Neuromark PET: A multivariate method for estimating whole brain fMRI guided intrinsic networks and connectomes from fMRI and PET data

Debbrata K. Saha, Anastasia Bohsali, Rekha Saha, Ihab Hajjar, Vince D. CalhounbioRxiv 2024 · DOI

TL;DR

This paper introduces "Neuromark PET," a novel method that leverages resting-state fMRI data to guide the analysis of whole-brain molecular connectomes derived from PET imaging. This approach allows for a more comprehensive comparison of functional and molecular brain networks, revealing both shared and modality-specific patterns, particularly in individuals with mild cognitive impairment (MCI).

What Problem Does This Paper Address?

Neuroimaging studies often analyze functional connectivity (fMRI) and molecular connectivity (PET) separately. While fMRI connectomes are well-established, integrating PET data into a whole-brain network analysis has been challenging. This is partly because PET is typically analyzed using a region-of-interest approach, overlooking the broader network structure. Consequently, direct comparisons between whole-brain PET and fMRI connectomes are rare, limiting our understanding of how these different types of brain connectivity relate to each other and to neurological conditions like MCI.

Methods

The core of this work is the "Neuromark PET" approach, which utilizes Spatially Constrained Independent Component Analysis (scICA). The key innovation is using pre-defined resting-state fMRI networks as spatial priors to guide the scICA decomposition of PET data. This "fMRI-guided" approach allows the PET analysis to identify networks that are analogous to those found in fMRI, creating a common framework for comparison. The authors applied scICA separately to both fMRI and PET data, using the fMRI components as templates for the PET analysis. They then computed functional network connectivity (FNC) for fMRI and molecular network connectivity (MNC) for PET, followed by comparative analyses of the resulting connectomes and spatial maps.

Datasets

The study utilized data from the CEDAR (Candesartan's Effects on Alzheimer's Disease and Related Biomarkers) clinical trial. This dataset included individuals with mild cognitive impairment (MCI) who underwent both MRI and Amyloid PET imaging. Data was collected at baseline and after 12 months, with participants stratified by current medication use and randomized to Candesartan or placebo. For the scICA prior templates, the authors used the Neuromark fMRI 1.0 template, derived from a large dataset comprising the Genomics Superstruct Project (GSP) and the Human Connectome Project (HCP).

Key Insights and Ideas

  • Bridging the Modality Gap: The fMRI-guided scICA approach successfully creates comparable whole-brain connectomes from both PET and fMRI data, enabling direct comparison of functional and molecular brain networks.
  • Complementary Information: While PET and fMRI connectomes show significant modular similarity, each modality also captures unique information. This suggests that a multimodal approach is crucial for a comprehensive understanding of brain connectivity.
  • MCI Network Signatures: The study identified distinct network patterns in individuals with MCI, highlighting differences in connectivity that might be specific to the modality used for analysis.
  • Treatment Effects: The analysis of medication versus placebo groups in MCI patients revealed both common and modality-specific treatment effects on brain connectomes, underscoring the nuanced insights gained from combining PET and fMRI.

Key Results

The study found significant modular similarity between PET and fMRI connectomes, indicating that both modalities capture aspects of brain network organization. However, modality-specific connectivity patterns were also observed. For instance, PET data showed strong positive connectivity between the Cognitive Control (CON) and Default Mode Network (DMN) domains, and between CON and Sensorimotor Network (SMN), while fMRI showed weaker negative connectivity in these areas. Conversely, fMRI revealed stronger positive connectivity between the Auditory Domain Network (ADN) and Sensorimotor Network (SCN), which was not as pronounced in PET.

In the MCI medication vs. placebo analysis, both PET and fMRI showed treatment effects. PET analysis indicated higher connectivity differences in the SMN-SCN, CON-SCN, CON-SMN, and DMN-SCN domains for those receiving medication. fMRI analysis showed higher connectivity differences in the VSN-SCN and CON-VSN domains for the medication group. Notably, some common connectivity differences were found across both modalities in specific domains.

Limitations and Future Work

The authors acknowledge the relatively small sample size used for the statistical comparisons, which limits the generalizability of the findings. They also note that the study was a feasibility study to evaluate the proposed framework for capturing group differences, not a clinical trial. Future work could involve generating PET-specific templates to constrain fMRI data and vice-versa, exploring nonlinear relationships within functional networks (as ICA assumes linearity), and replicating these findings with larger sample sizes. The authors suggest that their approach is generalizable and can be a powerful tool for future PET/fMRI relationship studies.

Summary updated

Abstract

Positron emission tomography (PET) and magnetic resonance imaging (MRI) are both widely used neuroimaging techniques to study brain functional and molecular connectivity. Although whole brain resting functional MRI (fMRI) connectomes (a matrix describing the inter-regional connectivity patterns) are widely used, the integration or association of whole brain molecular connectomes with PET data are rarely done. This likely stems from the fact that PET data is typically analyzed by using a region of interest approach, while whole brain spatial networks and their connectivity (covariation) receive much less attention. As a result, to date, there have been little focus on directly comparing whole brain PET and fMRI connectomes. In this study, we present a method that uses spatially constrained independent component analysis (scICA) (utilizing fMRI components as spatial priors) to estimate corresponding (Amyloid) PET and fMRI connectomes and examine the relationship between them using datasets that include individuals with mild cognitive impairment (MCI). Our results demonstrate highly modularized PET connectome patterns that complement those identified from resting fMRI. In particular, fMRI showed strong intra-domain connectivity with interdomain anticorrelation in sensorimotor and visual domains as well as default mode network. PET amyloid data showed similar strong intra-domain effects, but showed much higher correlations within cognitive control and default mode domains, as well as anticorrelation between cerebellum and other domains. The estimated fMRI informed PET networks have similar, but not identical, network spatial patterns to the resting fMRI networks, with the fMRI informed PET networks being slightly smoother and, in some cases, showing variations in subnodes. To further compare the two modalities, we also analyzed the differences between individuals with MCI receiving medication versus a placebo. Results show both common and modality specific treatment effects on fMRI and PET connectomes. From our fMRI analysis, we observed higher connectivity differences in various regions, such as the connection between the thalamus and middle occipital gyrus, as well as the insula and right middle occipital gyrus. Meanwhile, the PET analysis revealed increased activation between the anterior cingulate cortex and the left inferior parietal lobe, along with other regions, in individuals who received medication versus placebo. In sum, our novel approach identifies corresponding whole-brain fMRI informed PET and fMRI networks and connectomes. While we observed common patterns of network connectivity, our analysis of the MCI treatment and placebo groups revealed that each modality captures modality and group specific information about brain networks, highlighting differences between the two groups in both network expression and network connectivity.
BibTeX
@misc{saha2024neuromark,
  author = {Debbrata K. Saha and Anastasia Bohsali and Rekha Saha and Ihab Hajjar and Vince D. Calhoun},
  title = {{Neuromark PET: A multivariate method for estimating whole brain fMRI guided intrinsic networks and connectomes from fMRI and PET data}},
  year = {2024},
  doi = {10.1101/2024.01.10.575131},
  howpublished = {bioRxiv},
}
Figure