Precision neuroimaging

Testing a model voxel by voxel takes many hours of high-quality data from one brain. Other questions need a fast measurement at one location, or timing that fMRI alone cannot give. We design experiments, acquisition methods and analyses for these measurements, and prepare the resulting data for open release.

Approach

With colleagues, we have described the principles of what we call intensive fMRI: extensive sampling of cognitive phenomena in a few individuals, to support computational modeling of single voxels. It depends on experiments designed for a rich hypothesis space and on careful optimization of data quality (Kupers et al., 2024).

Our intensive dataset, Arrow of Time, has about 21 hours of 7T fMRI per participant (1.7 mm isotropic, repetition time 1.6 s), recorded while participants watched 2,160 brief naturalistic videos, plus retinotopy and motion-energy localizers.

Line-scanning fMRI gives up volume coverage to sample a single line through cortex at very high spatial resolution and sampling rate. MEG adds millisecond timing to fMRI’s spatial detail.

What we have found

Line-scanning in humans. Our 7T implementation sampled every 200 ms, and its t-statistics corresponded with those of gradient-echo EPI at the same temporal resolution and voxel volume (R = 0.6 ± 0.2) (Raimondo et al., 2021). Multi-echo readouts and denoising raised temporal signal-to-noise ratio from 4.0 to 36.2, and image-based navigators added motion correction (Raimondo et al., 2022). A spin-echo version gave a sharper line profile, at a cost in functional sensitivity (Raimondo et al., 2023).

Targeting a chosen location. We select a target on each participant’s cortical surface from its anatomy and pRF, place the line there in a second session, and tailor the stimuli to it. For most participants, pRF estimates from the line resembled the target’s; target and best-matching V1 location were on average about 5.5 mm apart (Heij et al., 2023).

Timing from fMRI and MEG. A forward model that predicts MEG responses from each individual’s fMRI visual field maps revealed a processing hierarchy across them: V1 captured most of the variance in the MEG sensors, and did so earlier than extrastriate areas (Eickhoff et al., 2026).

Cortical microstructure in depression. Quantitative 7T MRI of T1 (a proxy for intracortical myelin) and T2* (a proxy for iron) in 48 patients with major depressive disorder and 10 controls linked diagnosis and symptom severity to lower myelination in lateral orbitofrontal cortex, not to iron (Heij et al., 2024).

Body maps from large open datasets. In 7T data from 174 Human Connectome Project participants, a model that maps body-part and visual field tuning at once produced detailed body-part tuning maps from about an hour of rest per participant; during video watching, it revealed an array of somatotopic body maps across the dorsolateral visual system (Hedger et al., 2025). See also Topographic maps.

Open questions

How well do models fit to many hours of data in one brain generalize to new stimuli and to other people? Which questions need a dense, multipurpose dataset, and which a fast measurement aimed at one location? We are preparing Arrow of Time as an open resource; its first release is planned around the turn of the year 2026/27 (get notified).