Resting-State Networks as Simultaneously Measured with Functional MRI and PET

DISCUSSION

In the present study, a number of known RSNs could be identified in ¹⁸F-FDG PET data. Among them are the primary and secondary visual, sensorimotor, auditory, cerebellar, DMN, and CEN. Spatial similarity of these RSNs with those extracted from fMRI data in the same healthy subjects appeared to be overall modest. However, given that the 2 imaging modalities record different aspects of neural activity, the results do argue for the common neural substrate of RSNs.

Overall, the RSNs as known from fMRI were less accurately identified in PET data. That is, most networks consisted of clusters that were smaller, worse separated from each other or contained additional clusters. This observation is in line with previous studies in both humans (11) and animals (25). Further, roughly each third independent component (7/20) in the PET data appeared to be noise or could not be classified. Although in a few of these components (e.g., top and bottom in Fig. 2, Noise) major clusters were located in the brain, they were too discrete to enable assignment to a particular network. In contrast, only 3 noise components were found in fMRI data. Correspondingly, a higher number of meaningful RSNs could be extracted from fMRI than from PET data. As for the spatial similarity between the matched networks, it was at best moderate.

The less accurate identification of RSNs in PET data as well as its modest overlap with the fMRI-based network are not unexpected. First, ¹⁸F-FDG PET and BOLD fMRI capture partly different aspects of neural activity. While ¹⁸F-FDG PET directly measures activity-dependent glucose consumption, BOLD signal stems from interplay among cerebral blood flow, volume, and cerebral metabolic rate of oxygen (26). Second, ICA on fMRI data assigns voxels to the same component when they share similar temporal dynamics (i.e., covariance of BOLD-signal fluctuations), whereas ICA on PET data assigns voxels to the same component when they share similar energy levels (i.e., normalized ¹⁸F-FDG uptake) across subjects. Hence, it is far from obvious that similar RSNs can be detected by both imaging modalities.

As compared with fMRI, routinely acquired ¹⁸F-FDG PET data represent an averaged snapshot of tracer uptake over a certain time frame, typically 10–20 min, making a within-subject analysis of (static) PET images such as those in a BOLD time series impossible. ICA on such PET data thus identifies sources of intersubject covariance, making the results noisier and stronger dependent on the sample size. This is well in line with a previous study on a comparison between first- and second-level ICA on fMRI data (27). Specifically, the authors could extract the known RSNs from the intersubject covariation among highly distilled features such as amplitude maps. The results were noisier than those from the first-level ICA, still yielding strikingly similar patterns of intrinsic brain connectivity. It must be noted, however, that like fMRI ¹⁸F-FDG PET data can be acquired and analyzed dynamically (28). Spatial similarity of RSNs extracted from such data—when starting acquisition simultaneously with injection of ¹⁸F-FDG—may be higher because of the contribution of a blood flow signal. This interesting issue should be addressed by future studies.

The first attempt to extract RSNs from ¹⁸F-FDG PET data was performed by Di and Biswal (11). In contrast to our study, they quantified the networks from different groups of subjects and compared them only visually. Still, they found many of the known RSNs in the PET data. Interestingly, Di and Biswal had troubles identifying neurocognitive networks with anterior–posterior connectivity, for example, the left CEN (frontoparietal) and DMN. Our study in part supports this observation. That is, the CEN could be identified in our PET data as a single map, with fair to poor spatial overlap with the analogous fMRI-based networks. Although the DMN map consisted of rather small clusters, the pattern was anatomically clear, and the spatial correspondence with the fMRI-derived map was relatively high. The less accurate identification of these RSNs in Di and Biswal (11) may well be related to the fact that the image data under their study were part of a multicenter project with 3 different acquisition protocols (ADNI, http://www.adni-info.org). Besides, subjects in the study by Di and Biswal (11) were substantially older (mean age, 77.2 vs. 54.5 y in our study). Previously, we could readily identify the DMN in a group of young (mean age, 27.1 y) healthy subjects (12).

Apart from the dually detected RSNs, some networks were identified exclusively in fMRI or PET data. Among 7 fMRI-only RSNs, anterior insular and temporopolar do not belong to the established ones. Three further RSNs, (anterior) primary visual, lateral motor, and (anterior) DMNs are variants of those detected in both imaging modalities (Fig. 1). Hence, the salience and spatial attention networks as established ones could not be detected in the present PET data. Of note, the salience RSN was described in Di and Biswal (11). The discrepancy may be explained by a lower number of subjects or components in the present work. On the other hand, 4 RSNs were identified exclusively in the PET data. Along with an additional (secondary) visual RSN, there were 2 subcortical networks. These may have been missed in the fMRI data, because basal ganglia networks are reliably detected at a higher number of decompositions (29). Although 3 of the PET-only RSNs above have been described in the fMRI literature, we are not aware of any fMRI RSN encompassing the mesial parietal and prefrontal cortex as a single cluster. Interestingly, a very similar component has been reported previously (Fig. 1A (11); (30)). The latter study analyzed covariance in regional gray matter density in a group of healthy subjects. Along with the network above, the authors described frontal, posterior cingulate/precuneus, visual, basal ganglia, and cerebellar networks. Thus, the number of fMRI-similar RSNs was lower than in our study, despite a much larger cohort (n = 603) (30). These observations may indicate that the metabolic level of brain connectivity extends between structural and functional levels, expressing features of both. A biologic significance of the “mesial” network is yet to be established.

This study has several limitations. Because of the limited sample size, a higher number of components (in PET data) could not be calculated. For the same reason, we applied a rather liberal z threshold of 2.0 for visualization and quantification, bearing a risk of false-positives. Second, as part of the dedicated imaging protocol subjects were studied under different conditions with respect to visual perception, that is, with eyes open and closed (15). That study found that glucose metabolism in the visual cortex was higher in the eyes-open condition, contributing to a stronger functional connectivity between the visual and salience networks (15). Thus, this confound may have influenced the spatial similarity between the visual networks in the present study. Indeed, our additional analyses revealed a stronger metabolic connectivity within the visual network in the eyes-open condition (with no substantial difference in functional connectivity). Consequently, a higher similarity between the PET- and fMRI-based visual networks was found in the eyes-open subgroup. Finally, the Dice coefficient was used for the quantification of a spatial overlap between PET- and fMRI-based RSNs. This coefficient is a well-established measure of binary similarity and the most used metric in evaluating 3-dimensional medical image segmentation (31). Still, there are other similarity coefficients in the field available (31), both binary and continuous, that may have produced different results. Yet, this methodologic issue lies beyond the scope of our paper.

Besides the neurobiologic significance and a contribution to the developing field of cometomics (25), the present results may be of clinical value. Namely, the network information as obtained through ICA on PET data at the group level can be further used with diagnostic purposes at a single-subject level. This has been shown in many PET studies applying principle component analyses (32) and recently ICA (20,21). In particular, it is possible to calculate a degree to which each individual subject/image contributes to a given pattern/network. This so-called loading coefficient, as extracted from a single RSN or a combination of networks, can be used for single-subject analyses, for example, in a differential diagnosis of neurodegenerative disorders.