Attenuation Correction for Human PET/MRI Studies

Abstract

Attenuation correction has been one of the main methodological challenges in the integrated positron emission tomography and magnetic resonance imaging (PET/MRI) field. As standard transmission or computed tomography (CT) approaches are not available in integrated PET/MRI scanners, MR-based attenuation correction approaches had to be developed. Aspects that have to be considered for implementing accurate methods include the need to account for attenuation in bone tissue, normal and pathological lung and the MR hardware present in the PET field-of-view, to reduce the impact of subject motion, to minimize truncation and susceptibility artifacts, and to address issues related to the data acquisition and processing both on the PET and MRI sides. The standard MR-based attenuation correction techniques implemented by the PET/MRI equipment manufacturers and their impact on clinical and research PET data interpretation and quantification are first discussed. Next, the more advanced methods, including the latest generation deep learning-based approaches that have been proposed for further minimizing the attenuation correction related bias are discussed. Finally, a future perspective focused on the needed developments in the field is given.

1. Introduction

The integrated positron emission tomography (PET) and magnetic resonance imaging (MRI) human scanners introduced over the last decade have enabled investigators to begin to study the benefits of simultaneously acquiring the complementary information provided by the two modalities. To fully exploit this opportunity, the quality of the datasets collected using integrated PET/MRI scanners must be at least similar to that obtained with the equivalent stand-alone devices. On the hardware side, this required the development of novel types of PET and MRI technology to minimize the potential for mutual interference. On the methodological side, the main challenge in the PET/MRI field has been attenuation correction. Among the several corrections that must be applied to the raw PET data to estimate the radiotracer concentration, attenuation correction has the biggest impact on data quantification. To perform this correction, the attenuation of the 511 keV annihilation photons has to be either directly measured or estimated from the data obtained with another imaging modality. In the former approach used in stand-alone PET scanners, the attenuation along each line-of-response joining two detectors was measured with a rotating transmission source (filled with a positron or gamma-ray emitter). Although this method is considered the gold standard for PET attenuation correction as it measures the attenuation at the same energy as the annihilation photons, acquiring the data was slow and the measurements were noisy (Nakamoto et al 2002). Deriving this information from another modality gained popularity with the introduction of integrated PET and computed tomography (PET/CT) scanners (Beyer et al 2000). Several methods have been proposed for converting the CT’s Hounsfield unit (HU) values to linear attenuation coefficients at 511 keV (Kinahan et al 1998, Burger et al 2002, Bai et al 2003). Although the CT-based approach is not without limitations, it has become the accepted standard for PET attenuation correction. As neither approach is available in integrated PET/MRI scanners, this correction has to be performed using the MRI (and PET) data instead.

Several aspects must be considered when implementing an MR-based attenuation correction method (Fig. 1). The MR signal intensity, which depends on proton density and tissue relaxation times, does not directly reflect electron density, which is relevant for attenuation correction. This makes bone tissue difficult to visualize using the data acquired using conventional MR sequences. Similarly, MR imaging of lung tissue is challenging given its low proton density, susceptibility artifacts at the many air-tissue interfaces, tissue fraction effects, respiratory motion, etc. Subject motion leads to mismatches between the MR and the emission data. The patient’s arms are positioned along the body for PET/MRI studies and thus extend outside the uniform region of the MR field-of-view, leading to truncation artifacts. Foreign objects (e.g. dental implants and prostheses) can cause susceptibility artifacts in the MR images. Intravenous contrast agents often used for clinical purposes can change the contrast between different structures. Over the last decade, the equipment manufacturers and academic groups have proposed methods to address most of these challenges and reduce the impact of the remaining ones.

Figure 1:

Examples of challenges in MR-based attenuation correction: (A) Skull tissue cannot be easily distinguished from frontal sinus air cavities; (B) Lung has low and uniform signal intensity in the MR images acquired with conventional sequences; (C) The subject’s arms are truncated due to the limited MR transaxial field-of-view; (D) Respiration-induced liver motion leads to attenuation-emission data mismatches; (E) The radiofrequency MR head&neck coil array is placed between the subject’s head and the PET detectors; (F) The hip implant leads to susceptibility artifacts; (G) The MR contrast agent changes the signal intensity in various organs (e.g. aorta, kidneys, liver, etc.); (H) Fat-water inversion affects one of the legs (figure originally published in (Glocker et al 2016)); (I) A susceptibility artifact makes the lung appear connected to the background air; (J) Negative values in the prompts sinogram due to scatter overcorrection (figure originally published in (Heußer et al 2017)). The red arrows point to the specific issue or region affected in each case.

In this review, these challenges and their impact on the generation of the MR-based attenuation maps is briefly described, the standard methods supplied by the vendors available for clinical PET/MRI studies and their influence on PET data quantification are presented, the advanced methods proposed for further minimizing the attenuation correction related bias are discussed and a future perspective focused on the needed developments in the PET/MRI field is given.

2. Challenges in performing MR-based attenuation correction

2.1. Bone tissue MR imaging

Bone tissue has a very short T2* relaxation time (0.05–2 ms), which makes it virtually indistinguishable from air cavities in the images obtained using conventional MRI sequences (Fig. 1A). As bone tissue and air have the highest and lowest linear attenuation coefficients, their misclassification can introduce substantial bias in the resulting PET images. The bone-air and bone-soft tissue misclassifications are particularly relevant when imaging the brain (Catana et al 2010a, Andersen et al 2014, Hitz et al 2014), pelvis (Schramm et al 2015), spine and extremities (Samarin et al 2012), but could also affect the PET data quantification in posterior parts of the lungs (Seith et al 2017).

2.2. Lung tissue MR imaging

Lung tissue is also very difficult to image using MRI because of its short T2* relaxation time, low proton density, susceptibility artifacts and respiratory motion (Fig 1B). Additionally, there is large inter- and intra-subject variability in lung density during the respiratory cycle and in different pathological conditions (Marshall et al 2012, Seith et al 2017). An excellent lung-focused review that covers all these aspects and potential solutions in more detail was recently published (Lillington et al 2020).

2.3. Truncation artifacts due to the limited MR field-of-view

Geometrical distortions are present outside the uniform region of the MR field-of-view, which is usually limited to 45–50 cm in diameter (Fig. 1C). These spatial distortions are caused by B0 field inhomogeneity and non-linearity of the gradient field. Consequently, the MR-derived attenuation maps are truncated, leading to bias in PET quantification up to 15% globally and 50% locally, particularly in the thorax, abdomen, and pelvis in the case of larger subjects (Delso et al 2010). Additionally, truncation artifacts can also be introduced if the MR imaging field-of-view is smaller than the body part being imaged. The resulting wrap-around artifacts lead to incomplete attenuation maps and decreased MR image quality. Interestingly, even minimally cropping the tip of the nose can introduce large biases in brain studies performed with PET radiotracers that exhibit extremely high signal in the nasal cavity (unpublished observation). This is obviously not due to the minimal under attenuation correction of the missing tissue but due to substantial over scatter correction when the high counts erroneously assumed outside the incomplete attenuation map are used to scale the estimated scatter sinogram.

2.4. Subject motion

In addition to image blurring, subject motion (e.g. rigid-body head motion, non-rigid body respiratory and cardiac motion, peristalsis, bladder filling, positional readjustment, etc.) introduces a mismatch between the emission and MRI data acquired for attenuation correction purposes, which leads to biased quantification (Yang et al 2018a, Kolbitsch et al 2018, Bousse et al 2017, Fayad et al 2017) (Fig 1D). Although not the focus of this review, integrated PET/MRI scanners offer an unique opportunity to address this issue by using MR-derived motion estimates for PET data motion compensation (Catana 2015).

2.5. MRI hardware in the PET field-of-view

The radiofrequency MR coil arrays are often placed as close as possible to the subject and thus inside the PET field-of-view (Fig. 1E) and can interact with the 511 keV photons before they reach the PET detectors. In addition to redesigning the coil to minimally attenuate 511 keV photons, the equipment manufacturers provide the attenuation maps for the majority of the fixed RF coil arrays routinely used for clinical purposes (e.g. head and neck, spine) as well as for the patient table. These maps were obtained either from transmission or CT scans or generated from the CAD drawings of the coil (Eldib et al 2016, Paulus and Quick 2016). The hardware and human attenuation map are combined automatically for each bed position to generate the final map to be used for attenuation correction. Methods have also been proposed for reducing the bias (e.g. up to 15% in the adjacent regions) introduced by the flexible coils (Paulus et al 2012). A similar approach can be used for deriving and integrating the attenuation maps of custom-made application-specific RF coils (Sander et al 2015, Oehmigen et al 2016, 2018b, Farag et al 2020).

2.6. Foreign objects

The numerous foreign objects (e.g. dental implants, surgical clips and wires, orthopedic screws and plates, prosthetic devices, etc.) that can be present in the subject lead to susceptibility artifacts in the MR images (Hargreaves et al 2011) that propagate as signal voids in the corresponding attenuation maps (Fig 1F). A relatively simple solution to minimizing their effect is to fill the signal void with soft tissue (Ladefoged et al 2013, Lassen et al 2019). Alternatively, the location, shape and linear attenuation coefficient of the implant can be estimated using a joint reconstruction of the activity and attenuation algorithm (Fuin et al 2017). Alternatively, advanced MR sequences have been specifically implemented to minimize susceptibility artifacts (Ai et al 2012, Sutter et al 2012, Zho et al 2013).

2.7. MR contrast agents

As MR contrast agents are often used clinically, the possibility of them influencing the MR-based attenuation correction has been raised. Although their presence does not directly increase the 511 keV photon attenuation, they could lead to intensity changes in the MR images used for generating attenuation maps in integrated PET/MRI scanners (Fig. 1G). As will be discussed in the next sections, many of these methods rely on fixed thresholds or tissue contrasts for determining the contour of the body or segmenting various tissue classes. Studies performed to investigate the impact of ferumoxil, gadobutrol (Lois et al 2012) and ferumoxytol (Borra et al 2015, Muehe et al 2019) reported only relatively minor issues (e.g. stomach or liver tissue misclassification) in very specific scenarios. Liver tissue was also misclassified as lung tissue in the case of one liver transplantation patient due to iron overload likely caused by multiple blood transfusion (Büther et al 2017). The overall conclusion of these studies was that the administration of MR contrast agents does not require special considerations in clinical PET/MRI studies.

2.8. Fat-water tissue inversion

Intensity thresholds are also used to segment the various tissue classes from the MR images. Changes in the relative signal intensities due to raw data acquisition/processing factors can lead to errors in the tissue classification. For example, partial or complete fat-water tissue inversion has been reported in ~8% of the 283 patients who underwent brain and head&neck PET/MRI examinations (Fig. 1H). In the case of brain examination, the complete inversion can lead to a spatially varying bias pattern, with the errors increasing towards the center of the head. Partial tissue inversion can lead to artifacts in the resulting PET images (Ladefoged et al 2014).

2.9. Tissue classification/segmentation errors

Segmentation errors due to artifacts, reduced image quality or inadequate performance of the algorithms used for this purpose can result in misclassification of lung as air (Fig. 1I), liver as lung, air pockets in the stomach or bowel as lung or soft tissue (Keller et al 2013, Attenberger et al 2015). The artifacts caused by blood flow in the lung can lead to soft tissue/lung misclassification. Similarly, the aorta can be misclassified as air due to blood flow.

2.10. Halo artifacts

In the case of studies performed using highly specific radiotracers (e.g. [⁶⁸Ga]PSMA-11), the large differences between the radiotracer concentrations in the urinary bladder (and kidney) and the soft tissue background can lead to over scatter correction (Fig. 1J). Given the non-negativity constraint of the iterative image reconstruction algorithms available on integrated PET/MRI scanners, zero values are enforced in the voxels surrounding these organs, which manifests as a “halo” (or photopenic) artifact (Afshar-Oromieh et al 2014, Heußer et al 2017). Modifications of the scatter correction algorithms such as limiting the maximum scatter fraction to 40% (Heußer et al 2017), increasing the number of axial subsamples in the scatter estimation step and including an additional offset factor in the tail scaling step (Wangerin et al 2018) or using an unrenormalized absolute scatter correction (Lindemann et al 2019a) have been shown to minimize these artifacts.

3. Attenuation correction methods available for clinical PET/MRI

3.1. Vendor-provided MR-based attenuation correction methods

Philips Healthcare

An automatic, three-class MR-based attenuation correction procedure was implemented for the Ingenuity TF sequential PET/MRI scanner (Philips Healthcare, Cleveland, OH, USA) (Schulz et al 2011). Briefly, the MR data are acquired with a free-breathing 3D T1-weighted turbo spin-echo sequence (24 seconds acquisition time per bed position) using the body RF coil in multiple bed positions, each covering 6 cm slabs in the axial direction and with a transaxial field-of-view of 46 cm. To avoid the signal drop-off at the edges of the slabs from propagating in the attenuation map, a “stitching” algorithm was implemented to smooth the images in the axial direction. Next, the outer contour of the body is determined automatically based on a threshold obtained from the Laplace-weighted histogram (Wiemker and Pekar 2001) of all the voxels and using a slice-wise region growing algorithm. To segment the lung, a 2D region growing algorithm (based on a threshold obtained from the Laplace histogram) is applied to each of the 7 equally spaced coronal slices centered on the coronal slice with the maximum body cross-section. The pixels corresponding to the lung are determined as those belonging to clusters satisfying certain criteria (e.g. there should be only two clusters per slice, each with dimensions in a certain range, etc.). As the last step, the whole lungs are segmented using a 3D region growing algorithm (Sensakovic and Armato 2008). Predefined linear attenuation coefficients are assigned to soft and lung tissue (0.096 cm⁻¹ and 0.024 cm⁻¹, respectively) (Schulz et al 2011). An evaluation of this approach against transmission scanning revealed “reasonable quantitative accuracy” with average relative differences less than 10% in most structures except cerebellum, lung, and mediastinum (Schramm et al 2013a).

To minimize the truncation artifacts, potentially affected areas are first determined automatically. Next, the contour of the body is estimated from the non-attenuation corrected PET images. Finally, soft tissue linear attenuation coefficients are assigned to all the areas outside the truncated attenuation map but inside the body contour (Kalemis et al 2013). As artifacts (e.g. residual air cavities, background segmented as soft tissue) were still present in the truncation-corrected attenuation maps, using the attenuation corrected PET images (without truncation correction) and delineating the truncated areas in a volume-oriented fashion was proposed as an alternative to the vendor-provided approach (Schramm et al 2013b).

Siemens Healthineers

In the approach adopted for the Biograph mMR scanner (Siemens Healthineers, Erlangen, Germany), the whole-body is segmented into four tissue classes (background air, lung, fat and soft tissue) from the MR data collected using a 2-point Dixon volume interpolated breath-hold exam (VIBE) sequence (19 seconds acquisition time per bed position). Only the data in the thorax are acquired at breath-hold. In-phase and out-phase as well as water and fat images are generated at each bed position. The voxels corresponding to fat and soft tissue are obtained by applying thresholds to the fat and water images. The lungs are identified by connected component analysis of the internal air cavities. To minimize the misclassification of bone, heart, and aorta as air, a morphological closing filter is applied. Predefined linear attenuation coefficients are assigned to the background air, lung, fat and water classes identified (0, 0.022, 0.085, 0.100 cm⁻¹, respectively). Finally, a 3D Gaussian filter is applied to generate the attenuation map (Martinez-Moller et al 2009) that is combined with the hardware attenuation map during the reconstruction. In the initial evaluation, the largest bias in PET quantification was observed in bone lesions (8%±3% average changes in standardized uptake values, SUVs), followed by neck lesions (4%±2% average SUV changes) and lung lesions (2%±3% average SUV changes) (Martinez-Moller et al 2009). An additional work-in-progress method to generate the head attenuation map that relies on a dedicated MR sequence will be discussed in Section 4.1 as it was not initially approved for clinical use.

The truncation artifact was initially minimized by approximating the missing arms with a pair of cylinders and using the non-attenuation corrected PET images to guide their placement (Delso et al 2010). Subsequently, a modified maximum reconstruction of attenuation and activity (MLAA) algorithm was used to recover the missing data by alternatively updating the activity or attenuation map at each iteration, while maintaining the other one constant (Nuyts et al 1999, 2013). This method was shown to reduce the error in the SUV estimation from 15–50% to less than 5% (Nuyts et al 2010). As the performance of this algorithm is radiotracer dependent, a novel MR sequence was proposed for extending the transaxial field-of-view to 60 cm, an approach called B0 homogenization using gradient enhancement (HUGE) (Blumhagen et al 2013). The MR data collected with the HUGE and Dixon-VIBE sequences are combined to generate more accurate extended attenuation maps (Blumhagen et al 2014). One limitation of this approach was that 20 mm slice spacing was used to minimize the total acquisition time, which led to errors in the body contour estimation. Subsequently, the HUGE acquisition was combined with continuous table motion, which allowed the acquisition of all the data at the isocenter, where the distortions due to B0 inhomogeneity and gradient nonlinearity are known. This reduced the total time required for acquiring the HUGE data on both sides of the body to 40–90 seconds, depending on the size of the subject. A landmark-based registration was implemented to fuse the Dixon- and HUGE-derived attenuation maps. Smaller regional and global bias was noted for this method compared to the MLAA approach (Fig. 2) (Lindemann et al 2017).

Figure 2:

Examples of Dixon-based attenuation map truncation compensation using the HUGE and MLAA approaches and their impact on PET data quantification. Figures originally published in Medical Physics (Lindemann et al 2017).

A model-based segmentation algorithm was later proposed to add bone information to the Dixon-derived attenuation map (Paulus et al 2015). The bone model was obtained offline from 200 Dixon-VIBE MR images and bone mask pairs for each major body bone (i.e. left and right upper femur, left and right hip, spine, and skull). To generate the attenuation map for a new subject, the model MR image is registered to the subject MR image for each major bone individually. In the first stage, a landmark-based similarity registration (Zhan et al 2008) is performed using the out-of-phase and fat MR images. In the second stage, a more accurate intensity-based deformable registration is applied to the in-phase and fat MR images. Finally, the corresponding model bone mask is transferred to the subject space and the continuous-valued bone information is combined with the Dixon-based attenuation map. The first evaluation in 20 subjects showed the model-based method can reduce the bias in bony structures from the −25.5%±7.9% observed with the Dixon-based 4-compartment approach to −4.9%±6.7%. The two methods performed similarly in soft tissue (i.e. ~2.5%±2.5% and ~5%±5% bias in background regions and lesions, respectively) (Paulus et al 2015). For generating the head attenuation map using this model-based method, the linear attenuation coefficient assigned to soft tissue was later changed from 0.1 to 0.098 cm⁻¹. An evaluation in 16 subjects with suspected cognitive impairment showed a −0.3% whole-brain SUV_mean bias, with all but three of the brain regions of interest showing less than 5% bias. The authors hypothesized the larger biases in the frontal pole, rostral middle frontal and pars triangularis regions (10.5%, 6.1%, and 7.0%, respectively) were due to variability in frontal bone thickness and sinus pneumatization (Koesters et al 2016).

One of the remaining limitations of this method was that the Dixon images were not of clinical diagnostic quality as they were acquired in coronal orientation and reformatted in axial orientation with 4.1×2.6×3.1 mm³ resolution. Recently, the generalized autocalibrating partially parallel acquisition (GRAPPA) algorithm was extended using controlled aliasing in parallel imaging results in higher acceleration (CAIPIRINHA) to collect high-resolution (1.9×1.4×3.2 mm³) images in the transaxial orientation. The image quality of the CAIPIRINHA-accelerated breath-hold Dixon-VIBE 3D data acquired in 19 seconds per bed position was deemed “good” (4.2±0.8 on a 1–5 Likert scale) and of diagnostic quality (Freitag et al 2017).

After successfully testing the latest implementations of the model-based and HUGE approaches available in the software version E11P in 51 patients with oncological diseases, the authors concluded these methods are ready for clinical use (Oehmigen et al 2018a).

General Electric Healthcare

For the Signa time-of-flight PET/MRI scanner (GE Healthcare, Chicago, IL), the Dixon MR data are acquired using the built-in body radiofrequency coil arrays and a liver-accelerated volume acquisition (LAVA)-Flex sequence in 18 seconds. At each table position, the body contour is first determined and then tissue classes (e.g. lung and trachea or air pockets in the abdomen) are segmented using a phase-field-based approach. The Dixon fat-water images are next used to assign continuous values in the 0.086–0.100 cm⁻¹ range to the soft tissue voxels. A linear attenuation coefficient of 0.0180 cm⁻¹ is assigned to lung voxels. Truncation correction is performed from the time-of-flight non-attenuation corrected PET images. SUV_max biases of 5.7% and 1.5% were reported in the liver and in the FDG-avid features (Wollenweber et al 2013b).

GE supplies an atlas-based method to generate the head attenuation map from the data acquired using the same sequence. The probabilistic atlas was created from the CT images obtained from 50 subjects. All the images were first registered to the reference using an affine followed by a non-rigid B-spline registration and then averaged to obtain the bone atlas. The generation of the attenuation map for a new subject is performed starting from the in-phase LAVA images. The bone information is transformed from the atlas by performing a two-step registration of the bone-enhanced MR and the atlas-derived skull. The same transformation is applied to the non-segmented atlas to obtain the head contour, soft tissue and sinuses using intensity thresholding. Finally, the pseudo-CT obtained by combining the bone and soft tissue information is converted to an attenuation map (Wollenweber et al 2013a, Sekine et al 2016a). The pseudo-CT method tested in 13 subjects was shown to reduce the bias compared to the CT-based approach to 0.2% (Wollenweber et al 2013a). Similar magnitude whole-brain mean bias (2.19%±1.40%) was observed in a different group of 8 patients, with higher values in regions close to the skull base (e.g. cerebellum, temporal lobe, caudate nucleus, putamen, pallidum, thalamus, etc.) (Sekine et al 2016a).

The additional head attenuation map generation method based on the data acquired using a dedicated MR sequence will be discussed in Section 4.1.

United Imaging Healthcare

United Imaging Healthcare (Shanghai, China) recently introduced another integrated PET/MRI scanner called uPMR 790. A 3D Dixon sequence (TR/TE = 4.87/2.24 ms; flip angle = 10°; matrix size = 512×732; pixel size = 0.68×0.68 mm; slice thickness = 2mm; 5-fold acceleration with compressed sensing) is used for attenuation correction purposes. The data are acquired using the head and neck, body array and spine coils. The acquisition time is 70 seconds per bed position in the abdomen and thorax when using triggered acquisition and 35 seconds per bed position in the other regions. The Dixon in-phase are also used as T1-weighted images for diagnostic purposes.

To generate the head attenuation map, four tissue classes (i.e. air, soft tissue, fat and cortical bone) are segmented from the in-phase and out-of-phase images using a deep neural network (U-Net). For the body, the in-phase and out-of-phase images are used to calculate water- and fat-only images and then segmented into air, soft tissue and fat, lung, and bone. Lung segmentation is achieved through a deep neural network (U-Net) that does not require any manual input. A continuous-valued bone component is also generated using a pretrained deep neural network (ResNet). The linear attenuation coefficients used are 0.02, 0.080, 0.096 and 0.161 cm⁻¹ for lung, fat, soft tissue and skull, respectively (Fig. 3).

Figure 3:

Vendor-provided whole-body attenuation correction method. The in-phase, out-of-phase, water-only, fat-only MR images and corresponding attenuation map are shown for one representative subject. Images courtesy of United Imaging.

The attenuation maps for hardware components inside the PET field-of-view, including patient bed and solid MR coils (i.e. spine and head coils), were generated from CT scans and are included in the correction procedure, while flexible surface coils (i.e. body array coil) are ignored (Liu et al 2019).

3.2. Impact of standard MR-based attenuation correction on clinical and research PET data interpretation and quantification

The studies mentioned in the previous section were aimed primarily at assessing the performance of the MR-based attenuation correction methods from a technical perspective. In this section, examples of studies specifically aimed at studying the impact of attenuation correction on clinical and research data interpretation are discussed. Most of these studies followed a single-injection dual-scan protocol (i.e. PET/CT followed by PET/MR examination without an additional radiotracer administration). Hence, the contribution of other technical or biological factors to the observed changes in semiquantitative values could not always be disentangled from the attenuation correction effects. It should also be noted that the performance of MR-based attenuation correction methods was also assessed indirectly in numerous other studies (not discussed here) aimed at comparing the diagnostic performance of PET/CT and PET/MRI modalities in various clinical scenarios (Spick et al 2016).

Brain

Early studies have suggested the misclassification of bone tissue could have a large impact on PET data quantification in neurological studies as it introduces bias both locally, in the adjacent cortical regions and globally, throughout the whole brain (Catana et al 2010a). Indeed, when bone was misclassified as soft tissue in the first method approved for clinical use (Martinez-Moller et al 2009), several groups reported spatially varying bias (Andersen et al 2014) and region-dependent differences (Hitz et al 2014) in patients evaluated for suspected dementia, as well as substantial differences in kinetic parameters (Lassen et al 2017). Only after the model-based approach was introduced on the Biograph mMR scanner, attenuation correction was deemed clinically acceptable for lesions adjacent to bone structures as well as for tumors located within the brain (Rausch et al 2017).

The atlas-based method implemented on the Signa PET/MRI scanner was evaluated in 8 patients without pathological changes in the brain. The visual comparison of the MR- and CT-based attenuation maps revealed an underestimation of the linear attenuation coefficients in the mastoid processes, temporal bone, and base of the skull. A good agreement between the [¹⁸F]FDG uptake values was observed across the brain. The whole-brain [¹⁸F]FDG uptake was underestimated by 2.19%±1.40%, with larger relative changes observed in structures located in the cerebellum (3.69%±1.43%), and temporal lobe (3.25%±1.42%) (Sekine et al 2016a). The same method evaluated in 50 subjects with cognitive impairment who underwent [¹⁸F]FDG examinations showed a bias of 4.5%±6.1%, with overestimations and underestimations close to the vertex and in the cerebellum, respectively (Sgard et al 2019). In the context of Alzheimer’s disease, the diagnostic accuracy and predictive value of FDG PET after MR-based attenuation correction were assessed using simulated data generated by combining the PET error maps that reflect the differences between the PET images obtained using the atlas- and CT-based approaches from 47 patients with 203 ADNI datasets. Although a minor reduction in sensitivity was reported, the accuracy and specificity of the FDG PET obtained using the atlas-based attenuation correction method were maintained (Sekine et al 2020).

The impact of MR-based attenuation correction on the kinetic parameters obtained from dynamic data was also evaluated. For example, when using the Dixon-based approach (Martinez-Moller et al 2009), the PET kinetic parameters obtained from dynamic [¹⁸F]FDOPA data analyzed using the Gjedde-Patlak graphical method were overestimated by 5–9% and 7–10% when using the cerebellar cortex and the occipital pole as reference regions, respectively (Cabello et al 2019). In a different study, the total and specific volumes of distribution and occupancy of the serotonin transporter were calculated from the dynamic data attenuation corrected using several methods, including the Dixon-based and bone demonstrator (based on (Paulus et al 2015)). Using the former, a −11%±1.7 mean relative difference in the total volume of distribution with respect to the value obtained with the CT-based approach was observed in the temporal lobe. The differences in specific volumes of distribution were less than 9% but varied in the regions investigated. Using the latter approach, the largest relative differences in the total (−6.0%±1.6%) and specific (−11.1%±5.0%) volumes of distribution were observed in the temporal lobe. The largest relative differences in occupancy were observed in the anterior cingulate cortex for both methods (−8.7%±16.0% and −8.3%± 24.9%, respectively) (Rischka et al 2019).

Head and neck

PET/CT and PET/MRI examinations were performed in 14 patients with head and neck cancers after a single injection of [¹⁸F]FDG using the Philips Ingenuity TF PET/MR scanner and the 3-compartment model (i.e. air, lung and soft-tissue) (Schulz et al 2011). Very high correlations were observed between the PET/CT and PET/MRI SUV_max for the primary tumors, lymph nodes and metastatic lesions in the lung, liver and bone. Statistically significant higher values were obtained from the PET/MRI data for lymph nodes and metastatic lesions (Partovi et al 2014).

Carotid PET/MR imaging using the Biograph mMR scanner in patients with increased atherosclerosis risk showed a high correlation between the SUV_mean and SUV_max values obtained with the PET/MR and PET/CT, with a small (−0.14 for SUV_max) but statistically significant underestimation when using the former (Ripa et al 2013). The Dixon-based approach was used for attenuation correction. As pointed by the authors, the MR allowed the delineation of the carotid vessel wall, which enabled a more detailed analysis of the [¹⁸F]FDG accumulation in the vessel wall and the possibility of improving the quantification by performing MR-based partial volume effects correction.

A similar study performed using the Philips Ingenuity TF PET/MRI scanner and the default attenuation correction procedure demonstrated excellent correlation with the values obtained using the PET/CT scanner, with percent differences of 5.4% and 2.7% in the left and right carotid arteries, respectively (Bini et al 2015). Subsequently, the same group demonstrated the feasibility of adding a bone compartment to a continuous-valued version of the Dixon-based 4-compartment attenuation map. Although this approach resulted in the best correlation and lowest mean difference with respect to the CT-based method, the authors concluded the quantification errors due to bone tissue misclassification does not substantially affect the quantification in the carotid arteries and suggested the more straightforward default method is more appropriate for clinical use (Bini et al 2016).

Heart

The Dixon-based 4-compartment approach implemented on the Biograph mMR scanner was evaluated for cardiac PET/MRI studies by several groups. No differences in myocardial [¹⁸F]FDG uptake were observed in 10 myocardial infarction patients (Nensa et al 2013). In a different study, the data acquired with a chest phantom and 30 oncologic patients who underwent separate PET/CT and PET/MR examinations were used for the same purpose. The phantom study revealed a minimal 3% underestimation due to the presence of the phased arrays MR coil. For lung linear attenuation coefficients ranging from 0.0224 to 0.0289 cm⁻¹ (corresponding to HU values ranging from −767 to −699), the difference in myocardium SUV estimation was 2.2%. No statistically significant differences between the mean myocardial SUVs obtained by PET/CT and PET/MR were observed (Lau et al 2017). A similar study performed on the GE Signa PET/MRI scanner in 23 oncological patients showed a minimal 2.08% underestimation of the myocardial tracer uptake when using the default Dixon-based method compared to the values obtained by PET/CT. This bias was further reduced to 1.29% with the addition of time-of-flight information. Excellent correlation with narrow limits of agreement was noted between the two modalities on a segmental basis (Vontobel et al 2015).

The authors of a recent study performed in 20 consecutive heart failure patients who underwent myocardial perfusion and metabolism studies on the Biograph mMR scanner reported susceptibility artifacts in 50%, truncation artifacts in 100%, respiratory misalignment in 90% and tissue inversion in 30% of the patients. Large regional differences in uptake were observed after applying simple corrections for the susceptibility artifacts. Although no significant changes in the clinical assessment were seen in 85% of the patients, false-positive findings were reported in 3 patients when not accounting for these artifacts. The average misalignment due to respiratory motion was 7±4 mm, with misalignment larger than 10 mm leading to a severe underestimation of the myocardial uptake (i.e. up to 291% in the anterior wall) (Lassen et al 2019). In another study, mismatches between the attenuation map obtained using the breath-hold Dixon sequence and the emission data were also reported, leading to truncation of the PET data in the lateral wall. The authors proposed an alternative motion-robust multi-echo MR sequences that allows the generation of attenuation maps with multiple tissue classes (i.e. water, fat, lung and background air) (Robson et al 2020). It should be noted that minimizing the mismatch between the attenuation map and the emission data by performing MR-assisted PET motion correction is possible in an integrated PET/MRI scanner (Kolbitsch et al 2018), although implementing such an approach for patients with irregular breathing patterns or when using breath-hold protocols (common in cardiac MRI studies) is challenging due to the breathing mismatches between the PET and MRI data acquisitions.

Compared to the standard Dixon-VIBE approach, both the HUGE and MLAA-based truncation artifact correction were shown to lead to increases (5.4% ± 2.0% and 8.5% ± 3.4%, respectively) in the PET signal over the left ventricle in 32 patients. A limitation of this study was that the CT-based approach was not used as the gold standard. However, this allowed the authors to minimize the impact of other factors (e.g. different patient positioning for the PET/CT and PET/MR examinations, tracer kinetics in the case of sequential acquisitions, image registration) and focus only on the impact of truncation correction and bone segmentation (Lindemann et al 2019b).

Lung

The performance of the Dixon-based method implemented on the Biograph mMR scanner for the detection and characterization of pulmonary lesions was evaluated in 41 patients with known or suspected malignancy. The same number of lesions (including 10 lesions smaller than 1 cm in diameter) were detected in the PET images obtained with the two modalities. There was a strong correlation between the SUVs of lung lesions and no statistically significant differences between the two modalities (Rauscher et al 2014). In a different study, 121 oncologic patients were analyzed retrospectively. A strong positive correlation between the SUV_max and SUV_mean between PET/MRI and PET/CT was found but both measures were significantly higher for the former (Sawicki et al 2016a). On the other hand, the SUV_max obtained from 108 pulmonary tuberculosis lesions were significantly lower on PET/MR than PET/CT (Thomas et al 2017). A strong correlation between the SUV_max determined using PET/MR and PET/CT in 105 lung metastases (from breast cancer) was reported, with the former values being significantly higher (5.6 ± 2.8 vs 4.9 ± 1.8, p=0.001) (Sawicki et al 2016b).

In a different study, 11 patients with advanced non-small cell lung cancer underwent up to five PET/MRI examinations on the Biograph mMR scanner. This allowed the authors to assess the presence of artifacts in the Dixon-based attenuation maps, their impact on the reconstructed PET images and the scan-rescan SUV agreement. Critical artifacts (i.e. trachea artifact, tissue misclassification, body contour, metal, respiratory and lung border artifacts) that could introduce substantial bias were seen in 30% of the maps analyzed. As truncation artifacts were seen in all the cases, they were not even included in the analysis. The scan-rescan variation observed was within ±20% in 95% of the cases, with higher values in those affected by respiratory motion (Olin et al 2018). 25 lung cancer patients who underwent two examinations on the Biograph mMR scanner were included in another study aimed at investigating the frequency and test-retest (i.e. different time points, same imaging session) repeatability of attenuation correction artifacts. Like the earlier studies, truncation artifacts were presented in all the attenuation maps, while tissue inversion and susceptibility artifacts were seen in 52% and 12% of the attenuation maps, respectively. The susceptibility artifacts were seen in all the retest scans, while tissue inversion artifacts were reproduced in the retest scans in 37% of the patients. Correcting the attenuation maps for the missing bone (Paulus et al 2015) and truncation (Nuyts et al 2013) artifacts had minimal effect on the lung tumor SUVs. Susceptibility artifacts correction (Lassen et al 2019) led to up to 28.7% relative differences in SUV_max, with large test-retest variability (Kuttner et al 2019).

The results of analytic simulations suggest the quantification errors when using the 4-compartment attenuation maps that ignore bone can only minimally be reduced in lung tissue (from −3.4% ± 11.5% to −2.9% ± 7.1%) when incorporating the time-of-flight information (Fig. 4) (Mehranian and Zaidi 2015b).

Figure 4:

Relative change maps demonstrating the bias introduced by the 4-compartment model versus the CT-based approach and the impact of using the time-of-flight information for two patients with high (left) and low (right) body mass indexes. Figure originally published in the Journal of Nuclear Medicine (Mehranian and Zaidi 2015b)

Breast

A comparison of the quantification of breast cancer lesions was performed in 36 consecutive patients who underwent [¹⁸F]FDG PET/CT followed by PET/MRI examinations on the Biograph mMR scanner. No statistically significant differences in SUV_max, SUV_mean or metabolic tumor volume were seen in the 25 primary lesions visualized. SUV_max and SUV_mean were 54% and 26% higher in the PET/MRI than the PET/CT measurements (Pace et al 2014).

It should be noted that dedicated breast MR coils may be needed to obtain the highest MR performance required for the characterization of the primary tumor. A novel 16-channel RF breast coil was designed for and successfully tested with the Biograph mMR scanner in phantoms and 10 female patients diagnosed with local breast cancer (Oehmigen et al 2016).

Abdominopelvic organs

The impact of MR-based attenuation correction on the semiquantitative evaluation of lesions in abdominopelvic organs was evaluated by several groups. For example, 45 patients with abdominal and pelvic cancers underwent a single-injection dual-imaging protocol that included PET/CT and PET/MR examinations, using the Philips TF PET/MR scanner. Strong correlations were observed between the SUV_max and SUV_mean in the 49 lesions observed with both modalities. Both values were significantly higher for PET/CT (11.94±6.96 vs. 11.00±7.65, p=0.004 and 7.72±4.70 vs. 7.04±4.89, p=0.001 for SUV_max and SUV_mean, respectively). The SUV_mean of background organs (liver, spleen, muscle and bone marrow) were also higher for PET/CT (0.97±0.11 vs. 0.61±0.13, p=0.001) (Xin et al 2016).

Studies aimed at comparing the diagnostic performance of PET/CT and PET/MRI in lymphoma patients, reported strong correlation between the SUVs obtained from the PET data corrected using the 4-compartment attenuation correction method and the CT-based approach (Atkinson et al 2016, Giraudo et al 2016, Afaq et al 2017). These results are particularly relevant as a large number of nodal and extra-nodal lesions located throughout the body were included in the analyses.

18 patients with gynecological malignancies completed a single-injection dual-imaging protocol using the Biograph mMR scanner and the 4-compartment Dixon-based approach. Statistically significant differences in SUV_max (21.5±10.8 vs. 16.5±7.1), with decreased values were observed for PET/MRI (Schwartz et al 2018).

The model-based approach (similar to (Paulus et al 2015) but using CAIPIRINHA Dixon VIBE data) was assessed in 26 patients evaluated for prostate cancer. They were first scanned using the Biograph mMR scanner after the administration of ⁶⁸Ga-PSMA-11 and underwent an additional PET/CT examination (Domachevsky et al 2020). Statistically significant higher SUVs were observed after including the bone in the attenuation map in the prostatic lesions (12%, p=0.005) and normal prostatic tissue (8%, p=0.003). There was a trend for higher correlation with the CT-derived values for the 5- vs 4-compartment model in all the regions investigated (e.g. gluteus maximus, sacrum, normal prostatic tissue, intra prostatic cancer, metastatic lesions). Smaller but still statistically significant differences in SUV_max (2.5%, p<0.001) between the two MR-based approaches were reported in a different study in which primary and recurrent prostate cancer patients underwent [¹⁸F]fluciclovine PET/MRI studies. Interestingly, bone registration errors that led to bias in the 4.4–20.2% range were observed near the primary tumor in 64% of the cases (Elschot et al 2018).

Bone

In one of the largest studies aimed at comparing PET/CT and PET/MR for the evaluation of malignant bone lesions, 98 bone lesions (located in the cervical, thoracic or lumber spine, pelvis, upper extremity or shoulder, legs, ribs or sternum, etc.) and 630 regions of normal bone were analyzed. The study was performed using the Biograph mMR scanner and the 4-compartment Dixon-VIBE method. The SUV_mean and SUV_max obtained from the two modalities were significantly correlated, but they were significantly lower (12.4% ± 15.5% for bone lesions and 30.1% ± 27.5% for normal bone) for PET/MR. Based on these findings, the authors concluded that the misclassification of cortical bone does not have a clinical impact on the detection of [¹⁸F]FDG bone lesions, although they recommend using the same scanner for longitudinal studies (Eiber et al 2014). The diagnostic value of the morphological MRI data was considered superior to CT for providing anatomical correlates to the [¹⁸F]FDG lesions. The diagnostic confidence in the detection of bone metastases was also deemed superior for PET/MRI compared to PET/CT after the analysis of 86 [¹⁸F]FDG-positive bone lesions in patients with various malignant tumors (Samarin et al 2015).

Assigning a higher attenuation coefficient to the bone tissue identified as those voxels with a SUV higher than 2 from the [¹⁸F]NaF images led to reductions of SUV_max in pelvic and spinal lesions of 8.8% ± 2.7% and 8.1% ± 1.9%, respectively, compared to the 4-compartment approach (Schramm et al 2015). Similarly, a mean increase in SUV_max of 5.4%±6.4% (range −4.3% to 28.4%) was observed in 69 bone lesions after including the bone model and addressing the truncation artifacts using the HUGE method (Grafe et al 2020).

The results of analytic simulations suggest the quantification errors when using the 4-compartment attenuation maps can be reduced in bone tissue from −21.8% ± 2.9% to −15.3% ± 2.3% when incorporating the time-of-flight information (Fig. 4) (Mehranian and Zaidi 2015b).

4. Advanced attenuation correction methods for use in research PET/MRI studies

4.1. Direct bone tissue imaging using dedicated MR sequences

Several dedicated MR sequences have been proposed for imaging tissues with short T2* relaxation time: ultra-short echo time (UTE) (Reichert et al 2005, Robson and Bydder 2006), short-echo time (STE) and zero echo time (ZTE) (Weiger et al 2013, Wiesinger et al 2016b). The methods proposed for using the data acquired with these dedicated sequences for generating attenuation maps in the head, thorax and pelvis are discussed in this section.

Head

In the case of the UTE acquisitions, the free induction decay is sampled as soon as possible after the radiofrequency excitation (e.g. less than 200 μs) to minimize the signal loss. Early work using dual- and triple-echo UTE sequences (Catana et al 2010b, Keereman et al 2008, Berker et al 2012) allowed the generation of segmented head attenuation maps by identifying the voxels corresponding to bone tissue as those showing the largest intensity differences in the data acquired at different echo times. An approach based on dual-echo UTE data was also provided as a work-in-progress package on the Biograph mMR scanner. Subsequently, combining the dual-echo UTE and T1-weighted MR data using probabilistic atlases allowed the generation of substantially improved segmented attenuation maps (Poynton et al 2014). However, one of the main limitations of early generation UTE-based methods was that only three compartments (i.e. soft tissue, bone, and air cavities) could be identified. An approach to synthesize CT-like images from the UTE data used matching UTE-CT patches from a reference dataset. For each subject UTE patch, several corresponding CT reference patches were selected and combined based on a statistical model (Roy et al 2014). Alternatively, continuous-valued bone linear attenuation coefficients were obtained from the dual-echo UTE images by postprocessing the data obtained at each echo (Cabello et al 2015).

In another approach to generate continuous-valued linear attenuation coefficients for bone, regression analysis was performed to identify a five-parameter sigmoid model to map the parameters derived from the UTE data to HU values (Juttukonda et al 2015). To further reduce the tissue misclassification observed when global thresholds are applied to the UTE-derived parameters, a region-specific optimization of continuous linear attenuation coefficients based on UTE (RESOLUTE) was suggested (Ladefoged et al 2015). Specifically, different attenuation coefficients were automatically assigned to several tissue interfaces and classes (i.e. air, air/tissue, air/low-density bone, soft tissue, cerebral spinal fluid, brain, air/bone, tissue/bone, bone) (Fig. 5). To obtain continuous values for bone tissue, a mapping from R₂* measures (derived from the UTE images (Keereman et al 2010)) to HU values was obtained from 10 subjects. A quantitative evaluation of the method in a different group of 154 patients showed the average error in the brain was 0.1%±2.8% (mean absolute error 3.4%±1.6%), which was substantially better than the performance of the UTE- and Dixon-based approaches provided by the manufacturer.

Figure 5:

UTE-based head attenuation map generation. The attenuation maps generated for a representative subject using the vendor provided UTE- (A), RESOLUTE- (B), and CT-based (C) approaches are displayed in 3 orientations. Figure originally published in Physics in Medicine and Biology (Ladefoged et al 2015).

Continuous-valued attenuation maps were also obtained from dual-echo UTE and T1-weighted MR data using probabilistic atlases (Chen et al 2017), from UTE data by combining Gaussian mixture models with probabilistic tissue maps (Baran et al 2018) or from accelerated UTE and multi-echo Dixon acquisitions (Han et al 2020).

One limitation of UTE-based approaches is the long acquisition time. To address this issue, dual-echo ramped hybrid encoding was proposed to acquire the data in 35 seconds, while maintaining high spatial resolution (1 mm³). Using this approach, PET errors of less than 1% were reported (Jang et al 2018a).

Although promising results have been obtained using UTE data, these sequences are not yet widely available or used clinically. As the signal from bone can also be detected at short echo times (i.e. 0.5–1ms), STE sequences have been proposed as an alternative. These sequences allowed the generation of segmented attenuation maps using simple thresholding (Khateri et al 2013). For increased accuracy, the STE and Dixon data were later combined using a segmentation protocol based on fuzzy C-means clustering and morphological operations (Khateri et al 2015). More recently, advanced level-set segmentation was used to reliably generate segmented attenuation maps only from the MR data acquired at a single STE (Fathi Kazerooni et al 2017).

ZTE sequences were also developed for imaging bone tissue (Fig. 6). Instead of relying on the differences between the signal intensities at two or more echo times, ZTE sequences allow bone identification from a single echo proton density-weighted acquisition (Wiesinger et al 2016b). A careful analysis of the skull segmentation accuracy revealed the ZTE-derived bone masks were more similar to the CT-derived bone mask than previously reported for UTE acquisitions throughout the whole head, although some areas with a mixture of tissues (e.g. sinuses and inner ear) remained challenging (Delso et al 2015). One disadvantage of the first implementation of the ZTE sequence was the relatively long acquisition time (i.e. 172 seconds). Accelerated versions (i.e. less than 50 seconds) subsequently implemented by reducing the image resolution were shown to produce similar (Okazawa et al 2019) or better (Sekine et al 2016c) results than the atlas-based approach provided by the manufacturer. To decrease the tissue misclassification, a sinus/edge correction method consisting of morphological processing and edge detection was implemented. This method that does not rely on anatomical prior knowledge outperformed the standard ZTE- and atlas-based methods in terms of PET quantification accuracy and precision. Of particular relevance, the bias in the cerebellum was reduced from 8.1%±3.5% to 0.6%±2.7% after performing this additional correction (Yang et al 2017). A similar approach was shown to reduce the PET bias to −0.04%±1.68% (as well to allow MR-guided radiotherapy planning with absolute dose difference over the target volume of 0.23%±0.42%, which is well below the acceptance criteria) (Wiesinger et al 2018).

Figure 6:

ZTE-based head segmentation results in a representative subject. The RF bias‐ corrected, inverse log‐scaled images, corresponding segmentation results are shown in the first three rows and 3D renderings of the skull and the head contour are displayed in the last row. Figure originally published in Magnetic Resonance in Medicine (Wiesinger et al 2016b).

Other improvements have been suggested for further addressing the issues noted in the ZTE-based attenuation maps. For example, the relationship between the ZTE intensities and HU values was determined and used to generate continuous linear attenuation coefficients for the bone tissue to account for intra- and inter-subject variability in bone density. Additionally, to minimize the misclassification of the teeth, which have low intensities similar to air in the ZTE images, a teeth mask obtained by registering the Zubal phantom template was applied. Furthermore, the spatial resolution was increased from 2.4 to 1.6 mm to minimize partial volume effects. Interestingly, this method was also successfully used to generate an attenuation map for baboons (Khalifé et al 2017).

The ZTE-based approach was shown to have lower interindividual and inter-regional variabilities and increased accuracy compared to the default atlas-based method even in difficult scenarios such as evaluating the pituitary gland, located adjacent to the nasal passages and sphenoid sinus (Delso et al 2019). It was also successfully used to improve the PET data quantification and minimize the impact of metal artifacts in patients with oral cavity cancer (Tsujikawa et al 2020).

A comparison to transmission data revealed a high correlation between the SUVs obtained using the ZTE- and atlas-based attenuation correction methods provided by the manufacturer, although a 6–8% biased was also noted, suggesting the linear attenuation coefficients used in the MR-based approaches are higher than those obtained from the transmission images (Sousa et al 2018). The ZTE-based attenuation correction approach implemented on the Signa GE PET/MRI scanner is now approved for clinical use.

Thorax

Progress has also been made in optimizing UTE sequences for pulmonary applications (e.g. idiopathic pulmonary fibrosis and cystic fibrosis) (Torres et al 2019) and these methods could be used for MR-based attenuation correction. For example, a method to generate synthetic CT images of the thorax from UTE and Dixon data was proposed (Su et al 2019). As a first step, the air voxels outside and inside the subject are identified by thresholding the sum of the images acquired at the three echoes (0.14, 1.14 and 2.14 ms) and the water-fraction images, respectively. In the next step, fuzzy-C-means clustering is used to assign all the voxels to one of the six clusters based on the combination of the Dixon-water, Dixon-fat and water fraction values. A lung mask is generated next by thresholding the R2* (=1/T2*) images and applying morphologic operations. Finally, fixed HU values are assigned to the voxels in each of the six classes identified corresponding to air, lung, fat, soft tissue, spongeous and dense bone. As this study was performed in healthy volunteers, CT images were not available and reference CT images generated using a model-based method from the XCAT phantom were instead used for validation. The synthetic and reference thoracic CT images had a mean absolute difference of less than 50 HUs (Su et al 2019). Future studies are needed to validate MR-based attenuation correction methods in patients with lung parenchymal changes, especially given the recent development of a collagen-targeted radiotracer and the potential to combine these measurements with MR-derived parameters in idiopathic pulmonary fibrosis patients (Montesi et al 2019, Desogere et al 2017).

Pelvis

The data collected using dedicated MR sequences have also been used to generate pelvis attenuation maps. A patient-based method was proposed to segment the images obtained using an STE sequence into five classes (cortical bone, background air, internal air, fat and soft tissue). As a first step after applying MRI inhomogeneity correction, a spatial fuzzy C-means algorithm was applied to cluster the voxels into the bone-air, soft tissue and fat classes. The cortical bone internal air cavities are separated by analyzing the morphological characteristics of the objects identified in the bone-air cluster (e.g. the pelvis bony structures are symmetric with respect to the midline) (Shandiz et al 2017). Subsequently, the same group proposed using a time-resolved angiography with interleaved stochastic trajectories sequence to acquire higher-quality STE data in the pelvis and a hybrid segmentation protocol that also uses the longer echo time data. Numerical simulations suggest using the resulting segmented attenuation maps for PET attenuation correction would lead to mean relative errors of ~14% in the bone regions, substantially lower than the ~36% bias noted when misclassifying cortical bone (Shandiz et al 2018).

Translating the UTE- and ZTE-based methods to the pelvis was more difficult because of the larger field-of-view. A hybrid ZTE/Dixon-based method was implemented to generate continuous-valued attenuation maps (Fig. 7). Bone tissue was segmented from the ZTE data using automatic (i.e. filtering, rescaling, thresholding and morphologic) operations and manual correction. The normalized ZTE values corresponding to the segmented bone were mapped to HU values based on a model generated from seven co-registered CT and ZTE pelvis datasets. The quantification root-mean-square error was reduced compared to the Dixon-based approach from l1.02% and 7.79% for bone and soft tissue pelvic lesions to 3.28% and 3.94%, respectively (Leynes et al 2017). Preliminary results suggest a ZTE approach could also be used in other parts of the body (Wiesinger et al 2016a).

Figure 7:

ZTE-based pelvis attenuation map generation. Bone segmentation process (a) and 3D renderings of the CT- (b) and ZTE-segmented (c) bone are shown. Figures originally published in Medical Physics (Leynes et al 2017).

4.2. Atlas/template/patch/model-based methods

Atlas/template-based

Methods to generate attenuation maps from the MR images were proposed even before the first integrated PET/MRI human scanners were developed. For example, a template obtained by anatomic standardization of the transmission or CT images (Montandon and Zaidi 2005, Rota Kops and Herzog 2008) was brought into the subject space by registration to the morphological MR images. Local pattern recognition was combined with atlas registration to minimize the impact of image registration errors (Hofmann et al 2008). A multi-algorithm approach has allowed the use of a multimodality optical flow deformable model to warp a CT-based atlas to the patient-specific MR images (Schreibmann et al 2010).

More recently, an SPM-based method was implemented for generating head attenuation maps from a single morphological MRI dataset obtained with the MPRAGE sequence routinely collected in research neurological studies (Izquierdo-Garcia et al 2014). After intensity normalization, the MR images are segmented into six tissue classes using the “New Segment” SPM tool and then registered to a previously created template using a diffeomorphic non-rigid image registration algorithm (SPM DARTEL). The inverse transformation is applied to obtain the pseudo-CT images in the subject space. Using this approach, the voxel- and regional-based quantification errors compared to the scaled CT method were 3.87±5.0% and 2.74±2.28%, respectively. This method was also demonstrated to work in patients with modified anatomy (e.g. glioblastoma patients post-surgery) (Fig. 8) (Izquierdo-Garcia et al 2014). Another approach also uses the “new segment” SPM tool and the ICBM tissue probabilistic atlas to generate a bone probability map from the MPRAGE images. These maps are then postprocessed and combined with the Dixon-based maps to generate the final segmented attenuation map for the subject (Anazodo et al 2015).

Figure 8:

Head attenuation correction in a patient with a brain tumor and skull surgical alternations: MPRAGE images (A), pseudo-CT (B) and CT-based (C) attenuation maps. Figures originally published in the Journal of Nuclear Medicine (Izquierdo-Garcia et al 2014).

The training data could also be used to generate multiple atlases, consisting of pairs of CT and T1-weighted MR images. All the MRIs from the database are registered (affine followed by non-rigid registration) to the target MRI and the resulting transformation is used to map the corresponding atlas CTs to the target subject. The pseudo-CT is synthesized based on local morphological similarity measures that monitor the quality of the match between the target MRI and each of the registered MRIs. The CT value at each voxel is obtained from the atlas CTs by applying a weighted average, with the weights being related to the registration accuracy (Burgos et al 2014). The multi-atlas approach was shown to reduce the bias by 20% compared to the single-atlas method, with significant improvements being reported in the temporal lobe and cerebellum (Sekine et al 2016b). Similarly, the multi-atlas method performed slightly better than the single-atlas one (Izquierdo-Garcia et al 2014) in patients scanned with [¹⁸F]FDG and [¹⁸F]florbetapir, with reported average differences compared to the CT-based approach of less than 2% for both tracers (Burgos et al 2015).

The bias in the estimation of the volumes of distribution and nondisplaceable binding potential of [¹¹C]Cimbi-36 obtained when using the multi-atlas (Burgos et al 2014) approach was also evaluated. Kinetic modeling was performed from the dynamic data generated using the MR- and CT-based attenuation correction approaches with the simplified reference tissue model and the cerebellum as the reference region, and the standard two tissue compartment model. The percent differences in regional uptake values ranged from 1%±8% in the temporal lobe to −7%±−6% in the cerebellum. The percent differences in PET outcome measures were below 6% in all the regions investigated and no statistically significant bias was observed for any of the outcome parameters (Mansur et al 2018).

The impact of using the vendor-provided methods, as well as the dual-UTE (Cabello et al 2015) and multi-atlas (Burgos et al 2014) approaches were evaluated using dynamic [¹⁸F]FDOPA data analyzed using the Gjedde-Patlak graphical method. The PET kinetic parameters obtained using non-vendor provided methods agreed within 2% when using the cerebellar cortex as the reference region (and within 6% when using the occipital pole as the reference region) (Cabello et al 2019).

Other approaches have been suggested for performing the registration of the atlas to the target MRIs or for generating the target CT. For example, the deformed CT atlases were first segmented into three tissue classes, majority voting was used to assign a tissue class label to each voxel and the final value was determined by averaging the HU values of the CT atlases belonging to the majority class for that particular voxel (Mérida et al 2017, Merida et al 2015). Using this approach called MaxProb, the mean bias in FDG quantification was less than 1% for the majority of the brain regions analyzed and the results were better than those obtained using a single-atlas approach (Schreibmann et al 2010). Similarly, using dynamic [¹⁸F]MPPF data, regional errors ranging from −1 to +3% in the nondisplaceable binding potential were noted when using the multi-atlas method (Mérida et al 2017). In a different implementation, a deformable registration algorithm was used to register each of the ten atlas MRI volumes to the target MRI. The resulting displacement fields were applied to the corresponding atlas CTs. However, instead of simple averaging, the deformed CT volumes were iteratively registered to the voxel wise mean obtained at the previous iteration (Sjölund et al 2015).

Patch-based

An obvious limitation of atlas-based approaches is that they require computationally intensive image registration between the atlas(es) and target MRI volumes. As an alternative to intensity-based registration, the HU values can be estimated by identifying similar patterns of intensities (“patches”) in the neighborhood of a particular voxel. Analogous to atlas-based methods, the patch-based approaches use databases consisting of matching MR and CT patches. The patches can be obtained from MR data collected using UTE (Roy et al 2014) or standard MR sequences (Andreasen et al 2015, Torrado-Carvajal et al 2016). The assumption is that after intensity-normalization, the composition of tissues determines the pattern observed in the patches. Instead of selecting a single patch from the database, two “nearest” matching patches can be combined using a Bayesian framework (Roy et al 2014) or the similar patches can be extracted from the database averaged using similarity-weighs to assign the HU value (Andreasen et al 2015). Instead of computing the similarity between the subject and each of the target patches from the database, the search can be restricted to those patches in the neighborhood of a certain voxel (e.g. 11×11×11 voxels). The HU value can then be obtained as a weighted linear groupwise combination of these patches, with the weights reflecting nonlocal similarities between the voxels in the subject and target images (Fig. 9). As the search process is highly parallelizable, the computation time can be substantially decreased by implementing the algorithm in a graphical processing unit (Torrado-Carvajal et al 2016).

Figure 9:

Patched-based pseudo-CT synthesis method. The patient specific MR, CT and pseudo-CT images in three orientations are shown on the left. The corresponding PET images and relative change maps are shown on the right. Figures originally published in the Journal of Nuclear Medicine (Torrado-Carvajal et al 2016).

Model-based

A registration technique was initially proposed to include bone tissue in the attenuation maps by comparing the subject MRI to the CT from a database using similarity metrics (e.g. sex, height, weight, age, body and lung geometry, fat to lean tissue ratio, etc.) (Marshall et al 2013). The subject MRI segmented into three tissue classes (Marshall et al 2011) was registered to the most similar CT using rigid and nonrigid components. Finally, the bones from the registered CT were added to the segmented pseudo-CT. Using this approach, the relative errors observed in bone tissue was reduced from −37% to −8% to −3% to 4% (Marshall et al 2013).

The approach currently implemented on the Siemens Biograph mMR scanner and described in Section 3.1 also belongs to this category (Fig. 10) (Paulus et al 2015).

Figure 10:

Model-based approach for incorporating bone in the whole-body attenuation maps. Figure originally published in the Journal of Nuclear Medicine (Paulus et al 2015).

4.3. Emission-based techniques

A radiotracer-specific approach to minimize the impact of bone tissue misclassification is to segment it from the [¹⁸F]NaF emission data. Assigning a higher linear attenuation coefficient to the voxels segmented from the [¹⁸F]NaF images was shown to reduce the quantification errors in pelvic and spinal lesions (Schramm et al 2015) as well as carotid wall (Karakatsanis et al 2019).

Throughout this review, photon attenuation has been presented only as an undesirable effect. However, the fact that photons are attenuated means the emission data inherently contain information about the attenuation properties of the tissue. Therefore, approaches to simultaneously (or jointly) reconstruct the activity and attenuation have been suggested over the last four decades (Censor et al 1979, Lange and Carson 1984, Nuyts et al 1999, Salomon et al 2011). In the context of PET/MRI, the MR information can be used to guide this joint reconstruction. Combined with the time-of-flight information (also shown to reduce the quantification bias and impact of artifacts (Mehranian and Zaidi 2015b, Boellaard et al 2014)), the joint reconstruction techniques have been shown to improve the accuracy of the MR-based estimation of the attenuation maps in the lung (Mehranian and Zaidi 2015a) and brain (Mehranian et al 2017). The algorithms for MR-based attenuation map generation and joint estimation of activity and attenuation can also be used synergistically. Specifically, the MR-derived information can preferentially be used in soft tissue and the joint estimation be given more weight in regions with unreliable MR signal (e.g. around metallic implants, internal air, bone, lung, etc.). This method was shown to work not only with [¹⁸F]FDG but also with other radiotracers with more specific uptake such as [⁶⁸Ga]DOTATOC and [¹⁸F]Fluoride using data acquired with the GE Signa PET/MR scanner (Ahn et al 2018). A similar reconstruction technique was also proposed to improve the accuracy of the vendor-supplied approaches using non-time-of-flight brain data. The information obtained from the UTE-based attenuation map and the T1-weighted MR image intensity are used with a modified version of the maximum likelihood for transmission algorithm (MLTR) to obtain an attenuation map that is more consistent with the emission data. Furthermore, to minimize the cross-talk in the brain and address the scaling problem, information from the MLTR-derived attenuation map is combined with the T1-weighted MR image intensity (Benoit et al 2016).

A method that performs a joint reconstruction of radioactivity and attenuation from the non-time-of-flight emission data and an incomplete MR-based attenuation map was also proposed to determine the position, shape and linear attenuation coefficient for metallic implants. As mentioned in Section 2, the metallic implants lead to susceptibility artifacts in the MR images that propagate as signal voids in the corresponding MR-based attenuation maps. As the attenuation map has to be estimated only in the signal void and prior knowledge is available about the metal implants (i.e. zero activity and a linear attenuation coefficient higher than that of body tissue), the estimation of the attenuation map from the emission data becomes less ill-posed. This method was successfully tested in 11 subjects with various types of implants (e.g. hip chromium alloy and titanium replacements, femur replacement, titanium spine screws, dental implants) who underwent [¹⁸F]FDG-PET/MRI examinations on the Biograph mMR scanner (Fig. 11) (Fuin et al 2017).

Figure 11:

Implant PET-based attenuation map completion (IPAC) using a joint reconstruction of radioactivity and attenuation algorithm. Dixon-, CT-, and IPAC-derived attenuation maps of a right hip metallic endoprosthesis are shown on the left and the corresponding PET images on the right. Figures originally published in the Journal of Nuclear Medicine (Fuin et al 2017).

While transmission-based techniques are still considered the true gold standard for PET attenuation correction, traditional rotating transmission sources have not been integrated into the PET/MRI scanners due to obvious engineering challenges and the desire to reduce the radiation exposure. As such techniques would be valuable for improving and validating MR-based approaches, several stationary transmission sources have been suggested as alternatives. For example, simulations were performed to demonstrate that an annulus shaped transmission source filled with [¹⁸F]FDG that covers the entire axial field-of-view could be used to acquire transmission and emission data simultaneously in a scanner with time-of-flight capabilities (Mollet et al 2012). The time-of-flight information should allow, in principle, the identification of the events originating from the transmission source, although a final scaling was shown to be required (Defrise et al 2012). This method was subsequently shown to minimize the artifacts present in the MR-based 3-compartment method and reduce the bias in PET data quantification in lung, soft and bone tissue (Mollet et al 2014). Recent improvements in detector time-of-flight and count-rate performance could be used to minimize the cross-contamination of the transmission and emission data, and the dead-time effects and randoms cross-talk, respectively. In the case of scanners without time-of-flight capabilities (e.g. Biograph mMR), a single torus source filled with [¹⁸F]FDG was shown to allow the acquisition of highly accurate transmission images before radiotracer administration (Bowen et al 2016). As another alternative to incorporating a full transmission source in an integrated PET/MRI scanner, targeted positron beams generated from several fixed line sources can provide regional transmission data that can be combined with the estimated MR-based attenuation map using a modified MLAA algorithm (Watson 2014).

Finally, a transmission source that does not require any added hardware is provided by the intrinsic background radiation from Lu-176. This isotope has 2.6% abundance in the naturally occurring lutetium that is used to manufacture the scintillator crystals in current generation PET scanners. The 202 and 307 keV gamma-rays emitted as part of the Lu-176 beta decay could be detected simultaneously with the positron emission events in a time-of-flight scanner (Rothfuss et al 2014).

5. Latest generation deep learning-based attenuation correction methods

Although machine learning approaches have been suggested for generating attenuation maps from the early days of clinical PET/MRI (e.g. pattern recognition combined with atlas registration (Hofmann et al 2008, 2011), Gaussian mixture regression model (Johansson et al 2011), support vector regression technique (Navalpakkam et al 2013), structured random forests (Huynh et al 2016), multi-resolution regional learning (Shi et al 2017), feature matching with learned nonlinear local descriptors (Yang et al 2018b), etc.), the focus in this section will be on the recently introduced deep learning methods that are becoming increasingly popular as they have the potential to address many of the limitations of the previously discussed techniques, particularly in terms of accuracy and speed.

5.1. Head

In one of the first implementations, a 2D deep convolutional neural network adopted from the UNet architecture (Ronneberger et al 2015) was trained to generate synthetic CT images from T1-weighted MR images (Han 2017). Qualitatively, the synthetic and real CT images were remarkably similar, with the largest errors in the difference maps observed at the borders of the bone tissue and air cavities, likely due to high-intensity gradients at the interfaces between tissue classes and imperfect registration of the CT and MR training datasets. The average mean absolute error observed in the 18 test brain tumor subjects was significantly better than that obtained with an atlas-based method (Han 2017).

A different method to generate discrete-valued head attenuation maps from high-resolution contrast-enhanced T1-weighted MR images was developed using a deep convolutional auto-encoder network. The training data consisted of raw MR images and corresponding labels for air, bone and soft tissue obtained from unenhanced CT images. High labeling accuracy was reported for air, soft and bone tissue when comparing the resulting attenuation maps to those generated by segmenting the CT images. When using these maps for attenuation correction, the average errors reported were less than 1% for most brain regions, significantly smaller than those obtained with the Dixon-based 4-compartment approach and the registered CT-based template (Fig. 12) (Liu et al 2018a).

Figure 12:

Deep learning-based head attenuation map generation. The input MR, acquired CT and pseudo-CT images at different axial levels are shown. Figures originally published in Radiology (Liu et al 2018a).

As Dixon MR images are routinely acquired for attenuation correction purposes, a standard U-net architecture was proposed for generating continuous-valued attenuation maps from these data (Gong et al 2018). When ZTE MR images are also available, even more accurate attenuation maps can be obtained from both MR inputs using a modified U-net architecture based on group convolutional modules. The attenuation maps generated with both methods were more accurate than those obtained using the standard methods provided by the manufacturer. The mean relative errors in the reconstructed PET images were smaller than ~3% in all the brain regions analyzed, with the method using multiple contrasts producing the best results (Gong et al 2018). Alternatively, a framework proposed for improving the UTE-based attenuation maps obtained using dual-echo ramped hybrid encoding was shown to reduce the PET quantification errors to less than 1% (Jang et al 2018b). A 3D U-net convolutional neural network was also proposed for generating attenuation map from ZTE data and shown to introduce the smallest bias in PET data quantification when compared to the atlas- and standard ZTE-approaches (Blanc-Durand et al 2019). Another network based on the U-net architecture was shown to produce clinically acceptable attenuation maps in pediatric brain tumor patients, a particularly challenging patient population due to the age-related physiological differences in head shape and conformation and the presence of disease-specific morphological changes in the brain (Ladefoged et al 2019).

Alternatively, instead of generating the attenuation maps from the MR data, a deep convolutional encoder-decoder network was trained to generate pseudo-CT images from the uncorrected [¹⁸F]FDG-PET images (Liu et al 2018b). Using this approach, the mean PET image errors compared to the CT-based correction were below 2% for all the regions analyzed. Interestingly, the patients used for training the network were scanned with their arms either up or down, which suggests the method could be used to generate attenuation maps for other body regions, although this would likely require different models. A concern when using these architectures is overfitting. As a solution, generative adversarial networks (Goodfellow et al 2014) have successfully been used to translate PET into CT images (as well as for other tasks such as MR motion correction and PET image denoising). This framework consists of two distinct networks that are trained in competition, one to generate “fake” images that look realistic and the second one to binary classify the real and synthetic images. Qualitative and quantitative analyses showed the CT images generated using this framework are very similar to the ground truth (Armanious et al 2018). A slight underestimation (below 5%) of the [¹⁸F]FDG-PET SUVs was observed when performing the correction using the resulting attenuation maps (Armanious et al 2019). To regularize the synthetic CT generation procedure, an additional segmentation generative adversarial network component can be included in the model. This added network segments the synthetic CT into four tissue classes based on predefined thresholds, which leads to higher accuracy in deriving anatomical structures, albeit at the price of slightly biased CT intensity estimation compared to atlas-based methods (Arabi et al 2019).

Instead of focusing on synthesizing the most accurate pseudo-CT image, minimizing the bias in the PET images might be optimal. This was accomplished using a combination of two networks: one that generates multiple plausible CTs using a multi-hypothesis deep learning framework and a second one that predicts the residuals between the ground-truth and reconstructed PET images obtained using each of the plausible CTs. To train the second network the authors proposed using imitation learning, a process that approximates the PET image reconstruction (Kläser et al 2019).

An even more interesting possibility enabled by deep learning techniques is to perform the attenuation correction in the image space without the need to generate the attenuation maps as an intermediary step. This is particularly challenging because the attenuation map is also used for generating the scatter correction sinogram, which means both attenuation and scatter corrections must be performed in image space when the attenuation map is not available. The feasibility of such a joint correction approach was showed using [¹⁸F]FDG-PET data (Fig. 13). A deep convolutional neural network consisting of five encoder-decoder stages based on the U-Net architecture was trained using three-slice inputs. The mean relative difference across all regions and subjects with respect to the CT based approach was ~4%. Although ~15% (range −13.4 to 63.1%) variation was reported, higher than that observed for the MR-based approach, these preliminary results suggest such an approach could be an alternative to CT- or MR-based methods in the future (Yang et al 2019). While these methods have obvious advantages (e.g. they do not require acquisition of CT or MR images, are not affected by intermodality registration, etc.), they also have limitations such as being radiotracer-dependent, could fail for radiotracers with very specific and minimal global uptake, require very large training datasets to account for the presence of pathological changes and might not work for dynamic imaging. A potential solution in the case of radiotracers with highly specific binding such as those used to study the dopaminergic system is to use convolutional neural networks to generate attenuation maps from the noisy maps obtained with the maximum-likelihood activity and attenuation reconstruction algorithm (Hwang et al 2018).

Figure 13:

Deep learning-based joint attenuation and scatter correction (ASC) in image space. The standard and deep learning workflows are shown on the left. Examples of the PET images obtained using the CT (CT-ASC), deep learning (DCNN) and MR (MR-ASC) based approaches and the corresponding difference images are shown on the right. Figures originally published in Physics in Medicine and Biology (Yang et al 2019).

5.2. Pelvis

The pelvis is another area where MR-based approaches that do not properly account for bone attenuation could lead to substantial bias in PET data quantification and deep learning approaches have been proposed to minimize this bias. Although the pelvic area has more intersubject anatomic variability than the brain, the attenuation map generation task is facilitated by its relatively stationary nature in the sense that the pelvic organs are not affected by respiratory motion (albeit peristalsis, displacement of bowel air, bowel movements, bladder filling, can all lead to mismatches between the MR and PET data). In one of the first implementation, a 3D fully convolutional network was used to predict CT patches from input MR patches. 3D convolutions were chosen to better exploit the spatial information available in the data (Nie et al 2016).

In the context of PET/MRI, a deep convolutional neural network based on the U-net architecture was first used to generate synthetic CT image patches using three Dixon contrasts and the ZTE image as inputs. The impact on PET data quantification was evaluated for two radiotracers, [¹⁸F]FDG and [⁶⁸Ga]PSMA-11. The mean underestimations of the linear attenuation coefficients and PET SUVs were reduced by a factor of three and the standard deviation by a factor of two compared to the standard Dixon-based approach. A factor of four reduction in root-mean-square error in PET quantification was observed in bone lesions (Leynes et al 2018).

The obvious limitation of this method was that it required the acquisition of the ZTE data, which increases the total acquisition time by more than two minutes. Instead, pseudo-CT maps can be generated exclusively from the Dixon images (Fig. 14). Using a 2D network that takes four contrasts as inputs (water, fat, in-phase and out-of-phase Dixon-VIBE images), the mean absolute relative change in PET values in the pelvis area was 2.36%±3.15% (Torrado-Carvajal et al 2019). Alternatively, pelvis attenuation maps can be generated only from the diagnostic MR images (Bradshaw et al 2018). Specifically, axial T2- and T1-weighted images were used as inputs to a network developed starting from the multiscale 3D convolutional neural network with fully connected conditional random fields (Kamnitsas et al 2017) to generate discrete-valued attenuation maps. The reported root-mean-square error of the entire PET image was less than 5%. Even lower PET quantification errors and improved bone identification were reported when using a generative adversarial network (GAN) trained using augmented data. To increase the number of training datasets and minimize the errors introduce by inaccurate MR to CT registration, a challenging task in the pelvis, the authors generated augmented deformation fields for each of the 18 prostate cancer patients used for training the GAN (Pozaruk et al 2020).

Figure 14:

Deep learning-based pelvis attenuation map generation. Sagittal, coronal, and axial attenuation maps using the CT-, Dixon- and deep learning-based approaches are shown in the first three rows. The distribution of relative changes (RCs) between the corresponding PET images are shown in the lower two rows. Figures originally published in the Journal of Nuclear Medicine (Torrado-Carvajal et al 2019).

Similar methods to synthesize CT-like images from MR images have been developed for MR-only radiotherapy planning (Wolterink et al 2017, Maspero et al 2018, Fu et al 2019, Spadea et al 2019, Wang et al 2019, Klages et al 2020, Dinkla et al 2019). All these methods could in principle also be applied to PET attenuation correction in integrated PET/MRI scanners. Those methods that can be trained using unpaired MRI-CT datasets (Wolterink et al 2017, Jin et al 2019) are of particular interest in the body given the difficult intermodality registration task.

5.3. Other body regions

Only a limited number of deep learning methods have been proposed for generating attenuation maps outside the head and pelvis. For example, synthetic CT images of the abdomen (and pelvis) have been generated from the Dixon MR images (water, fat, in-phase and opposed-phase) using a transfer fuzzy clustering and active learning-based classification. Preliminary results obtained with this method suggest the quantification bias in the PET measurements is reduced to below 5% (Qian et al 2019). In another approach, a 3D convolutional neuronal network was used to generate whole-body attenuation maps from the imperfect maps obtained using the MLAA algorithm that also included the time-of-flight information. 3D patches were again used for training to preserve the spatial information and reduce the memory requirements. The resulting attenuation maps were less noisy than the original ones and allowed better identification of bone (Fig. 15). The peak signal-to-noise ratios and normalized root-mean-square errors for the voxel wise comparison were also better for the proposed method. The authors argued that using only the [¹⁸F]FDG emission data avoids introducing errors due to inaccurate intermodality registration or MR image segmentation, but also acknowledged the network would likely require additional training to be used for other radiotracers (Hwang et al 2019). An approach that does not require any structural information was also implemented to generate whole-body attenuation corrected PET images from the non-attenuation corrected ones. Specifically, a supervised 3D patch-based cycle-consistent GAN architecture was trained and validated using 25 whole-body [¹⁸F]FDG PET/CT datasets and tested using 30 additional datasets, all obtained from oncologic patients. Using this very promising approach, mean errors below 3.5% compared to the CT-based approach were reported in all the organs analyzed except lung (Dong et al 2020).

Figure 15:

Deep learning-based whole-body attenuation map generation from the MLAA-derived activity (A) and attenuation maps (B). The Dixon- (C), deep learning (D) and CT-based (E) attenuation maps are shown for a representative subject. Figures originally published in the Journal of Nuclear Medicine (Hwang et al 2019).

6. Reflections on the past, present and future of attenuation correction in the PET/MRI field

6.1. How far have MR-based attenuation correction techniques come in the last decade?

In the early days of clinical PET/MRI, the equipment manufacturers focused primarily on developing the hardware needed to minimize the mutual interference between the two modalities and mostly relied on academia to develop advanced attenuation correction approaches. The only method available when the first dedicated brain PET/MRI prototypes were installed consisted of fitting an ellipse to the contour of the head estimated from the non-attenuation corrected PET data and assigning a single linear attenuation coefficient to all the inner pixels. Similarly, the first approach implemented for whole-body attenuation correction three years later allowed only the generation of approximate attenuation maps, albeit from MR data. As the general opinion was that attenuation correction was the most important methodological challenge in the field, many academic groups have persevered in developing increasingly accurate approaches. Interestingly, even the standard for assessing the performance of these methods has changed over the years. While segmented CT was considered the silver standard initially when only a limited number of tissue classes could be segmented from the MR images, scaled CT and even transmission images are now used for this purpose. In addition to being used for many research studies, some of these more advanced approaches have been adopted by the equipment manufactures and are now available for clinical use. Knowing where we started a decade ago, makes it easier to appreciate being able to generate pseudo-CT maps of the head or pelvis from morphological MR data in a few seconds.

6.2. Are the vendor-provided methods sufficiently accurate for routine clinical and research applications?

At least the first part of this question can be answered in the affirmative as there are already several PET/MRI scanners approved for clinical use. Although the early attenuation correction approaches were not without limitations, the equipment manufacturers have managed to address them reasonably well and implemented more accurate methods to minimize the bias to clinically acceptable levels, which was confirmed by the majority of the studies performed to date, as discussed in Section 3.2 (and Section 4.1 for the ZTE-based approaches).

However, one can argue that settling for “routine” is equivalent to setting the bar too low for PET/MRI. As emphasized in numerous publications, PET/MRI could enable PET to go beyond qualitative or semiquantitative data analysis. In particular, PET/MRI could facilitate the adoption of true quantification, presently considered almost exclusively a research application, for the initial diagnosis or longitudinal assessment of therapy response (Catana et al 2018). This requires reducing the bias related to attenuation correction below the expected effect sizes. As first steps in this direction, cerebral glucose metabolic rates were estimated non-invasively using dynamic PET data and an automated MR-driven approach for calculating the [¹⁸F]FDG input function (Sundar et al 2019) and the whole-body [¹⁸F]FDG influx rates were measured from dynamic whole-body PET/MRI data (Johansson et al 2018). In the case of whole-body studies, the longer acquisition times required for MR imaging offers an unexpected opportunity for dynamic whole-body PET imaging, in addition to lowering the radiation exposure compared to the equivalent PET/CT examination (Rahmim et al 2019).

6.3. Which applications require even more accurate attenuation correction methods?

Quantitative imaging of bone lesions.

Notwithstanding all the focus on minimizing bone tissue misclassification in the MR-based attenuation maps, significant bias can still be present in bone lesions. Osteoblastic metastases in particular have been shown to have lower HU values than normal bone and enostoses (a type of benign bone lesion) (Ulano et al 2016). Furthermore, their attenuation properties can change during treatment, with higher HU values observed posttreatment (Elangovan and Sebro 2018). Interestingly, even low-dose CT performed for attenuation correction purposes had a lower performance than diagnostic quality CT in this setting, requiring a joint PET and CT prediction model to improve accuracy when differentiating enostoses from treated metastases (Lee et al 2020). In the case of PET/MRI, the attenuation in these lesions is still underestimated or overestimated when using 4- or 5-compartment MR-based methods, respectively.

Imaging pediatric patients.

PET/MRI has been suggested as a lower-radiation exposure modality for evaluating pediatric patients compared to PET/CT. However, very few attenuation correction methods have been specifically developed for or at least properly evaluated in this patient population, which is required given the higher anatomic variability and differences in bone and lung density (Bezrukov et al 2015). Pediatric atlases could be generated if a sufficient number of matching MR and CT data sets to reflect the variability are available (Bezrukov et al 2015). Alternatively, a deep learning approach that uses UTE data has been shown to produce accurate attenuation maps even in pediatric brain tumor patients (Ladefoged et al 2019).

Non-oncological lung imaging.

Although high resolution CT is the preferred imaging modality in the lung, providing excellent morphological and even physiological information, it is associated with high radiation exposure. PET on the other hand can provide more disease-specific measures particularly when using targeted radiotracers. For example, using a collagen-specific PET probe (Desogere et al 2017), higher collagen deposition was observed in idiopathic pulmonary fibrosis patients compared to healthy controls (Montesi et al 2019). As MRI could provide complementary readouts of morphology, physiology, metabolism and molecular processes, there is an interest in performing the data acquisition simultaneously using integrated PET/MRI scanners. However, pathological changes in the lung (e.g. fibrosis) lead to substantial changes in its attenuation properties that are not compensated by current lung attenuation map generation algorithms. Minimizing this subject- and disease stage-specific bias will require the optimization of dedicated MR sequences and implementation of advanced tissue segmentation algorithms to generate more accurate lung attenuation maps.

Machine learning multimodal imaging data integration and analysis.

Machine learning approaches for extracting additional information from multimodal images have great clinical translational potential. Although these advanced methods could be applied to the data acquired separately, an integrated PET/MRI scanner is ideal for comparing MR and PET derived image features to identify those that provide complementary information as it alleviates the limitations of software registration algorithms and minimizes the physiological changes between the two separate examinations. The most accurate attenuation correction methods are required to guarantee that PET data similar to that obtained using stand-alone devices are used as inputs for characterizing disease-specific imaging phenotype using machine learning approaches.

6.4. How can promising methods developed in academia by disseminated to other users?

In many cases this dissemination has been accomplished using a direct customer-to-customer transfer of the software packages required for generating the attenuation maps. For example, the SPM-based approach (Izquierdo-Garcia et al 2014) has been shared with at least a dozen of academic groups around the world. Ideally this transfer of knowledge should be bidirectional, in the sense that the additional data available at the “receiving” end should be used to further optimize the original method. Such an approach would be particularly useful for improving the performance of deep learning techniques and could be accomplished either by sharing the data (which is unfortunately increasingly more difficult) or using federated learning (Konečný et al 2016). Obviously, an even more efficient transfer to end users could be facilitated by the equipment manufacturers. First, they should provide the capability of importing custom-made attenuation maps on the scanner to allow the PET images to be reconstructed using the standard software. Second, academic-industrial partnership should be supported for addressing the remaining limitations of the vendor-provided approaches. The resulting techniques should be disseminated as work-in-progress packages to be tested more broadly. Finally, promising methods developed in academia should be adopted, further developed, implemented and distributed according to industry standards so they can be used for clinical purpose.

6.5. How can attenuation correction specific issues relevant to accreditation/qualification of integrated PET/MRI scanners for multicenter trials be addressed?

Relatively few studies have assessed the reproducibility of the attenuation map generation methods. As a first step in this direction, high interscan repeatability of the whole-body Dixon-based attenuation maps was reported in patients without tissue artifacts (Rausch et al 2016). Scan-rescan SUV variations smaller than ±20% were reported in 95% of the cases examined when the same method was used to generate thoracic attenuation maps (Olin et al 2018). The repeatability of the atlas- or UTE-based methods was also assessed. For example, comparable attenuation maps and PET volumes were obtained at three visits using the probabilistic atlas method (Chen et al 2017). Similar results were reported for the SPM-based method (Fig. 16) (Izquierdo-Garcia et al 2018).

Figure 16:

Inter-scanner repeatability of the head attenuation map generation method. The attenuation map obtained at three time points for a representative subject are shown on the left. The PET images obtained from the same emission data reconstructed with the three different maps are shown on the right. Figures originally published in IEEE Transactions on Radiation and Plasma Medical Sciences (Izquierdo-Garcia et al 2018).

Although this review focused almost exclusively on the methods proposed for generating attenuation maps to account for the attenuation in the subject, another factor that needs to be considered in this context is the challenging task of generating attenuation maps for the image quality phantoms. Although this approach has proved very useful in the case of PET/CT scanners, it is inadequate for PET/MRI as the algorithms optimized for generating human tissue attenuation maps cannot simply be applied to standard phantoms. Instead, this task requires dedicated phantom acquisition and processing protocols (Boellaard et al 2015) or novel fillable and easy to manufacture phantoms made from stable materials that mimic both the MR and 511 keV attenuation properties of human tissue (Rai et al 2018, Chandramohan et al 2020). Alternatively, approaches to evaluate the performance of attenuation correction methods that do not require scanning of phantoms could be implemented, albeit also requiring a revision of the criteria for qualifying hybrid scanners for a multicenter trial.

6.6. What are the priorities in the near future?

Reaching a consensus on the status of attenuation correction for various applications.

While attenuation correction will likely continue to be challenging in certain body regions (e.g. thorax), it is often considered a solved problem in others (e.g. head (Ladefoged et al 2016)). Nevertheless, as the early-generation vendor-provided methods were the only approaches available at many institutions (either due to lack of technical expertise or the requirement to use the vendor-provided methods for clinical purposes), the perception that head attenuation correction is still problematic has spread and still prevents PET/MRI from being included in multicenter studies. Substantial progress has also been made in the pelvis.

Establishing unified quantitative evaluation protocols.

In addition to the different silver/gold standards used for testing attenuation correction methods mentioned above, there was substantial variability in the number of subjects enrolled in the validation studies as well as in the radiotracers used for this purpose. The community needs to converge on a set of methods and metrics to be used to assess the performance of various methods for specific applications.

Performing multicenter studies to compare the performance of existing attenuation correction methods.

In one of the most comprehensive multicenter studies performed to date, eleven attenuation correction methods developed by academic groups and Siemens (Koesters et al 2016, Paulus et al 2015, Izquierdo-Garcia et al 2014, Burgos et al 2014, Merida et al 2015, Cabello et al 2015, Juttukonda et al 2015, Ladefoged et al 2015, Anazodo et al 2015) were compared using a large cohort of Alzheimer’s disease subjects who underwent PET studies using three different radiotracers and using a unified quantitative evaluation protocol (Fig. 17). The authors’ conclusion was “that the challenge of improving the accuracy of MR-[based attenuation correction] in adult brains with normal anatomy has been solved to an acceptable degree, with errors smaller than the quantification reproducibility in PET imaging.” (Ladefoged et al 2016). Similar studies should be performed for other applications or to assess the performance of attenuation correction in other body regions. Extending this work to a multivendor setting is also needed. Early steps in this direction were made by scanning the same volunteers on each of the scanners investigated (Beyer et al 2016). Furthermore, a cross-vendor validation was already performed when methods developed for one particular scanner were evaluated against CT data obtained from a scanner manufactured by a different vendor. Although vendor-specific formulas are used to convert the HU values to linear attenuation coefficients, the resulting attenuation map obtained from various PET/CT scanners are largely considered equivalent.

Figure 17:

Head attenuation maps generated using 11 different methods for a sample patient that minimized the difference of the overall brain error to the median error across all methods. Figure originally published in NeuroImage (Ladefoged et al 2016).

Integrating attenuation correction with other MR-based approaches for improving the PET data quantification.

Several other sources of information provided by MR can be leveraged to address confounding factors in PET (e.g. subject motion, limited spatial resolution, etc.). As an example, an efficient approach for MR-assisted PET data optimization (i.e. perform attenuation and motion correction and anatomy-aided reconstruction) for brain studies has been shown to enable more accurate assessment of pathological changes in dementia patients (Chen et al 2019). Similar pipelines for improving PET data quantification implemented for heart (Kolbitsch et al 2018), liver (Fuin et al 2018) and even lung (Fayad et al 2017) imaging could dramatically improve the performance of PET for a wide range of clinical and research applications.