The future of neuroimaging

Episode Summary

Executive Summary: Stanford neuroscientist Russ Poldrack explains how functional MRI maps brain activity indirectly through blood flow, how his lab uses it to study cognitive control and psychiatric disorders, and why reproducibility requires both larger datasets and more transparent, flexible analysis practices. The conversation also covers self-experimentation, brain network variability across people, open science infrastructure, and privacy/ethics in data sharing.

Main Topics: How functional MRI works (Priority: 5/5): Poldrack explains fMRI as an indirect measure of neural activity via changes in oxygenated and deoxygenated hemoglobin, enabling researchers to infer which brain regions are active during tasks or rest. Brain networks, resting-state connectivity, and individual variability (Priority: 5/5): The discussion covers large-scale brain networks, resting-state fMRI, and the discovery that while everyone has the same broad networks, the fine-grained organization and functional variants differ across individuals. Studying cognitive control and psychiatric disorders (Priority: 5/5): Poldrack describes lab work on stopping actions, switching tasks, and overriding habits to understand circuitry underlying control deficits relevant to mental health and neurological disorders. Self-experimentation and deep phenotyping (Priority: 4/5): He describes scanning himself repeatedly to collect enough data to map his own brain in detail, showing that dense within-person sampling can reveal stable parcels and unique network features. Reproducibility and analytic variability in neuroimaging (Priority: 5/5): The field’s reproducibility problems are attributed to small samples and many plausible analysis choices; Poldrack advocates larger datasets and multiverse analyses rather than one mandated pipeline. Open science, data sharing, and infrastructure (Priority: 4/5): The conversation highlights OpenNeuro, the Brain Imaging Data Structure (BIDS), and the norms needed to make brain data shareable, usable, and standardized across labs. Ethics, privacy, and limits of mind reading (Priority: 4/5): Poldrack addresses re-identification risks, consent, and the limits of current neuroimaging, emphasizing that fMRI cannot read a person's full train of thought or reliably detect lies.

Key Arguments: fMRI is powerful because it noninvasively tracks activity indirectly through blood-oxygen signals, allowing whole-brain study at millimeter scale. Resting-state fMRI revealed stable large-scale networks that coordinate brain function even when a person is not performing a task. Fine-grained brain organization varies across individuals, so averaging across people can hide meaningful functional variants. Many psychiatric and neurological disorders may involve impaired cognitive control, making stop-signal and task-switching paradigms scientifically and clinically important. Deep within-person sampling can produce highly reliable brain maps and reveal features that single-session studies miss. Reproducibility in neuroimaging improved through larger samples, richer within-person data, and better transparency about analysis choices. Analytic flexibility is a major source of irreproducibility; researchers should test results across many plausible pipelines instead of assuming one analysis is best. Open data standards like BIDS and repositories like OpenNeuro are essential for scalable sharing and reuse of neuroimaging data. Current imaging can provide information about what types of things someone is thinking about, but not detailed inner speech or mind reading. The most useful future application may be predicting which treatment will work for a given depression or OCD patient, not simply diagnosing disease from a scan.

Data Points: MRI field strength: 3 Tesla - Standard high-end hospital scanners used for typical fMRI measurements Spatial resolution at 3T: About 2 millimeters - Typical functional MRI resolution for mapping activity changes Higher-field scanner: 7 Tesla - New Stanford system expected to improve resolution toward about 1 millimeter Resting-state scan duration on Poldrack: 10 minutes - He repeatedly scanned himself while simply resting with eyes closed Self-scan frequency: Twice a week - Poldrack’s self-experimentation schedule during his study Self-scan day/time: Tuesdays and Thursdays at 7:30 a.m. - Regular schedule for his repeated MRI sessions OpenNeuro dataset size: Almost 90,000 subjects - Repository containing openly shared neuroimaging data Parcelation result from self-study: About 600 parcels - Fine-grained cortical patches identified from dense personal data Research group count in reproducibility study: 70 different groups - Teams independently analyzed the same dataset Alternative analysis pipelines: 240 plausible analyses - Number of statistical approaches assembled from the literature for one task Depression treatment first-try success rate: About one-third - Poldrack notes many depression treatments work only around 33% of the time initially Brain imaging study size example: About 10,000 kids - Adolescent Brain Cognitive Development Study imaging repeated over time Per-kid per-session imaging time: About one hour - Data collected in the ABCD study per session

Pivotal Quotes: "I don't really believe you can like read the contents off of a person's mind right now with the methods that we have." — Russ Poldrack: He explains the limits of current neuroimaging and why lie detection/mind reading claims are overstated "We need to look at the data across a range of plausible analyses and say which results are consistent, which aren't." — Russ Poldrack: His proposed response to analytic variability and reproducibility problems in neuroimaging "The first thing we learned is that when you have enough data, you can very reliably characterize not just these large-scale networks..." — Russ Poldrack: Describing what dense self-scanning revealed about reliable brain parcelation

Implications: For listeners and researchers, the episode shows that brain imaging is becoming more precise, more individualized, and more open, but also that careful standards, large datasets, and privacy protections are essential before clinical use can expand.