The history and future of resting-state functional magnetic resonance imaging

Abstract

Since the discovery of resting-state functional connectivity in the human brain, this neuroimaging approach has revolutionized the study of neural architecture. Once considered noise, the functional significance of spontaneous low-frequency fluctuations across large-scale brain networks has now been investigated in more than 25,000 publications. In this Review, we provide a historical overview and thoughts regarding potential future directions for resting-state functional MRI (rsfMRI) research, highlighting the most informative analytic approaches that have been developed to reveal the brain's intrinsic spatiotemporal organization. We review the collaborative efforts that have led to the widespread use of rsfMRI in neuroscience, with an emphasis on methodological innovations that have been made possible by contributions from electrical and biomedical engineering, physics, mathematics and computer science. We focus on key theoretical and methodological advances that will be necessary for further progress in the field, highlighting the need for further integration with new developments in whole-brain computational modelling, more sophisticated approaches to brain-behaviour mapping, greater mechanistic insights from concurrent measurement of neurophysiology, and greater appreciation of the problem of generalization failure in machine learning applications. We propose that rsfMRI has the potential for even greater clinical relevance when it is fully integrated with population neuroscience and global health initiatives in the service of precision psychiatry.

Fig. 1 |. Timeline of key events in the history of rsfMRI.

The first study to demonstrate functional connectivity in the human brain using rsfMRI was published in 1995. Nearly 30 years later, this approach has been used to identify reproducible large-scale functional brain networks, investigate lifespan changes in brain network configuration, and track clinical outcomes. Population neuroscience initiatives such as the Human Connectome Project and the Adolescent Brain Cognitive Development Study collect and disseminate large rsfMRI datasets for research purposes.

Fig. 2 |. Reproducible large-scale functional brain networks.

Multiple large-scale functional brain networks can be reliably identified using rsfMRI. Although there are inconsistencies in network nomenclature, most brain atlases derived from rsfMRI data include visual, somatomotor, attention, executive control, salience and default networks. Adapted with permission from ref. , American Physiological Society and from ref. , Springer Nature Ltd.

Fig. 3 |. rsfMRI data analysis workflow.

Pre-processing of rsfMRI data typically involves quality control and denoising prior to subject level analysis. As with task fMRI, group-level analyses are often performed using the metrics derived at the individual subject level. fALFF, fractional amplitude of low-frequency fluctuation; ReHo, regional homogeneity analysis; ROI, region of interest.

Fig. 4 |. Machine learning and precision psychiatry.

Metrics derived from rsfMRI features (for example, ICA components, region of interest time series and graph metrics) can be used within machine learning (ML) frameworks to predict phenotypes of interest, including cognitive, behavioural and disease-related outcomes.