Validation of a spatial normalization method using a principal component derived adaptive template for [¹⁸F]florbetaben PET

Abstract

Quantification may help in the context of amyloid-β positron emission tomography (PET). Quantification typically requires that PET images be spatially normalized, a process that can be subject to bias. We herein aimed to test whether a principal component approach (PCA) previously applied to [¹⁸F]flutemetamol PET extends to [¹⁸F]florbetaben. PCA was applied to [¹⁸F]florbetaben PET data for 132 subjects (70 Alzheimer dementia, 62 controls) and used to generate an adaptive synthetic template. Spatial normalization of [¹⁸F]florbetaben data using this approach was compared to that achieved using SPM12’s magnetic resonance (MR) imaging driven algorithm. The two registration methods showed high agreement and minimal difference in standardized uptake value ratios (SUVR) (R² = 0.997 using cerebellum as reference region and 0.996 using the pons). Our method allows for robust and accurate registration of [¹⁸F]florbetaben images to template space, without the need for an MR image, and may prove of value in clinical and research settings.

Introduction

Fibrillar amyloid-β (Aβ) plaques are a histological hallmark of Alzheimer’s disease (AD) [1] and can be quantified in vivo using the carbon-11 labelled Thioflavin-T derivative Pittsburgh compound-B ([¹¹C]PIB), and related fluorine-18 compounds, including [¹⁸F]flutemetamol [2,3], [¹⁸F]florbetapir [4,5] and [¹⁸F]florbetaben [6,7]. These PET ligands are approved for clinical use by the Food and Drug Administration and European Medicines Agency and have been shown to be of value with respect to diagnostic confidence and patient management in clinical routine practice [8].

At present, only visual assessment of uptake is approved, whereby an Aβ PET scan is classified as either negative (normal) or positive (abnormal) by a trained rater. Evidence suggests, however, that the incorporation of quantitative approaches may improve agreement across raters [9] and aid in the monitoring of treatment effects in anti-Aβ trials [10]. The most common of these approaches is the use of a standardized uptake value ratio (SUVR), a semi-quantitative method involving the normalization of tracer uptake within cortical regions by that within a reference tissue, such as the cerebellum or pons [11]. A requisite for the computation of SUVRs is the demarcation of anatomical regions of interest (ROIs) and the gold standard for this relies on the use of high resolution T1-weighted magnetic resonance (MR) imaging. However, access to MR is often limited in clinical settings, and, in elderly individuals, contraindications for MR imaging are not uncommon [12]. Consequently, PET driven approaches have been developed, using probabilistic regional atlases [13].

A challenge inherent to ¹⁸F-labelled Aβ ligands is that nonspecific binding to white matter is seen regardless of cortical Aβ levels. As a result, uptake patterns across images can result in a systematic bias when a PET driven registration method is used. In an attempt to address this problem, we recently developed an automated PET only registration method using an adaptive template derived from a principal decomposition of [¹⁸F]flutemetamol PET images [14]. As this method allows for robust and accurate normalization of [¹⁸F]flutemetamol images without the need for MRI, it may simplify the clinical use of quantification with Aβ PET. In the present study we here aimed to validate this approach using [¹⁸F]florbetaben PET given its approval for commercial use and due to previous findings showing differences in cortical and white matter retention between [¹⁸F]flutemetamol and [¹⁸F]florbetaben [15,16].

Materials and methods

Subjects

The study population consisted of 132 subjects from an open-label, multicenter non-randomized phase 2 clinical study [17]. 70 subjects had a clinical diagnosis of probable AD, based on the National Institute of Neurological and Communicative Diseases and Stroke-Alzheimer’s Disease and Related Disorders Association (NINCDS-ADRDA) criteria and the revised Diagnostic and Statistical Manual of Mental Disorders IV [18,19]. The remaining 62 subjects were cognitively unimpaired healthy controls, as indexed by a clinical dementia rating (CDR) score of zero, a Mini-Mental Status Examination (MMSE) score ≥ 28 and a z-score ≥ -1.00 for each subject of the CERAD neuropsychological test battery [20]. Controls also underwent structural brain scans with MR imaging; these were judged as “age-appropriate (normal)” and included ratings of cerebral atrophy [21] and cerebrovascular disease [22]. All participants provided written, informed consent. Our study was done in accordance with the Declaration of Helsinki after approval of the local ethics committees and radiation protection authorities of all participating centers.

Adaptive template generation

First, [¹⁸F]florbetaben images were co-registered to their corresponding MR images and spatially normalized to the Montreal Neurological Institute T1 template using the MR driven approach included in SPM12 (http://www.fil.ion.ucl.ac.uk/spm). This approach allows for a global deformation field using a linear combination of low-frequency basis functions. Two sets of SUVR images were then created using a composite cortical region-encompassing brain regions typically showing high Aβ load in AD, including frontal, temporal and parietal cortices, precuneus, anterior striatum, and insular cortex-and the pons and cerebellar cortex as reference tissues [23]. Complete details pertaining to the creation of the synthetic template can be found in the original publication [14]. In brief, principal component images were calculated by singular value decomposition of the SUVR images for all subjects. A synthetic template, I_Synthetic, could then be modelled by a linear combination of the first principal component image, I_PC1, and the second principal component image, I_PC2, according to:

I_Synthetic = I_PC1 + w I_PC2

where a negative value of w generates a template with an appearance towards the Aβ-negative range and a positive value of w generates a template with an appearance towards the Aβ-postive appearance. The synthetic [¹⁸F]florbetaben template could now be utilized by the registration algorithm described in the original publication [14] which incorporates both the weight (w) and parameters for spatial transformation in the optimization. This allows the registration method to iteratively find the best set of spatial transformation parameters for a given patient’s [¹⁸F]florbetaben scan to fit the optimal template for this particular scan. Once converged, refinement of the registration of the brainstem and cerebellum is performed.

Refined reference region registration

Due to the low spatial resolution of PET, a refined local rigid-body registration approach was implemented for the pons and cerebellum. A binary mask covering the brain stem and cerebellum was first created. The binary mask was then smoothed using a 3-dimensional Gaussian kernel. The voxel intensities of the smoothed mask were then used as weights for the local rigid-body registration; a zero-value voxel of the smoothed mask hence leaves the corresponding voxel in the subject’s image unaffected, while a voxel with value one will give a full contribution of the calculated rigid-body transform. This ensures that there are no discontinuities in the final registered image.

Spatial normalization and comparison with MR based registration

After co-registration to their corresponding MR scans, [¹⁸F]florbetaben images were spatially normalized to template space using two approaches: first, using spatial transforms derived from the MR driven registration in SPM12 and, second, using the principal component template registration method. In order to evaluate our method, coefficients of determination (R²) were calculated between SUVR values derived using the principal component template and the SPM12 MR driven registration method. Absolute differences between SUVR values were also calculated using both approaches.

Results

Table 1 shows the main demographic variables for the study population. Spatial normalization was successfully achieved for all [¹⁸F]florbetaben images, with no manual corrections required. Representative [¹⁸F]florbetaben images are shown in Figure 1. Comparison of quantification results using the [¹⁸F]florbetaben-driven principal component generated adaptive template registration and MR-driven SPM12 registration showed good agreement, with high R² values using both the cerebellum (0.997, P < 0.001) and pons (0.996, P < 0.001) as reference regions (Figure 2). Mean absolute differences between SUVRs calculated using both methods were low using the cerebellar cortex (AD = 1.702%, controls = 1.843%) and pons (AD = 1.724%, controls = 1.659%). The first and second principal component derived images, corresponding approximately to the average of all images and the difference between Aβ-positive and Aβ-negative images, are shown in Figure 3A and 3B, respectively, with a selection of generated adaptive templates shown in Figure 3C.

Table 1.

Demographic variables for the study cohort

	AD	Controls	P *
N	70	62
Age, years	70.07 ± 7.68	68.10 ± 6.58	..
Female, N (%)	31 (44%)	37 (60%)	..
Education (years)	12.39 (7.68)	14.54 (3.56)	..
MMSE	22.64 (2.61)	29.26 (0.77)	< 0.001
Word-list memory	10.76 (5.02)	22.53 (3.70)	< 0.001
Word-list recall	2.27 (2.11)	7.95 (1.65)	< 0.001

Data is presented as mean ± standard deviation or as n (%). MMSE = mini-mental state examination.

^*Group differences were tested for significance with the two-sided Fisher test for ordinal and the Wilcoxon test for continuous variables.

.. = No significant difference between groups.

Figure 1.

Representative coronal [¹⁸F]florbateben images showing an Aβ-negative (left) and an Aβ-positive (right) scans.

Figure 2.

Scatterplot showing the comparison of [¹⁸F]florbateben SUVR values between PCA and SPM12 driven approaches using the cerebellum (A) and pons (B) as reference regions. R² values (P < 0.001): cerebellum, 0.997; pons, 0.996. HC = healthy controls, AD = Alzheimer’s disease dementia.

Figure 3.

The first and second principal components are shown in (A) and (B); these represent, respectively, the average of all images and the difference between between Aβ-positive and negative images. Synthetic template images showing characteristic [¹⁸F]florbetaben uptake pattern going from most negative (upper left) to most positive case (lower right) are shown in (C).

Discussion

Building on the previous publication [14], in which the proposed PET driven adaptive template registration method was originally described using [¹⁸F]flutemetamol, we here show that our approach works equally well using [¹⁸F]florbetaben PET and a similar population.

In light of recent findings that clearly support Aβ imaging having a significant impact on the clinical management of patients with mild cognitive impairment or dementia [24], the clinical use of Aβ PET is likely to increase. Though currently available commercial Aβ tracers are approved for visual reads only, relevant levels of variability in both rater accuracy and between-reader agreement have been reported [25,26]. While this topic has yet to be sufficiently explored, there is increasing evidence to suggest, increasing evidence to suggest that quantitation should be added to visual read in certain situations (e.g. inexperienced reader, borderline scans) [27]. Existing data indicates that quantitation improves both metrics [9,28] with several commercial software packages available to calculate SUVRs, for example. Automated quantitation using the proposed registration method may thus aid in increasing reader certainty and further the clinical adoption of Aβ imaging, including within the context of future clinical trials testing anti-Aβ compounds in individuals who are Aβ PET negative but whose brain Aβ levels are rising [29]. Furthermore, as the proposed method performs as well as SPM12’s MR based approach, it may simplify study protocols by removing the requirement for a separate T1-MR scan.

By comparison to other existing methods that employ adaptive [30] and principal component derived templates [31] for use with Aβ PET data, our approach carries a low computation cost (~20 seconds to process a single scan, as compared to > 6 hours for the method by Fripp and colleagues) and appears to generate a more accurate template. In addition, by contrast to the method developed Lundqvist and colleagues [30], for which a patent has been filed, our method is non-proprietary meaning that it can be easily distributed across interested parties within the field, potentially facilitating research into its application with Aβ tracers in different settings. As our primary focus is the potential use of this method in clinical settings, we have not evaluated our method using [¹¹C]-Pittsburgh compound B ([¹¹C]PiB) due its characteristic short half-life precluding its use clinically. As we have previously shown that our method works well with [¹⁸F]flutemetamol [14]-essentially the ¹⁸F-labelled version of PiB [32], with the two tracers shown to match closely in both controls and AD [33]-we think our method would work equally well with [¹¹C]PiB. Though the availability of PET/MRI systems, allowing for the simultaneous acquisition of PET and MRI data, would circumvent the need for PET driven methods, the number of such platforms is quite low in comparison to stand-alone PET or PET/CT scanners [34].

Limitations of this study include the lack of neuropathological confirmation in subjects with a clinical diagnosis of probable AD. [¹⁸F]florbetaben PET, however, has previously been shown to have high sensitivity and specificity for detecting histopathology-confirmed neuritic Aβ plaque pathology [35]. Moreover, due the absence of cases with borderline changes that are difficult to classify visually, we were unable to examine the added benefit of our adaptive template method, and quantification in general, over visual read; this is, however, the subject of ongoing work using larger data sets. In addition, we did not assess the performance of our method in patients with non-AD neurodegenerative disorders such as frontotemporal lobar degeneration, which can be characterized by marked focal atrophy [36]. Finally, future studies are also required to address the extent to which the proposed method may apply to other PET tracers, including those for tau.

Conclusion

Our findings validate those originally reported for [¹⁸F]flutemetamol PET, indicating that the proposed method, which allows for a robust and accurate PET driven normalization procedure, applies equally well to [¹⁸F]florbetaben. The proposed method stands as a promising strategy that may simplify the implementation of quantification in clinical settings.

Future studies are required to address this and the extent to which this method may apply to other PET tracers, including those for tau.

Disclosure of conflict of interest

None.