The future of fMRI in cognitive neuroscience - Russell Poldrack

Post date: Jun 14, 2016 2:45:26 PM

The future of fMRI in cognitive neuroscience:

A summary of Russell Poldrack’s talk at the Wellcome Trust Centre for Neuroimaging Brain meeting, 10/06/2016

By Annie Brookman (@nniebrookman) and Lucía Magis-Weinberg (@luciamawe)

Russell Poldrack (Stanford University) gave a talk on the future of functional magnetic resonance imaging (fMRI) in cognitive neuroscience, which fitted within the wider context of the current “crisis” described in psychological science. As such, it echoed recent concerns about the lack of reproducibility, with a particular focus on fMRI studies. Neuroimaging studies have particular pitfalls that make them the “perfect storm for irreproducibility”: given high cost and other limitations, these studies are likely to be underpowered to detect small effect sizes. In addition, there is very high methodological flexibility in data processing and analysis, as well as incomplete reporting of methods due to space constraints in journals. Once a skeptic of study preregistration, Poldrack spoke of the benefits of adopting this practice for all fMRI studies in his own lab, and encouraged others to do the same. This practice can help to guard against the statistical problems encountered in cognitive neuroscience, including many forking paths and p-hacking.

The current approach in fMRI looks at which brain areas are activated during different tasks. Poldrack used the example of the anterior cingulate, which is activated in about a third of fMRI studies – knowing this tells us very little about the precise role of the anterior cingulate. Rather, in order to move the field forward, we need to see if we can predict the psychological task participants performed from the brain activation.

This approach is the direction Poldrack’s lab is taking and he argues that looking at data in this way can tell us more about our psychological theories (Poldrack, 2012). For a given psychological concept, such as cognitive control, we can use different tasks, inside the scanner, which are expected to measure it. We can then see if the similarity in the functional brain images of these tasks relates to the theorised ontological (a higher order map or knowledge base) similarity of these tasks. If we find that the images are not related, then this suggests that our understanding of cognitive control is wrong (Figure 1). Further, it is important to consider how well experimental tasks map onto the concepts we are measuring. For example, which specific elements of cognitive control does the n-back task tap into (i.e. inhibition, updating)?

Figure 1: Ontology-based analysis of brain data activations across studies and experimental paradigms (source).

Poldrack referred to other sciences where new concepts have emerged over the years, based on experimental work and technological advances. In psychological science on the other hand, we’re still using the same terms and asking the same questions that were being asked by William James in 1890 in Principles of Psychology. Moving towards ontology-based meta-analyses which use fMRI data collected in many labs to distinguish between competing models of cognitive processes, may enable us to really test these psychological concepts, and come up with new terms that better describe cognition (see Figure 1).

Reflecting on the past 25 years of the development of fMRI to study human cognition, Poldrack recognised that fMRI has come a long way. However, he stressed that to push the field forward, important improvements to fMRI methods and study design are needed, to ensure reproducibility and mapping onto ontological concepts. Poldrack mentioned a series of tools and practices which may help improve neuroimaging research:

1) Problem: Majority of fMRI experiments are underpowered

a. Neuropowertools à calculate the necessary sample size for an fMRI study to obtain a certain predefined level of statistical power

b. OpenfMRI and Neurovault à public repositories to promote data sharing and allow meta-analyses with greater combined power

c. Neurosynth à a platform for automated synthesis of fMRI studies

2) Problem: Incomplete reporting of methods + methodological flexibility

a. Nipypeà platform to create a workflow of data analysis to keep track of all the steps and decisions taken

b. COBIDAS report à guideline on best practices and reporting produced by the Organisation for Human Brain Mapping

c. Introduction of default settings in programmes to increase similarity across studies in both methods and analysis

3) Problem: Lack of reproducibility

a. Encourage data and code sharing through platforms such as Github

b. Virtual machines à for complex analysis that cannot be easily reproduced on standard computers

c. Introduce incentives for sharing data online and preregistering