Variability and Uncertainty of ¹⁸F-FDG PET Imaging Protocols for Assessing Inflammation in Atherosclerosis: Suggestions for Improvement

DISCUSSION

Our analysis of the literature showed the huge variability in the ¹⁸F-FDG PET imaging protocols used to assess the inflammation associated with the plaques present in the arterial bed. The differences first concerned the acquisition protocol. This variability could be due to the fact that about one third of the studies (15/49) were retrospective studies using data that were initially acquired for cancer investigation and then revisited to investigate vascular inflammation. Yet, even when these 15 studies were removed, the variation in I_counts was still 37 ± 24 MBq/kg min (against 35 ± 27 MBq/kg min for all 49 studies). Differences also related to the reconstruction protocols. Iterative reconstruction was used in half of the studies mentioning the reconstruction algorithm, but one third of the reports did not mention which algorithm was used to produce the images. Again, these differences might be due to the fact that part of the studies were initially designed for tumor imaging, so that images were not necessarily reconstructed using a protocol suitable for accurate quantification in subcentimeter lesions. Key parameters such as the spatial resolution and the voxel size were not even mentioned in almost 70% of the articles. This lack of appropriate reporting prevented a sound comparison of the results given that spatial resolution and voxel size highly affect the severity of PVE and tissue fraction effect (5), which in turn strongly biases the measurements in small lesions such as plaques. Last, protocols highly varied in the metrics used to assess the severity of the disease. This high variability reveals a lack of consensus regarding which parameter should be measured and how this should be done. Most studies (89%) measured either an SUV or a TBR but still differed in the way these metrics were estimated. The precise definition and size of the region used to make the measurement were reported in less than 10% of the articles. Only 6% of the studies used metrics that gave an overall assessment of the disease by combining measurements performed in different locations of single or different targeted segments.

These observations demonstrate the need for harmonization and systematic reporting of ¹⁸F-FDG PET imaging procedures of vascular wall inflammation to facilitate the comparison of results between studies. Also, such a comparison should account for the fact that in retrospective studies involving oncology patients, there might be confounding radiation- and chemotherapy-related vascular inflammation and alterations in the metabolic milieu related to the underlying malignancy itself. Therefore, SUV arterial wall signals assessed in oncology subjects will remain difficult to compare with those measured in noncancer patients.

Using 2 simulated lesions, we observed that the bias in SUV estimates was large whatever the acquisition and reconstruction protocol. This large bias is due to the small thickness of the inflammatory part of the atherosclerotic lesions (typically ∼1 mm) associated with a small longitudinal extent (typically 1–3 cm) with respect to the spatial resolution in the reconstructed images (∼5 mm). In particular, the lesion thickness was well below 3 times the FWHM of the PSF characterizing the imaging system, making it impossible to properly recover the activity concentration without any sophisticated PVE correction.

Despite these high errors, we identified various approaches that minimized the biases. The parameter that most affected the error was the number of iterations used in reconstruction. The use of too-few iterations maximizes the bias, especially for SUV_max, whereas SUV_mean is the least sensitive to that number. The benefit of using an increased ENI was observed for all reconstruction schemes (Figs. 4 and 5), without excessive increase in the variability of SUV estimates. Using a high ENI is especially important when modeling the PSF in the reconstruction, because such a modeling reduces the speed of convergence. Yet, in most articles surveyed as part of our literature analysis, most reconstruction protocols used fewer than 60 ENIs. For the largest lesion reconstructed with a 1-mm voxel size and PSF modeling, varying the ENIs from 60 to 150 increased SUV_max from 3.8 to 5.1.

Postfiltering always had a detrimental effect on SUV recovery, as it further blurred images, hence amplifying PVE. Although using a low iteration number and postfiltering improved visual image quality (Fig. 3), this was at the expense of quantification accuracy, suggesting that the images that are used for visual interpretation should not be the same as the ones from which measurements are performed.

Using finely sampled reconstructed images associated with PSF modeling also systematically reduced the bias in uptake estimates, although the effect was not as high as that brought by increasing the number of iterations or removing postfiltering. When small voxels are used, noise regularization is absolutely needed to avoid excessive noise in the reconstructed images as shown by comparing Figures 3E and 3F. Such noise regularization can be introduced by modeling the PSF during the reconstruction process, as shown by comparing the orange and light blue curves in Figures 4 and 5. The impact of the voxel size was the highest for SUV_max and the lowest for SUV_mean. It was almost negligible for the large lesion.

The acquisition duration modified the measured values to a small extent: increasing the acquisition duration decreased the variability of the error for a given bias. This was especially true for SUV_max.

The metrics systematically affected the resulting error, with the smallest biases observed with SUV_max and the largest bias observed with SUV_mean.

Overall, for an SD in the measurement less than 0.5 SUV units, the lowest bias was always estimated using the protocol involving an 8-min acquisition and a reconstruction including a PSF model using a 1 × 1 × 1 mm voxel size and no postfiltering, with at least 120 ENIs when using an OSEM approach. This was true whatever the metrics and the lesion. Yet, even with these optimized parameters, the mean error is SUV estimate remained high because of PVE. Several strategies can be considered to reduce these errors: improving the spatial resolution of the PET images based on hardware development and implementing explicit partial-volume correction (7). Also, in our study, we did not simulate motion, which introduces extra PVE, so motion should be compensated as well for increased accuracy (35). PVE correction is difficult to implement because the inflammatory part of the lesion cannot be seen in CT. PET/MR protocols might substantially ease the implementation of PVE by facilitating the delineation of plaque components (25). Improvements in these 3 directions will certainly enhance the value of plaque imaging using ¹⁸F-FDG PET. Still, even before such new features become practical in the clinics, the systematic and precise reporting of the acquisition, processing, and measurement protocols would considerably facilitate the reviewing of the literature and the interpretation of the data and might help explain contradictory findings (31,36).

Given the high biases affecting the measurements, one can wonder how ¹⁸F-FDG PET plaque imaging has been found useful to identify patients at high risk of plaque rupture (37). Although the bias is high, the variability of the bias is not so high, especially because the number of iterations is always rather low in published reports. Thus, for a given protocol, the magnitude of the error will be related to the lesion size and contrast. Even if the measurement is wrong, it bears information regarding the lesion features. The measured metrics describe a combination of activity concentration and lesion size (5,7). For a given lesion size, despite the bias, the measured value will reflect the activity concentration. For a given activity concentration, the measured value will be larger for larger lesions. When both the size and the activity concentration decrease, the measured metrics decrease as well. As the measured metrics reflect a combination of activity concentration and lesion size, they still convey useful information regarding the lesion.

In this article, we analyzed SUV only. The use of TBR has been justified by indicating that it was a blood-corrected uptake measurement (31,38). Yet, blood activity actually adds to wall activity due to the spatial blurring of the PET images (imperfect spatial resolution) and to the tissue fraction effect. Mathematically, we have:Eq. 2where OWA stands for observed wall activity, RWA stands for real wall activity, BA stands for blood activity, and a and b are 2 unknown weighting factors. The weight a is less than 1 because of the spill-out of wall activity in neighboring regions. Even if BA could be precisely measured, the resulting so-called TBR used in current articles would be:Eq. 3which is an accurate estimate of RWA only if b = 0 and a/BA = 1, that is, OWA = BA.RWA, as if the blood contamination was a multiplicative process instead of an additive one. Therefore, there is no legitimate rationale for using TBR instead of SUV. In addition, the use of TBR increases the variability of the measurements because of the biologic and measurement variability of the blood uptake, as actually underlined or suggested in a few articles (37,39). The variance of TBR is indeed the sum of the variance in OWA and of the variance of the estimated BA, making TBR less reproducible than SUV.

Last, our study focused on ¹⁸F-FDG as a marker of the inflammatory component of atherosclerosis. Other tracers targeting plaque features or some other components of an atherosclerotic lesion, such as the promising ¹⁸F-fluoride (40), would also lead to focal uptake of spatial extent similar to ¹⁸F-FDG, suggesting that the observations reported in this paper would have some relevance for those tracers.