The future of neuroimaging

Episode Summary

Executive Summary: The episode explores how functional MRI reveals brain networks, individual differences, and disease-related patterns, while also addressing major reproducibility challenges in neuroimaging. Russ Poldrack explains resting-state and task-based fMRI, self-experimentation, open data, and multiverse analysis as ways to make findings more reliable and clinically useful, especially for psychiatric diagnosis and treatment prediction.

Main Topics: How fMRI works: Poldrack explains that fMRI indirectly measures brain activity through blood-oxygen-level changes, using magnetic properties of oxygenated and deoxygenated hemoglobin to infer where activity is occurring. Task-based vs. resting-state imaging: The discussion contrasts classic task fMRI with resting-state fMRI, which reveals correlated large-scale brain networks even when a person is doing nothing in the scanner. Brain networks, parcels, and individual variation: Poldrack describes how the brain is organized into large-scale networks and finer parcels, and how individual brains can differ in the exact placement of functional boundaries and connections. Clinical applications in mental health: The episode highlights research linking altered salience-network size with depression and OCD symptoms, with an emphasis on eventual treatment prediction and targeting rather than simple diagnosis. Reproducibility and analytic variability: Poldrack argues that inconsistent methods across labs can produce conflicting results, motivating larger datasets, deeper sampling per participant, and multiverse analysis. Open science and data sharing: The conversation covers open repositories like OpenNeuro, standardized file organization via BIDS, and the need to balance transparency with privacy protections.

Key Arguments: Functional MRI is a powerful indirect tool for mapping brain function because blood-flow changes reliably track neural activity at a coarse spatial scale. Resting-state fMRI can identify stable, large-scale functional networks without requiring task performance. The human brain shares common network architecture across people, but fine-grained organization varies substantially by individual. Some psychiatric conditions may be better understood as changes in brain network organization or traits than as simple symptom-state markers. Reproducibility problems in neuroimaging often stem from small samples and highly variable analytic choices rather than from one single flawed method. Large datasets, repeated scans on the same person, and multiverse analysis improve robustness by revealing which findings survive many plausible analytic paths. Open data and standardized formats are essential for making neuroimaging more transparent, reusable, and reproducible. Current neuroimaging cannot read a person's exact thoughts; it can infer broad categories of mental activity but not inner speech or full trains of thought.

Data Points: MRI field strength: 3 Tesla - Standard high-end hospital MRI scanners discussed as the basis for typical fMRI measurements. Spatial resolution at 3T: About 2 millimeters - Typical resolution Poldrack said is achievable in standard functional MRI. Spatial resolution at 7T: About 1 millimeter - Higher-field 7 Tesla scanners can improve resolution toward about a millimeter. Resting data collection per session: 10 minutes - Poldrack’s self-study included 10 minutes of resting-state data each scan. Self-scanning frequency: Twice a week - He scanned himself regularly on Tuesday and Thursday mornings. Self-scanning start time: 7:30 a.m. - The time he regularly entered the scanner for his self-experimentation. Self-study duration per visit: 30 minutes - Each self-scan session included about half an hour in the MRI scanner. Functional parcels found in self-study: About 600 - Parcelation of Poldrack’s brain identified roughly 600 small functional patches. Number of networks in common brain organization: Approximately 13-20 - Poldrack described resting-state large-scale networks as numbering around the teens to low twenties depending on the definition. Analytic groups in reproducibility study: 70 - A dataset was shared with 70 research groups who independently analyzed it. Plausible analyses in a task-study audit: 240 - Poldrack described a paper where 240 plausible analytic pipelines were identified. Depression treatment effectiveness: About one-third - He said the first treatment for depression works about a third of the time. OpenNeuro dataset size: Almost 90,000 subjects - Poldrack described the OpenNeuro repository as containing nearly 90,000 participants' worth of data. ABCD study sample: About 10,000 kids - Example of a large-scale repeated-imaging study used to improve reproducibility. ABCD per-session scan time: About 1 hour per child per session - Used to illustrate the tradeoff between depth and breadth in imaging studies.

Pivotal Quotes: "we 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: On the limits of what neuroimaging can infer about thought and mental content "what we need to do to make the research more reproducible" — Russ Poldrack: On the field’s response to inconsistent findings in brain imaging "you can do the analysis any way you want, but then you need to look at how sensitive the results are to changing various features of that analysis" — Russ Poldrack: On his approach to multiverse analysis and analytic transparency

Implications: For listeners and the field, the episode shows that fMRI is increasingly useful for mapping brain organization and may someday help guide psychiatric treatment, but only if researchers embrace larger samples, standardized sharing, and rigorous analysis to avoid false findings.