Significance
The novel opportunities afforded by simultaneous acquisition of PET and MR imaging data on modern hybrid scanners come at the cost of increased difficulty in reconstructing quantitatively correct PET images and parametric maps. This study contributes to solving this problem by quantitatively comparing recently proposed MRAC methods. The results will help users of PET-MR hybrid systems to choose the correction method that is most appropriate for their practice requirements. Specifically, we reconstructed individual parametric [11C]PE2I BPND and R1 images generated by a voxel-wise kinetic modelling approach applied on a dynamic [11C]PE2I PET scan using four methods. The main differences with most other reported MRAC evaluation studies are three-fold: (1) acquisition of dynamic data instead of static data; (2) use of individual 68Ge-transmission AC maps with a direct measurement of 511 keV AC as reference method in contrast to CT-AC; (3) parametric images providing functional information, related to DAT availability and overall brain activity, whereas standard uptake values (SUV) or ratios to a reference region (SUVr) reflect mostly tracer uptake. The study is pertinent for quantitative dynamic PET/MR imaging in research, but also for our clinical practice because dynamic [11C]PE2I PET scanning, followed by a kinetic modelling approach, is implemented in the routine clinical evaluation for differential diagnosis of parkinsonian disorders at Uppsala University Hospital.
Main results
The evaluation of attenuation maps showed that ZTE- and MaxProb-MRAC agreed best with 68Ge-AC in brain soft tissue. In bone, Atlas-MRAC showed the lowest and ZTE-MRAC the highest bias. While absolute differences were small in soft tissue (less than 10%), they were larger in bone (up to 16%), with Altas-AC showing the lowest and ZTE-MRAC the highest bias. The bias in bone may in part be due to the spatial resolution of the different AC images, with skull being more smoothed in the Atlas- and 68Ge-AC images than in Maxprob- and ZTE-AC images. Since the bone VOI was based on ZTE, it inherently showed lower AC values in Atlas and 68Ge-AC maps.
For BPND in striatal regions, mean bias across AC methods ranged from about − 12 to + 9%. For R1, the mean bias ranged from about + 1 to + 3% in various clusters of regions across the brain for all methods. Values within limits of -/+ 5% are generally regarded as acceptable. Importantly, differences in accuracy and precision between MRAC data sets can in part be explained by differences in shapes of the time activity curves and its SDs.
ZTE-MRAC showed low BPND bias for all striatal regions (-0.9 to 1.8%). The vendor-provided Atlas-MRAC method appeared to perform well for estimation of BPND-values in putamen, but a severe and significant underestimation was found in BPND values in caudate (12.1%). In contrast, for MaxProb-MRAC the mean bias was also relatively low in caudate (2.1%) while a substantial, but non-significant, bias was observed in putamen (9.2%). However, MaxProb-MRAC showed higher variability in bias for all regions (see Fig. 4A), and in both estimates of BPND and R1, indicating a lower precision than ZTE- and Atlas-MRAC. Methods with both a high accuracy and precision are most desirable as this will lead to more robust and reproducible outcome parameters. As such, the higher accuracy of MaxProb in terms of attenuation maps did not translate to higher accuracy for BPND estimates.
Previously, we evaluated Atlas- and ZTE-MRAC compared to 68Ge MRAC using static brain [11C]PE2I PET images from the same group of participants with suspected parkinsonian disorders [32], but using only regional SUV and SUVR (SUV ratios to cerebellar grey matter) as semi-quantitative outcome parameters. In this prior work, ZTE-MRAC was also found to be more reliable than Atlas-MRAC in terms of accuracy, precision, and degree of agreement compared to 68Ge MRAC. The same clusters of brain regions across the brain were used for R1 in the current study, allowing some comparisons. The largest mean regional biases were found when using SUV, i.e. varying between 6 and 10% for both Atlas- and ZTE-MRAC. This bias decreased considerably when normalizing SUV to cerebellum, resulting in regional bias percentages of 2.9 to 4.8% for Atlas-MRAC and − 1.3–2.6% for ZTE-MRAC.
Only a few prior brain PET/MRI studies have investigated the effects of MRAC methods on accuracy and precision in the context of kinetic modelling of dynamic PET [33, 35–37, 49]. Schramm et al. [49] evaluated ZTE-MRAC and Atlas-MRAC compared to CT-based AC in eight healthy volunteers using a Signa PET/MRI scanner and [18F]PE2I. BPND in striatal regions was estimated using SRTM with cerebellum as reference region. In striatum, the mean bias was 1.5% for ZTE-MRAC and 3.8% for Atlas-MRAC. These results are in line with our study for ZTE-MRAC, but not for Atlas-MRAC where we found a bias of -5.4%. Our Atlas-MRAC results revealed that especially in caudate, the BPND values were severely underestimated for all participants (Fig. 4–5, Table 1), leading to a range of individual biases from − 11.0 to -1.0% compared to -1.0 to + 9.4 reported by Schramm et al. [49]. The discrepancy in results might be due to differences in the type of participants (patients vs healthy subjects) and methodological aspects such as tracer properties ([11C]PE2I vs [18F]FE-PE2I), definition of brain regions (using Hammers maximum probability atlas in PMOD [50, 51]), and most importantly the reference standard used (68Ge vs CT). However, both studies point out distinctly that ZTE-MRAC performed better than Atlas-MRAC.
Merida et al. [33] assessed the value of a multi-atlas approach (MaxProb-MRAC) compared to a single atlas with CT-based AC as reference method. The study used dynamic PET data from seven healthy participants acquired on a PET/CT scanner without TOF and using the serotonin 5-HT1A receptor tracer [18F]MPPF. SRTM was used for estimation of BPND across the brain, with cerebellar grey matter defined with the Hammers maximum probability atlas [50, 51] and excluding cerebellar vermis which may contain 5-HT1A receptors. Overall, bias of BPND ranged from − 2.5–5.0% for MaxProb-MRAC and from − 9.3–3.3% for their single-atlas MRAC. For MaxProb, the range of the bias was in good agreement with our results, as was the substantial increase in bias when using a single-atlas MRAC. Furthermore, their single-atlas MRAC caused a severe underestimation of BPND values, as seen for Atlas-MRAC in our study. However, these comparisons between studies need to be interpreted with care because of differences in data acquisition, tracer kinetics and specified VOIs. It should also be noted that this study’s single-atlas MRAC implementation is different from Atlas-MRAC in the present work.
Previous reports indicate that the effect of different AC methods over time after injection may be dependent on the varying distribution of the tracer [33, 34, 47, 48]. In addition, any misalignment-induced errors in attenuation correction are not corrected by post-reconstruction frame-by-frame realignment. Although identical motion correction was used for all four AC methods, misalignment-induced errors can vary for different AC methods. These could be two reasons for the differences in accuracy and precision of outcome parameters of kinetic analysis based on the different AC methods. Hence, relative bias and its SD in SUV values over time for striatal regions and cerebellum after implementation of Atlas-, ZTE- and MaxProb-MRAC were assessed. Both relative bias and SD varied over time in different ways for the different regions and methods, especially in caudate and consequently striatum (Fig. 9). MaxProb showed distinct changes over time for the later frames. These findings are in line with the hypothesis that various MRAC methods are affected differently by the radioactivity distribution in the brain at different time points [33]. In the present study, the biases on TACs have different temporal shapes between regions for Atlas-MRAC and MaxProb, which partially explain their bias in R1 and BPND. For example, for MaxProb, the bias curves for cerebellum and striatum have similar shapes, and bias in the resulting BPND is low (2.1%) whereas bias curves in cerebellum and putamen have different shapes, resulting in a larger bias in BPND for putamen (9.2%) despite the bias for putamen in the TACs themselves being much lower than that of the other two methods. For ZTE, TACs show large, but constant biases over time (around 10% for striatum and 5% for cerebellum). These quantification errors are then compensated when using cerebellum as reference region in kinetic modelling, or SUV ratios over cerebellum. It is noteworthy that the constant-but-high biases on ZTE TACs could have a higher impact when kinetic modelling is performed with arterial input functions or with a reference region other than the cerebellum, where the error compensation between two regions may not take place. In that case, Maxprob-AC may be the optimal method because of its lower absolute bias.
Based on our results it should be noted that the choice of MRAC method may depend on amongst other the aim of the PET examinations, type of participants, tracer and quantification method. The aim of the study determines the requested accuracy and precision of the quantification method required for interpretation of the results. When considering absolute values, MaxProb’s low absolute TAC bias may be particularly important. For ratio or reference region-based methods ZTE’s higher but temporally consistent TAC bias will be an advantage. In case of participants with an abnormal anatomy, e.g. post-trauma or -operative patients, acquisition-based methods like ZTE-MRAC will be preferable. In contrast to stand-alone PET and PET/CT systems it may be difficult to develop a standard MRAC routine for AC in integrated PET/MR systems suitable for all circumstances and tracers. Instead the choice of MRAC method for a specific investigation need to be based on balancing the pros and cons of each approach.
Methodological Considerations
There are five major methodological considerations which need to be addressed as they might affect comparisons between MRAC methods and other studies. First, we used 68Ge-transmission AC maps with a direct voxel-wise measurement of AC in contrast to CT-AC. A 68Ge transmission scan measures attenuation directly at 511 keV, whereas CT-AC requires a conversion from Hounsfield units to 511 keV attenuation coefficients. Hence, the 68Ge-AC attenuation maps used in the present work can be regarded as the true reference standard for AC. A comparison of 68Ge-AC and CT-AC showed higher radioactivity concentrations when using CT-AC, although of small magnitude, the differences were consistent and significant [8]. This bias should be noted when considering quantitative analysis and comparisons across different imaging modalities. Today CT-AC is more widely accepted in clinical practice, but this might be due to a higher signal to noise ratio and shorter acquisition times compared with 68Ge AC. Second, the AC of 0.097 cm− 1 measured for soft tissue for 68Ge-MRAC is modestly lower than that for the other investigated MRAC methods, i.e. 0.100 cm− 1. The soft tissue value of 0.100 cm− 1 was the same as used in clinical practice for CT-AC. In the previous study with static [11C]PE2I PET data [32] the effect of different AC values was investigated. The accuracy of SUV for various brain regions appeared to be considerably higher for both Atlas- and ZTE-MRAC when scaling all AC maps to 0.097 cm− 1, but the precision was scarcely affected. When SUV was normalized to cerebellum (SUVR), the accuracy and precision were similar before and after scaling. Even though in the present study outcome parameters are normalized to cerebellum, we cannot exclude the possibility that differences in AC values might have affected the magnitude of the measures of accuracy. Third, AC of sinus regions is vulnerable to misclassification of tissue types and inaccurate registrations of templates/atlas as they are composed of bone, soft tissue, and air [3, 28]. When this mixture of structures is classified as bone, the resulting positive bias can affect cerebellum if the sinus region in question is close [3]. In our study, parametric R1 and BPND images were generated from the dynamic [11C]PE2I scan using a kinetic modelling approach with cerebellum as reference region. In theory, biased AC in sinus regions could therefore have affected the accuracy of the estimated R1 and BPND values in various regions. However, we believe this effect is modest. For R1, which displays CBF at an early stage, only a minor bias was noted for all MRAC methods. On the other hand, BPND is related to the specific binding of [11C]PE2I to DAT, while the cerebellum is devoid of DAT. Accurate registration of templates/atlases might be hampered by the difficulty of matching skull boundaries in sinus regions. Furthermore, the interindividual variability is large for sinus regions. Inaccurate registrations might confound parameter estimations, particularly in subcortical regions. Fourth, similar to 68Ge-transmission AC maps, ZTE-MRAC is based on an individual scan in contrast to single- and multiple atlas MRAC approaches. One advantage of a specific AC map for every individual is that it does not depend on a priori anatomical information and assumptions [10, 49]. Multiple atlases should be more reliable than a single atlas, as shown earlier by Merida et al. [33] and also found in our comparison of Atlas- and MaxProb-MRAC. However, methods based on an individual scan are likely to be suitable even for patients with an abnormal brain anatomy (e.g. brain cyst or hydrocephalus) or after surgery (e.g. patients with epilepsy surgery or traumatic brain injury). Fifth, the vendor-provided ZTE- and Atlas-MRAC are under continuous development and reflected the state of the art at the time when the study was conducted. At the time of this writing, the version of ZTE-MRAC as used in the present work (MP26) is commercially available [52], but further developments have already been proposed [53, 54]. Some may have already been implemented but were outside the scope of this study.
Limitations
Our study has some limitations related to design and methods. As for the previous study, we had to deal with differences in axial FOV between scanners related to AC maps for 68Ge-MRAC (155 mm), MaxProb-MRAC (216 mm) and Atlas- and ZTE-MRAC (250 mm). For this reason, the missing information of the AC maps for 68Ge (top and neck regions) and MaxProb (neck regions) were completed with corresponding information from the Atlas AC map. Consequently, the lines of response associated with the added parts might have introduced a bias towards Atlas-MRAC, but more in regions outside than inside the brain. Another limitation is related to head movements during the entire dynamic PET acquisition. The MRI sequences for AC are acquired during the first five minutes of the PET scan to avoid misalignment between MRAC and MRI data. Since frame by frame motion correction was done using identical transformation parameters for all four dynamic datasets any remaining movement induced errors should be the same for all four AC methods. However, any motion-induced errors in attenuation correction are not corrected for using post-reconstruction motion corrected methods. The effects of these may differ for the different AC methods, and this may have affected our results. Further we had only data from a cohort of patients with parkinsonism, some of whom had severe atrophy. In this case MaxProb-MRAC may mislabel cerebral spinal fluid as bone, which could cause a positive bias in cerebellum as a result of inaccurate bone estimation [18], but this positive bias is not seen in our results. It might have been beneficial to study a healthy cohort in addition to the patient cohort, as effects of MRAC method and disease (alterations in DAT availability and overall brain function) were entangled in the current design. It should also be noted that the tracer used, [11C]PE2I, allowed only investigation of the impact of the various MRACs on BPND values in striatal regions which have a high DAT density. From an evaluation perspective it would be beneficial to investigate a tracer for neuroreceptors or transporters widely distributed across the brain in addition to the dopaminergic system. Despite these limitations, this study provides important information on the accuracy and precision of four MRAC methods in a clinical setting with use of dynamic data.
Future developments
For the future development of MRAC methods in dynamic brain PET, it would be important to consider diverse neurological applications in larger patient cohorts with specific tracers. The current study comprised data from only nine patients with parkinsonism, which was enough to attain the aim of the study. However, small study populations do not allow subgrouping based on, e.g., the magnitude of accuracy and precision, subtypes of parkinsonian disorders, or other relevant numerical or categorical measures. Application to more complex cases will show whether the developed MRAC methods are likely to perform well in clinical practice. For example, differential diagnosis of parkinsonian disorders might be extremely challenging at an early stage due to the difficulty of distinguishing typical Parkinson’s disease and the atypical subtypes, and the confounding effect of comorbidity with other neurodegenerative and chronic diseases (e.g. dementia and diabetes). For this, it is essential that quantitative values be accurate and precise. Another promising direction is the further development of an individual mapping of linear attenuation coefficients, as suggested by Visvikis et al. [3]. Like CT-AC, most of the reported MRAC-methods disregarded possible heterogeneity in tissue by using AC maps with fixed values assigned to a limited number of tissue types. More detailed attenuation maps might be developed by optimizing the simultaneous acquired information from MRI and PET combined with templates/atlas and machine learning [55, 56]. Advanced MRAC methods would certainly further promote the clinical use of integrated PET/MRI scanners.