Synopsis

Blood Oxygenation Level Dependent functional MRI (BOLD fMRI) is a powerful non-invasive probe of the brain in health and disease. Deoxyhemoglobin serves as an endogenous paramagnetic susceptibility contrast agent, sensitizing MRI data to local hemodynamic changes concomitant to neuronal activity. Reactive hyperemia provides increases in brain perfusion that are more generous than needed to meet demands for increased oxygen, causing reductions in local concentrations of deoxyhemoglobin, which improves local homogeneity of the static magnetic field, thereby increasing gradient-echo magnetic resonance signals from water protons. Well, that’s how we get our BOLD data. So, how should we analyze it?

People, let’s be careful out there! Given recent concerns about a “replication crisis” in science and the prevalence of so-called “p-hacking” (or “data dredging”), the field of fMRI may be highly susceptible to such worries, because of factors including the large dimensionality of the data, the popularity of massively univariate analyses, and a lack of attention to “known unknowns”.

This talk will address steps that you can take to ensure that your conclusions are valid. We will review:

• The importance of quality assurance. If you never look at your raw data, how do you know what you’re missing? There could be distortions and voids in your raw images, denying you useful signals from various regions of the brain, even though they are within the imaging field of view.
• The importance of knowing what question you’re trying to answer. What is the difference between prediction and association? Between hypothesis generation and hypothesis testing?
• The importance of not fooling yourself. Why is “double dipping” dangerous? Instead of just reporting correlations, can you report cross-validation in an independent sample?
• The importance of understanding the limitations of massively univariate analyses. The general linear model (GLM) is an excellent machine for testing your a priori temporal hypotheses, but it cannot discover structure (in the data) that you had not anticipated.
• The importance of understanding the fundamental limitations of correlational neuroimaging. For example, in a clinical population with a specific pathology, differences in BOLD fMRI outcomes (compared with healthy controls) could be due to: a. Primary effects of disease; b. Secondary effects, due to living with disease; c. Compensation for, or adaptation to, disease; d. Therapy (e.g., pharmacological) for disease. In general, fMRI cannot distinguish between these possibilities.
• The importance of understanding the limitations of data pre-processing. For example, group analyses require spatial normalization (the transformation of subject-level fMRI data into labeled neuro-anatomical space, by warping individual brains onto an atlas or template), which is imperfect, especially in cases of manifest pathology such as brain tumors.
• The importance of understanding limitations of data acquisition strategies. In the case of task fMRI, can BOLD fMRI distinguish between excitatory and inhibitory neurotransmitters? In the case of resting fMRI, can we be confident that low-frequency signals are real, rather than temporally-aliased artifacts resulting from undersampled physiological processes such as respiration and cardiac pulsations?

Why did the course organizers choose to devote a talk to these “issues”? Perhaps because, as Feynman wrote, "...it is our responsibility as scientists, knowing the great progress which comes from a satisfactory philosophy of ignorance, the great progress which is the fruit of freedom of thought, to proclaim the value of this freedom; to teach how doubt is not to be feared but welcomed and discussed; and to demand this freedom as our duty to all coming generations."

Acknowledgements

References