40 percent of fMRI signals do not correspond to actual brain activity

Original Article ↗ Hacker News Discussion ↗

Scope and headline issues

  • Many commenters say the headline is misleading: it implies “MRI” broadly, but the finding concerns functional MRI (fMRI) and specifically the BOLD signal.
  • Structural MRI doesn’t measure brain activity at all; it’s imaging anatomy. Some note structural MRI is also statistically abused with tiny sample sizes.

What fMRI/BOLD measures

  • Explanations clarify that fMRI tracks changes in blood oxygenation (BOLD: blood‑oxygenation‑level dependent), not neuronal firing directly.
  • The standard assumption: more local neural activity → more metabolism → more blood flow/oxygenation → larger BOLD signal.

Dead salmon and statistical abuse

  • The famous “dead salmon” fMRI paper is repeatedly cited: it showed you can get “significant activations” in a dead fish if you don’t correct for multiple comparisons.
  • Participants stress that the lesson was “you must do proper statistics,” not “fMRI is nonsense.”

New paper’s claim and its interpretation

  • The reported result that ~40% of increased fMRI signal corresponds to decreased neuronal activity is seen as challenging the simple interpretation “more BOLD = more activation.”
  • Some argue this isn’t shocking to experts: the coupling between BOLD and neural activity has long been debated; non‑neuronal processes and inhibitory activity also drive metabolism.
  • Others point out the new study validates conventional BOLD using another model‑based MRI measure, which itself rests on assumptions and is not a perfect ground truth (PET would be closer but is costly and invasive).

Reliability, reproducibility, and pipelines

  • Several comments emphasize very poor test–retest reliability for many task‑based fMRI paradigms, implying many studies are underpowered and not reproducible.
  • Site/machine differences, numerous preprocessing choices, and low signal‑to‑noise are cited as major issues; tools like fMRIprep and statistical harmonization (e.g., COMBAT) try to mitigate this.
  • Some argue “almost all” cognitive fMRI is unreliable; others counter that, with large samples, good tasks, strict noise handling, and cross‑modal confirmation, robust findings exist (especially methods papers and basic sensory/motor work).

Clinical and pop‑science misuse

  • Commenters criticize commercial and media use of colorful brain images (fMRI, SPECT) as diagnostic “mind reading” or personality typing, calling it “non‑invasive phrenology” and “wallet biopsy.”
  • Influencer doctors selling expensive scans without strong evidence are cited as examples of pseudoscientific overreach that this kind of research helps push back against.

Machine learning and overinterpretation

  • Experiences from BMI/EEG/fMRI startups highlight how deep learning can find patterns in noise and artifact if not rigorously validated, yet such work is easily hyped as “AI can read your thoughts.”
  • Overall sentiment: fMRI remains a powerful but extremely indirect, noisy, and fragile tool whose results require cautious, statistically rigorous interpretation—especially when generalized to claims about “brain activation,” cognition, or diagnosis.