We characterised the Marseille database of [18F]FDG PET-CT and T1 MRI and then compared it to the recently described CERMEP iDB-MRXFDG database [1]. We investigated the harmonisation of both databases in the context of the example application of detection of epileptogenic lesions in patients with FCD.
The Marseille database has previously been used in various contexts [24, 33] but has not undergone detailed characterisation as CERMEP iDB-MRXFDG [1].
In contrast to structural MRI, there is a dearth of databases of healthy controls with [18F]FDG-PET images published or made available to the scientific community (see introduction), especially in the younger age range which is essential for research into the epilepsies [18], disorders of consciousness [34, 35], and normal brain function [36,37,38]. The Marseille database meets this need since it is composed of subjects ranging from 21 to 78 years old (50 ± 17 mean ± SD) and can be used in a large panel of applications. Leave-one-out SPM analysis was conducted to ensure the normality within the database, resulting in 4 subjects (7%) being removed from further analyses, in line with Waterschoot et al. [6] where 4% of subjects were removed via a similar process.
While overall the two databases were comparable in terms of [18F]FDG uptake patterns, we also found differences due to various factors.
First, global image appearance was different, with greater spatial resolution and better contrast in MRXFDG (Fig. 2). We demonstrated that this was due to different smoothness of the reconstructed images, expected from differences in acquisition systems and reconstruction algorithms. Reducing the differences in the image resolution (visually and quantitatively, see Table 2) was successfully achieved by smoothing the MRXFDG database with a larger Gaussian kernel, similarly to previous work [13, 14].
Second, different SUVs were obtained for the two databases (Fig. 5, row 1). As SUV is semi-quantitative, differences are expected between scanners and acquisition protocols, in particular for uptake delay and acquisition duration. For the Marseille database, the same dose was injected for all participants, which will explain the higher inter-subject variability compared to adjusting dose by weight, as in MRXFDG [1]. We also observed a deviation of the boxplots towards high SUV values for the Marseille database, especially in the occipital lobe, likely corresponding to more visual stimulation during the tracer uptake period. This was also borne out by the voxel-based comparison and occurred despite similar instructions, namely to rest with eyes closed. Finally, for MRXFDG, we observed three outliers with lower regional SUVs. They also had reduced inter-region variability suggesting they had not followed the fasting instruction, making the PET data less informative. Normalizing within-image via SUVr (Fig. 5, row 2) reduced between-subject SUVr variability for the Marseille database, as variability due to fixed-dose injection was compensated for. The MRXFDG database had higher inter-region SUVr variability, in particular in regions of the frontal lobe and central structures. This could be a consequence of the longer uptake period for MRXFDG (50 min vs. 30 min for Marseille) and the higher spatial resolution for MRXFDG which will enhance contrast between regions.
While intra-regional COV distribution was similar for the two databases (Fig. 5, row 3), we note: (1) intra-regional COV for the caudate nucleus in MRXFDG was higher than for all other regions, with a number of outliers. This may be due to the regional segmentation of MRXFDG via a single registration of a maximum probability atlas derived from the Hammers Atlas Database [26, 27] to the subject space which can result in caudate nucleus labels including voxels of the surrounding lateral ventricles, whereas the Marseille database was segmented via the 30 individual atlases from the Hammers Atlas Database (i.e., identical region definitions) but with the more accurate multi-atlas approach MAPER [30]. (2) Temporal lobe regions had higher intra-regional COV, higher variability, and a larger number of outliers for the Marseille database. This may be due to the shorter field of view of the scanner used for the Marseille database (15.7 cm vs. 22 cm for MRXFDG) which may produce higher COV in the temporal lobe which is closer to the edge of the field of view [39].
Third, correlations between age and [18F]FDG uptake were shown in each database, with positive correlations in the white matter (likely due to SPM normalizing to the global mean) and negative correlations in or near fissures widening with age. Leaving out subjects ≥ 60 years old minimised these in each database.
We used individualised smoothing of the two databases, based on the images themselves, ahead of SPM analysis. Another approach is to use standardised phantoms to derive a smoothing factor [40] to account for different reconstruction algorithms, later expanded to additionally account for different hardware (EQ.PET, https://marketing.webassets.siemens-healthineers.com/1800000002580275/475f0ba8d270/EQ-PET_WP_1800000002580275.pdf). This phantom-derived approach has also been applied for comparing half-body (i.e., non-brain) images across different reconstructions and PET systems [41]. However, as these methods only work on spatial resolution via Gaussian smoothing, brain PET with its strongly varying SUVs would still require normalisation for global uptake, as performed via a full factorial design in this study.
We used the insights gained by the database analyses to explore combination of databases in an experiment designed to detect known abnormalities (FCD) (Table 3). Likely due to the global intensity differences, simply combining both databases decreased performance compared to each database on its own despite the much larger number of scans. Accounting for differences in intensity between the two databases in a full factorial design outperformed either database used separately. This was true for any type of accounting for age (via linear or quadratic covariates or by removing participants over 60 years old). However, excluding the oldest subjects in the databases informed by the age correlation experiments yielded the best results in locating FCDs (up to 71%); when all control participants were used, using age as a linear covariate was preferable to a quadratic fit but remained inferior to the exclusion of older controls.
Other studies have detected histologically proven MRI-negative FCD on PET SPM5 analysis against 30 controls in 72% [17], in line with our study. Detection rates on visual analysis were 78% for PET-only and 95% after coregistration with MR [17]. In another larger study from the same group, visual analysis blind to all electroclinical data led to a detection rate of 44% increasing to 71% with electroclinical data and to 83% after coregistration with MRI [18], suggesting the potential of using SPM or similar analyses of PET with larger databases jointly with MR data to increase detection rates in future studies [42].
Differences between databases could potentially be reduced by adapting the reconstructions to be more similar—for example, MRXFDG could have been re-reconstructed without the spatially varying point spread function. However, as PET hardware [43] and PET reconstructions [44] continue to evolve, such differences are unavoidable, and we opted to use both databases as provided. It is reassuring that we found similar performance for patient data acquired on three different systems, especially as we chose not to repeat the resolution matching for each of the three patient scanners which would be prohibitively time-consuming in practice.
We only investigated two [18F]FDG brain PET databases. As explained above, few are publicly available, and often the focus is on the older age group used for dementia studies. We found the best performance when restricting the age range, suggesting that e.g. the ADNI database exclusively containing data from ≥ 60 year-olds would be less suitable for applications like epilepsy workup or basic science studies in younger people.
We used a single sample of patient data as part of a Service Evaluation. However, there was variability from using three scanners, and the sample had the advantage of a histological ground truth.
Our work contributes to the understanding of data structure and descriptive statistics of [18F]FDG PET databases. These have been studied less than MR databases but are crucial for detectability of pathology as shown here. Fully characterising [18F]FDG databases enables principled choices for combining them. In our work we have explored some harmonisation methods and shown a substantial improvement in the detectability of lesions, using FCD as an example. Future work can now focus on combining multiple databases and perfecting methods to decrease differences between databases, e.g. regional normalisation. In addition, while SPM is a well-understood analytical method, our work can be used beyond mass univariate testing towards combining databases in machine learning or artificial intelligence (AI) applications. As an example, we have recently created large databases of simulated healthy and FCD-typical brain FDG PET images from MR phantoms [45] and used them to create higher-quality PET images from clinical standard-resolution PET via convolutional neuronal networks, resulting in better quality metrics and FCD detectability. Healthy control databases and AI methods can also be used to simulate pseudo-healthy FDG PET from patients’ MRIs which can then be subtracted from the real clinical PET to reveal lesions [42, 46]. Real databases such as the ones examined in this paper fulfill the need to take into account the physics of image acquisition and reconstruction and inform of descriptive image statistics encountered in practice.
