This study aimed to investigate the use of rCBF images derived from dynamic PIB PET scans as an alternative to FDG through an SSM/PCA analysis for classifying AD patients versus HC subjects. The metabolic pattern generated by FDG images has been identified as an appropriate tool to identify AD patients [2, 20, 23]. Due to the correlation between metabolism and blood flow [33], a similar characteristic AD pattern was expected from rCBF images [20]. Furthermore, rCBF measures have been shown to correlate, at least partially, with cognitive impairment in AD patients [34], increasing the importance of these images in AD classification. Therefore, the use of amyloid-derived rCBF images as a proxy for an FDG scan is an attractive alternative that may reduce study costs, decrease patient discomfort, and minimize radiation exposure because both rCBF and specific binding information can be driven from a single dynamic PET scan.

The generated DPs presented a cortical decrease in flow [R1 and ePIB(20–130 s)] and metabolism (FDG) in AD patients when compared to HC subjects. As previously mentioned, this similarity between patterns was expected due to the existent relationship between metabolism and blood flow [5, 9] and due to previous studies using voxel-based univariate analysis of the images [9]. Still, some differences between the patterns were found in regions that are already known to be hyperperfused, which is consistent with previously published results [9, 20, 35]. However, the most interesting point to observe from this analysis is the difference in pattern from the ePIB(1–8 min) as compared to the other methods (Fig. 1). This time interval between 1 and 8 min has been recommended as the time interval with the best visual correlation with FDG, based on the correlation of its regional values and those of FDG images [16]. However, the generated AD-DP in this study shows a pattern more closely related to the amyloid deposition pattern of AD patients, which showed increased signal in grey matter cortical regions in scans acquired later during the scan [20]. Furthermore, a previously published SPM voxel-based analysis of the same images showed that this time interval was also not able to differentiate between patients and controls [9]. This result is consistent with the hypothesis that this time interval is too long, and the signal is already affected by Aβ binding, resulting in an image that does not reflect purely rCBF [9], Therefore, for the remaining of this discussion, ePIB(1–8 min) data will no longer be addressed.

The distribution presented in Fig. 4 showed that all methods presented significantly different scores between the AD and HC groups. However, only FDG was capable of distinguishing the group of AD patients from MCI+ and MCI− subjects. Even though MCI + is also known as ‘MCI due to Alzheimer’, no method was able to generate subject scores that were statistically different between MCI+ and MCI− or HC groups. This could be due to the fact that rCBF images are less sensitive to subtle changes than metabolic scans [9] and, therefore, metabolism images might be able to capture subtle changes that are not reflected by rCBF. However, a larger dataset of patients might increase the stability of the DP, which may allow for these subtler changes to be captured. Furthermore, an independent testing group of AD and HC subjects could better assess the accuracy of the generated AD-DP. Other automated methods for image assessment for AD classification have shown to be more sensitive to assess disease progression of MCI patients [1, 36]. Moreover, the same set of subjects used in this study were previously analysed using an automated tool for assessment of AD, which resulted in a good contrast between AD patients and HC subjects but could not distinguish between MCI subjects as well [10]. Therefore, these patients were added to this analysis to evaluate the use of the SSM/PCA technique to evaluate their AD-DP expression.

Moreover, Fig. 4, in combination with Fig. 5, shows a smaller range of rCBF scores when compared to metabolism. This suggests that the reduction in FDG uptake in AD patients is greater than the reduction in R1 and ePIB(20–130 s) when compared to HC subjects. Therefore, the AD-DP expression through the subject scores results in smaller values for rCBF methods and ensues a greater bias for larger scores when compared with FDG. Moreover, the more extensive range of scores shown in Fig. 4, the higher AUC, and the largest correlation to metabolism scores indicate that R1 might be the most suited rCBF method for generating an AD-DP through SSM/PCA as a proxy for FDG scans. In addition, R1 images provide a pure measure of relative radiotracer flow due to the inherent quantitative aspect of parametric images, while SUVR shows a mixture of specific and non-specific binding, free tracer in tissue, and blood signal, which might have affected the ePIB(20–130 s) performance for subject identification [37]. Finally, Fig. 5 shows that the bias on rCBF scores when compared to FDG is negative for higher scores and positive for lower scores, which further indicates that rCBF scores are not as robust as the FDG scores when differentiating between AD and HC. Further studies using other techniques for estimating rCBF images are worthy of exploration, as some of them might improve results even for MCI subjects. However, as SUVR is the most frequently used approach in clinical research due to its simplicity, it was the one chosen for this study.

Although the results presented in the previous section suggest a good correlation of AD-DPs and subject scores between rCBF and FDG, it is important to mention that, from a physiological point of view, there is no perfect equivalence between them. However, these results sustain the use of SSM/PCA as a classification technique to be used not only with FDG PET scans and AD, but also for other types of images and diseases. Furthermore, this analysis was performed using PIB as a radiotracer, but similar results can be expected for other 18F-labelled amyloid radiotracers such as [18F]Florbetapir, [18F]Florbetaben, and [18F]Flutemetamol [38]. Yet, further research is necessary to confirm this. Moreover, this study was performed on a relatively small sample of subjects. Larger cohorts might yield more accurate results by providing a more stable pattern, which might find statistically significant differences where this study found only a trend, such as a significant difference between AD and MCI+ patients. In addition, longitudinal datasets might be useful for analysing the efficiency of SSM/PCA in predicting the conversion of MCI subjects to AD. Finally, SSM/PCA might be an interesting technique to be further explored for the identification of patients both in the clinic and in research settings, since it allows for testing of new subjects that are not related to the ones used to generate the DP.

Amyloid PET imaging is already used in clinical settings for the assessment of deposits to identify AD patients. A dynamic scan may provide rCBF images that, in some cases, might be enough for an assessment on neurodegeneration, dismissing the need of a second FDG PET scan for this analysis. Performing a single scan reduces patient burden and exposure to radiation, and costs in terms of FDG production and scanner time. Furthermore, in a clinical setting, the assessment of PET images is mostly performed visually by an expert neuroradiologist or neurologist. Previous studies have found that visual assessment relies on the reader’s experience and even then, the agreement between the readers is not always optimal [39]. An automated technique such as the SSM/PCA has the potential of resulting in more consistent and accurate subject assessment. Moreover, by providing scores that quantify the expression of the DP in a subject, this technique may be used for classifying subjects in stages. Finally, SSM/PCA allows for the generation of a centre-specific DP that can be generated with both FDG and rCBF images. This enables the comparison of rCBF images with more than just an FDG pre-set database that is available in software packages available for the clinic. However, the technique proposed in this study requires a local dataset of patients and controls for the generation of the disease characteristic patterns, as this technique has been shown to be sensitive to scanner and reconstruction effects [40], which may not be available in all imaging centres.

While rCBF images might be used as a surrogate for FDG PET in more extreme cases, it has been shown not to fully capture the smaller changes in patients when compared to controls. This is a key point when identifying patients in a preclinical phase, for example. Amyloid imaging is the first recommended imaging modality (after structural imaging) in cases of suspected AD diagnosis [41]. However, in cases where amyloid imaging is inconclusive, to exclude possible amyloid positive co-pathologies, to make a short-term prognosis, or to make a better assessment of the stage of the disease, FDG PET is still recommended. By performing a dynamic scan instead of a static one, it is possible to generate rCBF from the same scan, avoiding putting the subject through the burden of a second PET scan.