2026-08-15
Quantitative fMRI in 40 healthy subjects shows about 40% of voxels with significant BOLD changes have oxygen metabolism moving in the opposite direction, concentrated in the default mode network; these voxels regulate demand via oxygen extraction rather than blood flow, challenging the canonical reading of BOLD.
fMRI does not measure neural activity; it measures blood oxygen. Neurons consume oxygen, vessels dilate in response, deoxygenated hemoglobin falls, and the BOLD signal rises. The whole interpretation rests on a canonical model: blood flow always outpaces metabolism (the ratio, n-ratio, sits around 2-4), so signal up means activity up and signal down means activity down.
PET validated that equation in sensory cortex, but nobody had tested it across the entire cortex. Counterexamples kept accumulating: animal studies found blood-flow changes with minimal BOLD response, and reports of negative BOLD with unchanged or increased metabolism. For a tool that thousands of papers a year depend on, this is a foundation-level question.
The Technical University of Munich team's design: measure BOLD and quantitative fMRI in the same session, quantify blood flow (CBF), blood volume (CBV), and oxygen extraction fraction (OEF), compute voxel-wise oxygen metabolism (CMRO2) via Fick's principle, and check whether the direction of the BOLD signal agrees with the direction of metabolism.
Forty healthy subjects (47 enrolled, 7 excluded for data quality) each completed four conditions in one session: mental arithmetic (CALC), autobiographical memory recall (MEM), a low-load baseline (CTRL), and rest. Arithmetic activates task-positive networks and suppresses the default mode network (DMN); memory does the reverse, so the same regions contribute both positive and negative BOLD responses.
The quantitative part combines multiparametric qBOLD with pseudocontinuous arterial spin labeling: T2 and T2 maps yield the deoxyhemoglobin-linked relaxation rate R2', DSC contrast agent gives CBV, pCASL gives CBF, and Fick's principle (arterial oxygen content × OEF × CBF) yields voxel-wise CMRO2. The MRI protocol is noninvasive and replaces the 15O-PET gold standard, which requires an on-site cyclotron and three short-half-life radiotracers.
Analysis uses partial least squares for group statistics, then compares the data voxel by voxel against the classic Davis hemodynamic model, which predicts that whenever CBF rises less than CMRO2 (n-ratio < 1), the same metabolic increase produces a negative BOLD signal. The paper measures how much of the cortex actually lives in that regime.
| Measurement | Result |
| Positive-BOLD voxels with opposite metabolism | 31% (CALC vs CTRL) |
| Negative-BOLD voxels with opposite metabolism | 66% (also 66% within the DMN) |
| Replication (harmonized voxel grids, N=10) | 40% discordant, both signs |
| Pooled over all significant BOLD voxels | about 40% discordant |
Concordant voxels show the canonical response (n-ratio 2.0/1.6) and meet oxygen demand through CBF changes (87% of explained variance); discordant voxels have insufficient CBF responses and rely on OEF changes (58%), with lower baseline OEF and metabolism, indicating a larger oxygen buffer. Baseline OEF explains over 68% of baseline CMRO2 variance.
The confound checks are thorough. The same voxels show a canonical response for positive BOLD under MEM but a discordant one for negative BOLD under CALC, ruling out voxel-specific artifacts. A 10-subject replication with harmonized voxel matrices reproduces the result. Excluding voxels with high vascular contribution still leaves 11-29% (positive) and 68-78% (negative) discordant.
For neuroimaging this is a quantitative audit of the BOLD interpretation framework. In roughly 40% of significant voxels, "signal decrease means activity decrease" is exactly backwards, and negative BOLD is the least trustworthy of all (two-thirds discordant), concentrated in the DMN, the network with the most reported deactivations. Task-versus-baseline subtraction logic distorts systematically in these regions.
For AI-adjacent readers the implication is direct: fMRI is the standard format for neural datasets feeding brain decoding, brain-computer interfaces, and models of visual cortex computation. Treating BOLD as a proxy for neural activity means training on directionally flipped labels in the discordant voxels.
The authors' alternative is quantitative fMRI: voxel-wise CMRO2, OEF, and CBF, more reliable for both absolute and relative changes. The acquisition burden is real but far below 15O-PET.
The authors list their own: mqBOLD absolute quantification of CBF, OEF, and CMRO2 carries systematic biases (affecting cross-subject comparison, not task effects), and DSC-based CBV reflects total rather than venous blood volume. The main study has 40 subjects (30 for MEM), all right-handed healthy adults; whether the findings hold in patient groups, whose hemodynamics are often altered, is untested.
My additional reservations: discordance is defined within the Davis model framework, whose parameters (α=0.38, M=11.2%) are estimated from the same data, a partial circularity, though the Fick-principle and Davis-model routes give similar proportions. CMRO2 is a product of four measured quantities, so error propagation is nontrivial and voxel-level SNR is managed by median aggregation. Only two cognitive tasks are studied; the whole-cortex claim extrapolates from them. And between CMRO2 and neural activity there remains a gap no MRI method has closed: that metabolism direction equals activity direction is still an assumption.