Visual task · Study & Methods

Where the recordings come from

The recordings belong to Studyforrest, an open collection of brain imaging and behavioral data. This analysis uses its higher visual area localizer: 3T fMRI acquired during image-category blocks, with 15 participants and four runs each. This is a distinct acquisition within Studyforrest; the movie-viewing recordings are not the inputs to this analysis.

The source inputs are the publisher's motion-aligned images and accompanying stimulus-event records. Anatomical mapping to MNI space, cortical ROI extraction and the temporal preparation described below were performed locally.

• Studyforrest project — the broader study and its aims.
• Data overview — the visual localizer and other available acquisitions.
• Sengupta and colleagues, 2016: visual-localizer data descriptor — acquisition, stimulus design and validation.
• Aligned-image repository and documentation — the distributed aligned derivatives.

Study design and scientific questions

The Studyforrest visual-category localizer contains 15 participants with four approximately five-minute runs each. Each run presents two blocks from each of six categories: faces, bodies, houses, objects, scenes and scrambled images. Participants detect image repetitions, and brief fixation periods separate the blocks.

The primary comparison pools visual stimulation against inter-block fixation. Category differences are a secondary question. A rank-7 model is learned across all 60 continuous runs without task labels; stimulus timing is used afterward to summarize its expressions.

Visual systems include retinotopic maps, category-selective responses and distributed interactions. The localizer design provides experimental contrasts; broad atlas labels describe a different scale of organization.

Signal preparation and timing

Publisher motion-aligned scans are mapped using BOLD-to-T1 rigid and T1-to-MNI affine transforms, then averaged within Schaefer 100 cortical parcels. Motion regression, detrending and a 0.008 Hz high-pass precede per-run, per-ROI min–max scaling. Task responses are retained.

Measured image events define the stimulus-block envelopes. BOLD responses lag and extend beyond the stimulus boundaries. Comparing 0-, 4- and 6-second offsets checks how conclusions depend on assigning samples to conditions; it does not estimate neural response onset. Runs stay continuous, and unrelated blocks are not joined to train separate models.

How regional signals become network dynamics

Each observation is a vector of signals from 100 cortical ROIs. Short windows summarize the recent past and subsequent observations. Spectral decomposition compresses their relationships into a small predictive space, and the model maps that representation back to patterns of regional relationships.

One basis is learned across the study, with independent recordings kept separate at their boundaries. The spatial network matrices are shared; their expression changes through time and can differ across participants or runs. A common basis makes these comparisons possible, while potentially reducing effects outside the retained space.

The modeled regional second moment is M(t) = W₀ + Σₖ xₖ(t)Wₖ. Each Wₖ describes a fixed spatial pattern and xₖ(t) its evolving coefficient. Multiple networks can contribute simultaneously. The fitted operator describes predictive relationships between coordinates; its signed entries are not transition probabilities.

Reading network expression and opposing poles

A network can contain two sets of ROI-pair weights with opposite signs. Their spatial footprints may emphasize different functional systems. One coefficient scales both poles together: a positive coefficient preserves the stored signs, and a negative coefficient reverses them. The opposing anatomy can be functionally informative, but its signs alone do not identify excitation, inhibition or cognitive states.

Average network coefficient describes the prevailing orientation of the pattern during a period. Contribution magnitude measures its size regardless of orientation. Large contributions in opposite directions can cancel in the signed average, so the two measures answer different questions.

Network 1 acts as a mean/bias-like pattern and is set aside in routine figures while retained in the model. It is distinct from the explicit constant coordinate. Network numbers and signs belong to one fitted study and rank; correspondence across separate fits requires an explicit comparison.

Connecting learned networks with functional anatomy

Schaefer parcels carry assignments to seven broad cortical systems: visual, somatomotor, dorsal attention, salience/ventral attention, limbic, frontoparietal control and default mode. The finer 17-system reference is available for exploration.

For each learned network, positive and negative edge magnitudes are summarized separately by off-diagonal row sums at each ROI. Regional Pearson similarity compares that spatial footprint with each reference system. The weight-share bar shows the proportion of the pole’s regional weight assigned to that system. A high resemblance and a large weight share are complementary evidence of localization, rather than interchangeable scores.

All edges contribute to reference matching. Reducing visible edges in the atlas changes the display only. A match is a descriptive anatomical association, not a validated functional label or a significance test.

Using the main views

• Functional systems: select a system or network to connect the reference anatomy with the positive and negative poles in the relationship graph.
• Network matches: click a score to inspect its network, pole and reference in the atlas. Click a system header to sort; select a ranked ROI to inspect its connections.
• Temporal comparisons: inspect individual recordings behind the group summaries. A time-varying coefficient changes the contribution of a fixed pattern; it does not change that pattern’s anatomy.

Quality checks and interpretation limits

Motion and cortical-coverage flags are retained for inspection. Full-cohort measurements and alignment examples appear in Preparation & QC. Detailed visual alignment review remains a limitation, and a processing flag is not evidence of a participant’s functional abnormality.

Participant-balanced condition summaries prevent recordings with more samples from dominating the group. Directional agreement describes how often participants shift in the same direction; it is not a significance test. Inter-block fixation retains stimulus history and is not equivalent to an independent resting-state scan.

Across-runs comparisons ask whether expression trajectories repeat within a participant. Shared spatial anatomy is imposed by the common basis; repeatable temporal expression is a separate result.

Scope of the evidence

The models and anatomical associations are exploratory and in-sample. Held-out prediction, spatial-null comparisons and repeated-fit stability would be needed to assess generalization. The two studies have separate acquisition and processing histories, so similar system labels do not make their fitted coefficients directly comparable.