Impact of point spread function modelling and time of flight on FDG uptake measurements in lung lesions using alternative filtering strategies

[¹⁸F]2-Fluoro-2-deoxy-d-glucose (FDG) positron emission tomography (PET) has been shown to play a key role in the management of patients with non-small cell lung cancer in terms of staging and prognosis [1]-[5] and monitoring response to therapy [6]. In these applications, the uptake of FDG expressed as standardised uptake value (SUV) is of key importance, with SUV_max being the most commonly reported measure [7]. The use of SUV_max for discrimination between benign and malignancy for soft tissue masses and lymph nodes has been demonstrated for lung cancer patients [8],[9] and changes in SUV_max used as an indicator of response to therapy [10].

While the use of SUV_max is commonplace, it is known to be sensitive to both reconstruction parameters [11] and the amount of statistical image noise, leading to poorer test-retest consistency relative to other SUV-based metrics [12],[13]. Consequently, alternative metrics such as SUV_peak [14] and total lesion glycolysis (TLG), the product of SUV_mean and metabolic tumour volume derived from the PET images, have been suggested for use, particularly in monitoring response to therapy [6],[15]. Recently, TLG has also been shown to offer superior prognostic information than SUV_max [16]-[20].

In recent years, there have been significant advances in iterative image reconstruction algorithms and scanner hardware. Consequently, reconstruction algorithms that include point spread function (PSF) modelling [20],[21] and time of flight (TOF) [22] have become commercially available on PET/CT scanners, with TOF also available on PET/MR [23].

The use of PSF modelling, with and without TOF, has been shown to improve signal-to-noise ratio (SNR) [24]-[27] and lesion detectability [28]-[30] partly through decreasing voxel variance. However, the implementation of PSF modelling, both within projection space and image space, from different manufacturers and also academic institutions has been shown to produce Gibbs artefacts [21],[31]-[35] (Nick Vennart, personal communication). In patient imaging, the Gibbs artefact, combined with reduced partial volume effects, has a significant impact on SUV_max [36]-[38]. This is particularly evident with minimal or no post-reconstruction filtering, which has been shown in phantom studies with numerical observers to provide greater lesion detectability [28]-[30]. Changes to SUV_max as a consequence of PSF modelling present a challenge as changes to defined local practice for reporting may be required such as changing the thresholds used for the discrimination of malignancy. The scanner used in this study has been part of a multi-site network of scanners for routine FDG oncology imaging since 2009. SUV_max is the reported uptake metric, and the consensus amongst local reporting clinicians within the network is that lesions with SUV_max > 5.0 are considered highly suspicious of malignant disease.

It is necessary, in practice, to smooth clinical images to provide image quality that is deemed acceptable for clinical reporting. This degrades the spatial resolution but increases signal to noise. The degree of smoothing applied at any given centre is heavily influenced by the experience and personal preferences of the reporting clinicians, informed by the advice of physicists providing scientific support. Where several PET scanners serve the same patient population, it is also advantageous to match imaging performance across the network in terms of visual image quality and quantitative characteristics.

A trade-off curve of signal enhancement versus noise reduction when using PSF and/or TOF algorithms can be established by applying a range of reconstruction post-filters. It has been demonstrated that it is possible to match SUV_max from PSF-based reconstruction with traditional non-PSF algorithms by applying a particular post-filter. Lasnon et al. [39] showed that a 7.0-mm full-width-half-maximum (FWHM) post-filter with PSF reconstruction gave comparable recovery coefficients in phantom data to non-PSF reconstructions and brought the recovery coefficients in line with European recommendations [40]. Another study proposed the application of a post-filter for the purpose of quantification [41]. This study also demonstrated that despite a spatially dependent PSF, this approach of using a single post-filter choice was adequate for all lesions irrespective of their location in the field of view. The application of a relatively broad post-filter to PSF modelling images may seem counterintuitive as it will undo the improvements in partial volume effect, but there are likely to be other benefits that have not been reported such as a reduction in voxel variance in the images.

Another potential solution may be to use alternative uptake metrics to SUV_max. One study [37] suggested that TLG may be more stable when comparing PSF to non-PSF reconstruction, but this study only assessed ten lung lesions. Another study [38] has suggested the move to SUV_mean based upon a 50% isocontour of SUV_max. To our knowledge, there are currently no studies that investigate the impact of these reconstructions with PSF modelling and TOF on TLG and SUV_peak.

The primary aim of this study was to evaluate the impact of PSF modelling and TOF on SUV_max-based lesion classification as implemented at the local institution. This was performed using Siemens reconstruction software including implementations for TOF and PSF modelling (HD, UHD). Implementations of reconstruction algorithms can differ, and therefore, the results might be specific to HD and UHD; however, we feel it is likely that findings may be generalisable to other reconstruction implementations with similar philosophies. Any change in FDG uptake measurements across different reconstruction protocols can hopefully allow other centres to assess how such changes may impact their approaches to lesion classification. Two set criteria for post-filtering the images were assessed based upon characteristic locations on a signal enhancement versus noise reduction trade-off curve. These two points are 1) matching image noise (voxel variance) which was expected to enhance signal and 2) matching signal (SUV_max) which, based on previous studies [39],[41], was anticipated to require greater levels of post-filtering and hence reduce image noise. This latter approach is aimed to be particularly relevant to centres that wish to maintain uptake quantification for practical purposes, which is particularly important in multi-site imaging networks. In addition, this work aimed to expand on the results of previous studies [36]-[38] with the addition of TOF, evaluation of other uptake metrics such as SUV_peak and TLG, and determining gains in SNR for the two strategies.

The deployment of PSF and TOF-based reconstruction methods into routine clinical practice for FDG imaging presents a challenge, particularly in centres or collaborative imaging networks with a defined protocol for classification of malignancy based upon SUV data. To our knowledge, this is the first study that has evaluated the performance of PSF and TOF-based reconstruction algorithms with two post-filtering strategies based on the objective criteria of matched image noise (voxel COV) or matched SUV_max, quantifying the impact on SUV_max, SUV_peak, TLG and SNR_(T-L). Specific findings are applicable to Siemens HD and ultraHD reconstruction algorithms using the parameters applied in the study.

It is clear from the data in Tables 1 and 2 and Figure 3 that quantification differences occur in the phantom data for all algorithms applied in this study. There are several factors that will contribute to the differences: the effect of statistical noise, partial volume effect, the size (and hence number of voxels) of the region of interest and, for the HD and UHD algorithms, Gibbs artefacts. The contributions from these factors to the measurements of SUV_max will differ as reconstruction parameters are varied. We believe that the interactions between the various factors are complex and not completely separable. As such, we do not feel that it is possible to identify one single phenomenon as the source of quantification differences for any of the algorithms used.

It can be seen that overestimation occurs for all four reconstruction algorithms (Table 1) and requires the application of a post-filter to reduce this (Table 2). The smaller filter kernel applied to HD and UHD to match noise combined with voxel correlation leads to a lesser reduction of this overestimation. It can be seen that there appears to be a particular size of object where an overestimation with HD and UHD is particularly prominent with no or minimal levels of post-filtering, which, in part, may be due to overlapping Gibbs edge artefacts. Despite this, it can be seen from HD recovery data in Table 2 that, with matched voxel variance, there is very little dependence of recovery on sphere size for the 13- to 37-mm spheres, which is a desirable property. This highlights the importance of establishing a full understanding of the impact of these algorithms, and it is the duty of medical physics experts to educate clinicians on changes expected to quantification.

Ideally, the implementation of PSF modelling would not lead to Gibbs artefacts, but given the necessary compromises for PET imaging with limited statistics, an improvement in one area such as in image resolution is almost certainly going to lead to a deterioration in other aspects. Overall, whether the changes are desirable is application dependent, with our data showing smaller absolute errors for smaller spheres (but not for large spheres) and reduced dependency on quantification with lesion size.

Matching image noise produces marked increases in SUV_max, particularly with PSF reconstructions, that are potentially clinically significant, depending on local practice. This highlights the pitfalls of using uptake metrics such as SUV_max, that are so sensitive to partial volume effects and reconstruction parameters, with fixed thresholds for malignancy. The largest increases in SUV_max occur for small lesions, which typically have low SUV_max (less than 5), which is consistent with other studies [36],[37]. One potential solution may be to modify thresholds based on estimated tumour volume. It would be useful to extend the matching of SUV_max to smaller objects, but this is not possible due to the limitation of the current NEMA phantom, with 10 mm being the diameter of the smallest sphere insert. It is these small lesions, with SUV_max close to the typical cut-offs for discrimination of benign and malignant disease, that are arguably the most critical lesions for lung cancer staging as they are likely to be possible additional pulmonary nodules or lymph nodes. Determining whether a lymph node is malignant, particularly those in the mediastinum, has a considerable influence on the overall staging and will play a major role in patient management. This change in SUV_max is expected to require an adaptation of locally used thresholds for discrimination of disease. It was also noted from the phantom studies that variability of SUV_max was worse for PSF-based algorithms in the small spheres, which suggests worse test-retest performance in clinical data. This is suspected to be due to increased inter-voxel correlation that is introduced when using PSF-based algorithms [21]. This increased correlation results in a reduction of voxel variance (and hence the voxel COV as used in this study as a noise metric), but it has been shown to potentially result in larger variability of uptake metrics within small ROIs [45]. We feel that the impact of PSF modelling on variability for clinical data has yet to be explored fully, and while this is beyond the scope of this study, it is recommended that caution is observed when applying PSF modelling for assessing response to treatment with follow-up scans. Despite this, the reduced levels of post-filtering required with PSF and PSF + TOF have been shown to improve lesion visualisation [28]-[30].

With matched voxel COV, SUV_peak experiences similar differences to those seen for SUV_max, albeit to a lesser extent. Quantification of peak uptake implicitly includes an additional filtering operation with a spherical kernel. The small mean relative differences for TLG suggests that it is a relatively robust uptake metric when comparing against OSEM images for either filtering strategies. The large degree of variability seen in the relative changes, as highlighted by the confidence intervals in Tables 4 and 5, may be concerning. However, it should also be noted that the total range of TLG observed in this study is approximately a full order of magnitude greater than SUV_max and SUV_peak. The use of TLG has been reported in assessment of therapy response and, recently, for prognosis in a small number of studies. The increased stability of TLG with a volume delineation based on a percentage of SUV_max suggests the metric may be more appropriate than SUV_max for staging and prognosis as the evidence base for this metric is established. We believe that this is the first time that the dependence of TLG on reconstruction algorithm has been explored in the literature.

Alternatively, post-filters for PSF and TOF-based algorithms can be determined to give SUV_max that, according to this institute's practice, would not alter the outcome of the study. For all lesions with borderline SUV_max for suspicion of malignancy, relative changes with PSF and TOF-based reconstructions were less than 20%.

Matching SUV_max between PSF-based algorithms and OSEM has been demonstrated previously [39]. However, our study has also shown that matching SUV_max will significantly reduce the voxel variance in the image compared with OSEM, which we believe has yet to be demonstrated quantitatively. Combined with increased voxel correlation, this reduction of voxel variance alters the image appearance quite considerably and may be perceived as over-smoothing of images. Findings from this study are based upon an image matrix of 256 × 256 voxels, whereas other centres may use different parameters such as 200 × 200 or 400 × 400 voxels, which are common choices on the mCT due to the system's intrinsic 400 × 400 matrix. We believe that, when Gaussian post-filtering is applied, the dependence of both image noise and SUV_max on matrix choice is diminished. It has also been shown that the thickness of the walls of the fillable spheres of the NEMA phantom has an impact on SUV_max quantification [46],[47]. This is only seen to cause appreciable error with low sphere-to-background contrast and small spheres, and hence, we expect that the impact on the test objects used in this study is likely to be minimal.

It is noted that the degree of post-filtering for the HD and UHD algorithms (6.6 and 6.5 mm, respectively) will reduce spatial resolution for these PSF-based algorithms that are intended to provide superior spatial resolution. However, we feel that this approach may be beneficial when deploying a new PET/CT scanner to an existing clinical setting, comparing patient scans for follow-up with other systems or supporting the transition to a ‘new’ imaging facility with a catalogue or library of images with higher resolution.

In this study, the addition of TOF increased the variation in ratio values of image voxel variance for both phantom and patient data with either matched noise or matched SUV_max. In the patient data only, TOF appeared to introduce a slight positive bias and greater distribution of differences in the SUV_max data. This was not seen in the phantom studies and the cause of this is unclear. It could be due to a dependence on patient size, as TOF is associated with SNR gains proportional to the diameter of object [48]. However, in this study and others [20], this did not appear to apply in lung images where the majority of tissue in the image has low density with very low uptake of FDG.

We believe this is the first study to demonstrate SNR gains with PSF and/or TOF using lesion uptake as a measure of signal with two different criteria for choosing post-filtering. A recent study has shown reductions in voxel variance and gains in SNR but measured only in uniform areas of uptake with patient livers [27]. One study has evaluated SNR gains using lesion uptake as the signal [25] but only comparing images reconstructed with PSF and PSF + TOF, with the intention to demonstrate the SNR gains brought on by TOF. It was expected that SNR gains would be seen for PSF and TOF-based algorithms compared with conventional OSEM. However, it was not anticipated that the gains in SNR would be greater when parameters are chosen to match SUV_max. This may be of particular relevance for low-contrast lesions elsewhere in the body, such as the abdomen, which do not have the inherent high lesion to background contrast of lung lesions. The notion that increased levels of post-filtering may be superior in terms of SNR gains seems slightly at odds with published work on lesion detection that suggest less post-filtering results in optimal lesion detection [28],[29]. This may be due to fact that the definition of SNR in this study is not a direct indicator of lesion detectability.

There are two limitations with this study where future work is planned. Firstly, no histological correlation with FDG uptake measured in the lesions was performed as in other studies [36]. Therefore, it is not possible to determine cut-off values and diagnostic accuracy of the uptake metrics in the two strategies of implementation. This is arguably outside the scope of this study as the purpose was not to determine such data. Secondly, we have only assessed lung lesions, and from other studies [25], it is likely that reconstruction will perform differently in other areas of the body.

The effect of PSF and TOF-based reconstruction on quantification, particularly SUV_max, has limited their introduction into routine clinical use despite demonstrated improvements in lesion detectability. This study extends existing studies [39] which have shown that the impact on SUV_max can be addressed with appropriate post-filters, by demonstrating that the same approach can be used for reconstructions with TOF reconstructions and also with alternative uptake metrics such as SUV_peak or TLG. Furthermore, we have demonstrated that this additional filtering to match SUV_max actually provides added gains in SNR over parameters to match image voxel COV. However, if the additional smoothing is visually undesirable, an alternative methodology can be used which performs the additional filtering required to match SUV_max only for quantification and is not visualised [41].

This work evaluated the impact of reconstructions that include PSF modelling and/or TOF on lesion classification according to a local protocol by assessing changes in FDG uptake measurements. Two objective strategies for post-filtering were investigated: matching image voxel COV versus matching SUV_max. For matched voxel COV, considerable increases in SUV_max and SUV_peak were observed compared with OSEM. Using post-filters to match SUV_max reduced the discrepancies of either SUV_max or SUV_peak across reconstructions, particularly with PSF modelling. This also resulted in a considerable reduction in voxel variance. Some small discrepancies in patient data still remained when TOF was incorporated, which was not seen in phantom data, warranting further investigation. The TLG metric appears to be more robust in either scheme of post-filtering despite a slightly larger variation in the amount of change, which may be less of a problem considering the large range of TLG data observed. This suggests TLG may be a more suitable metric to adopt instead of SUV_max as the evidence base develops. Gains in SNR were seen in both implementations with the greatest gains seen for matched SUV_max post-filters.