disadvantages, APDs have been used to replace PMTs because they are unaffected by magnetic fields. Their introduction has paved the way for new hybrid scanner designs that include photon detectors in the main MR magnetic field (14). In some instances, relatively short light fibers between scintillators and APDs have been used to move components that have a low magnetic permeability, such as front-end electronics, out of the FOV to reduce magnetic field disturbances (15). With the scintillator’s low level of light energy flux and the low intrinsic amplification of APDs, a charge-sensitive preamplifier is required to convert the electron charge into a measureable voltage signal. The charge-sensitive preamplifier integrates the APD charge over several nanoseconds to achieve high conversion factors, with the disadvantage of pile-up events at high counting rates. The lower gain in APD-based systems requires more advanced front-end electronics. Therefore, most systems use an application-specific integrated circuit to help meet the electrical requirements, complexity, compactness, and performance. Because of the noise and the time variation in the electron drift during the avalanche multiplication process of APDs, the timing resolution is limited to approximately 1.8 ns in a block readout (16), which hinders the use of APDs in time-of-flight applications.

Approximately 10 y ago, the first Geiger-mode APD (G-APD) arrays were studied for low-light-level detection in high-energy physics. A single photon can trigger a self-sustaining ionization cascade, also called Geiger discharge, analogous to the operation of Geiger-Müller tubes. G-APD arrays feature a gain comparable to that for PMTs; they are therefore commonly known as silicon photomultipliers (SiPM). Other names include solid-state photomultipliers and multipixel photon counters.

The signal generated by a single G-APD does not carry any information about the energy or quantity of incoming particles. In PET scintillation detectors, γ-photon energy information is carried via the number of photons generated by the scintillator rather than by the photon wavelength or energy. To discern the amount of incident light, SiPMs consist of a grid of small G-APDs, or cells, at a level of 100 to 10,000 per mm². All individual cells are connected in parallel, and their signals are combined. The number of cells per surface area increases the dynamic range of the detector, and its output can provide information about the quantity of photons.

The main noise contribution in an SiPM output signal is from the so-called dark counts, which are signals generated by the detector itself. These signals are partly due to the optical crosstalk between cells, but they are mainly thermally generated. The dark counts increase by a factor of approximately 2 every 8°C (17). Also, the SiPM gain is strongly affected by the temperature (18). This is especially relevant for its application in an MR scanner because rapid gradient switching, the application of radiofrequency pulses, and the confined environment can induce large temperature variations. Therefore, temperature control is of particular importance for SiPM detectors in PET/MR.

Because of the high gain and fast response times of SiPMs, front-end electronics can be straightforward and do not necessarily require charge-sensitive preamplifiers and application-specific integrated circuits. Wehner et al. (19) recently showed an insert based on SiPM technology with fully front-end digitizers. The signals from PET detectors are directly digitized and processed within the magnetic field. The single-photon detection capability of G-APDs enables triggering on the first photo electron, allowing for coincidence timing resolutions of 100 ps full width at half maximum or lower (20). Placement of SiPMs in strong magnetic fields has no effect on the energy and time resolutions (21). Because of their favorable properties, SiPMs are likely to replace PMTs, in PET/CT scanners as well, as demonstrated by a system recently introduced by Philips featuring digital SiPMs. More details about the 3 types of PET detectors have been listed by Lecomte (22).

Shielding

PET detector systems require sensitive preamplifiers and front-end electronics, which are also highly sensitive to electromagnetic interferences, such as the radiofrequency pulses generated by MR. To prevent MR sequences or gradient switching from degrading the PET signal quality or from generating false events, the electronic components should be placed at a distance or should be shielded (23). The integration of a PET system in an MR scanner requires a careful design of the radiofrequency shielding to avoid introducing eddy currents. These currents are introduced in conductors by changing magnetic fields, for example, applied gradients. Eddy currents create an additional magnetic field that can induce local magnetic field inhomogeneities as well as affect the gradient speed (slew rate) and the spatial encoding ability of the system (23). Furthermore, the PET components should consist of materials with a low magnetic susceptibility. Most integrated circuits fulfill this requirement, but components such as resistors, inductors, capacitors, or connectors should be equipped with nonmagnetic terminals to reduce artifacts.

Most PET systems do not cover the object or patient entirely; as a result, some of the radioactivity remains outside the FOV. Naturally, events caused by this radiation cannot be resolved because no coincidences will be detected. However, this radiation will increase the rate of scattered and random events. γ-ray shielding can be used to reduce this effect, but there are specific limitations on the materials used inside an MR scanner (24). Shielding materials should clearly be of high density to sufficiently attenuate γ photons, but shielding should have low conductivity to prevent the induction of eddy currents and to avoid influencing the radiofrequency signals. Moreover, the material should have a low magnetic susceptibility to guarantee a homogeneous magnetic field. Materials that fulfill these characteristics are lead and tungsten powders in combination with epoxy to reduce the conductivity (24).

Magnet

PET images have been acquired at different magnetic field strengths, from the 0.15-T permanent magnet setup used by Yamamoto et al. (25) to the 9.4-T systems for preclinical (26) and clinical applications (27). Hybrid systems usually consist of a commercially available stand-alone MR scanner combined with a custom built PET system. A notable exception was built with an only slightly modified, commercially available PMT-based PET system, but it features a 1-T split-magnet specifically designed for combined PET/MR (28). Another noteworthy example operates a PET scanner in a field-cycled MR system (29). During PET acquisition, the magnet is ramped down to enable normal operation of the PMT-based PET detectors.

The additional components required for PET acquisition result in a reduction in the bore size in hybrid systems. For example, the Siemens mMR system is based on a 70-cm, wide-bore MR scanner; however, the installed PET hardware reduces the bore diameter to 60 cm (9). In addition, the added components may distort the main magnetic field in simultaneous systems. A distorted magnetic field can decrease the MR image quality, especially for sequences that particularly rely on field uniformity, such as MR spectroscopy and echo planar imaging. However, several studies have shown that with sufficient shimming, the presence of additional PET hardware does not significantly affect the magnetic field homogeneity (8,9,30).

Gradient and Radiofrequency Coils

The gradient system is an essential component of an MR system. In the integrated or insert-based PET/MR scanners that have been described thus far, the PET detectors, electronics, and shielding reside within the gradient coils. This design can reduce the effectiveness of the gradients, decreasing the signal-to-noise ratio in the MR images (14). Rapid switching of the gradient coils, particularly evident in sequences such as echo planar imaging, produces resistive heat when the coils are energized (31). The heat can cause additional warming of the PET components. Because the operation of PET detectors, such as SiPMs, is strongly temperature-dependent, strict temperature control is required. Thus far, proposed hybrid systems have not used gradient coils specifically designed for combined PET/MR, with the exception of a gradient system for the split magnet proposed by Poole et al. (32). The gradient system and the magnet are split into 2 parts to house the PET detectors.

An integrated radiofrequency body coil can be used for whole-body imaging in MR scanners. However, because only weak radiofrequency signals emanate from the tissues during MR acquisition, and this particular coil is positioned relatively far away, this setup results in a low signal-to-noise ratio. Higher signal-to-noise ratios in MR images can be achieved with radiofrequency receiver coils that are placed closer to the object of interest.

All coils placed in the FOV significantly attenuate radiation. It is worthwhile to optimize the coil design in PET/MR to reduce the attenuation losses. Still, even optimized coils attenuate and scatter radiation to some extent. For accurate quantification, attenuation maps of the coil elements in the FOV should be estimated and used during reconstruction. Because of a loss in counts, the global activity concentration will be underestimated when the coil is not accounted for in the attenuation correction (AC), but errors will be especially prevalent near the coil (33–35). The problem with the coils and other objects in the FOV, such as the bed, is that they provide no MR signal although they attenuate radiation. Coils with a fixed position can easily be measured beforehand, for example, with CT, and used as a template in the AC (34,36). In a similar manner, the bed can be included in the AC. However, coil placement should be fairly accurate because small misalignments could already create image artifacts and incorrect quantification (33,36). Additionally, the CT scan of the coils may have artifacts due to the metal components, which in turn can affect the PET images (34). An even more difficult challenge is attenuation due to flexible coils that can be placed in variable positions and orientations. In most cases, these coils are not included in the AC, but not including them could introduce significant deviations in the activity concentration. Attempts have been made to identify the coils using ultrashort echo time sequences or markers attached to the coil that are visible by both MR and CT (37).

Specially designed, low-attenuation coils in combination with accurate localization techniques can improve quantification of PET images. The standard coils available for MR have never been designed specifically for PET/MR; instead, they have been designed for optimal signal-to-noise ratio. PET/MR coils should have low attenuation for γ photons while still performing adequately. The optimization of radiofrequency coils for use in PET/MR is still a relatively young field, but initial studies show promising results (38). The homogeneity of the radiofrequency field could be influenced by the additional PET components, affecting the flip angle. Only very small radiofrequency field map differences have been reported in preclinical (30) and clinical (8) insert-based PET/MR systems.

Clinical Imaging

In brain imaging, the cranial bone significantly contributes to attenuation of the radioactivity in the brain because most lines of response intersect with the bone. Therefore, MR-based AC methods designed for brain imaging need to account for bone attenuation (44,45). However, it is not possible to delineate bone tissue solely from the intensity of single voxels with standard MR sequences because of its short T2 relaxation times.

With ultrashort echo time sequences, signal can still be acquired from bone and used to create attenuation maps. Proposed methods segment ultrashort echo time brain data into 3 (46) or 4 (45) tissue classes, including bone tissue. Using such methods, a high correlation with the reference PET data using CT-based AC can be achieved, but regions with fine structure such as nasal cavities and cerebrospinal fluid are occasionally misclassified. AC methods for whole-body imaging need to address several challenges. Bone detection with MR is difficult because the current ultrashort echo time sequences are too time consuming to be applied for whole-body acquisitions in the clinical setting. However, unlike for brain imaging, whole-body applications are less affected when bone tissue is not properly discriminated from soft tissue in attenuation maps. Therefore, multiple groups have developed segmentation-based methods for whole-body MR-based AC that do not account for the bone without compromising the quantification accuracy for regions that are not close to osseous tissue. Lung segmentation needs to be addressed because the lung density between patients is highly variable, which can affect the quantification results (47). However, the lung signal intensity is generally low in images acquired with conventional MR sequences.

Martinez-Möller et al. (48) proposed a segmentation-based method that discriminates between air, lungs, fat, and soft tissue with intensity-based segmentation using fixed thresholds and connected component analysis to detect the lungs. The AC for the Biograph mMR system is based on this method. Figure 2 shows an attenuation map that was created on this system and the corresponding PET image for an example patient. Approaches that perform segmentation of 3 classes without distinguishing between fat and nonfat soft tissues have also been developed. The AC for the Ingenuity TF PET/MR system is based on such a method (49).

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 2.

PET/MR image of patient with enteric melanoma metastasis (arrow): MR-based attenuation map (A), T2-weighted short-τ inversion recovery sequence (B), ¹⁸F-FDG PET image (C), and PET overlaid on short-τ inversion recovery image (D). Data were acquired on a Biograph mMR system.

Atlas-based methods predict bone tissue without requiring specialized MR sequences. The attenuation maps can be created by deforming a single template image to match the patient anatomy (50) or averaging multiple coregistered attenuation template datasets (51). These methods can work well in brain imaging, but regions with high anatomic variability, such as the sinuses, prove to be challenging. Machine learning-based approaches offer another way to account for bone attenuation by learning a mapping from the MR to either the linear attenuation coefficient or CT data (44).

Atlas approaches or machine learning methods are more challenging for whole-body imaging than for brain imaging because of the high anatomic variability between patients and differences in the body position. Although segmentation-based methods that do not account for bone generally provide accurate quantification results for soft-tissue lesions (52), such methods can be complemented with bone prediction using an atlas or machine learning–based method.

Finally, the transaxial FOV of MR scanners is smaller than the patient cross section, which can lead to truncation artifacts in the MR images at the arm areas. These artifacts propagate to the attenuation maps created from the MR data if they are not adequately corrected. The truncation artifacts can be corrected using reconstructed PET images to derive the body outline (53) or by applying alternate emission and attenuation reconstruction techniques to estimate the missing parts of the attenuation map (54). Blumhagen et al. (55) proposed a method for extending the FOV via homogenization gradient enhancement, which might be an alternative to truncation artifact reduction without resorting to the PET data. Watson (56) proposed an approach for correcting truncation artifacts using supplemental transmission line sources placed around the periphery, whereas Mollet et al. (57) obtained a full attenuation map using an external radiation source. In addition to truncation artifacts, attenuation maps created from MR data can be affected by susceptibility artifacts in MR images. Metallic objects can cause nearby large image distortions or signal loss and affect the attenuation maps. Other intensity variations in MR images can also cause tissue misclassification, whereby, for example, the lungs may be falsely classified as air (58). A review of various MR-based AC techniques was given by Bezrukov et al. (59).

Preclinical Imaging

PET systems for preclinical small-animal imaging have less stringent attenuation map requirements due to the smaller subject size and shorter lines of response. An evaluation by Konik et al. (60) suggested that a uniform attenuation map computed by segmenting uncorrected PET data could provide sufficient accuracy below 6% for subjects smaller than 4 cm in diameter. Deriving attenuation maps from MR data is potentially more robust than segmenting uncorrected PET data, especially if highly specific tracers, such as ¹¹C-raclopride, are used. In studies with larger animals, additional tissue classes can be considered.