Studyforrest: visual blocks, brain signals, and a route to network dynamics¶

Can learned brain networks distinguish what a person is viewing? This dataset provides a controlled counterpart to resting-state analysis. Participants view blocks of faces, bodies, houses, small objects, outdoor scenes, or scrambled images and detect image repetitions. Repeated blocks let us ask whether network expression changes reliably with stimulus category, rather than merely fluctuating over time.

Sources: studyforrest data overview , localizer data descriptor , aligned scan documentation , published visual ROIs .

What does each run contain?¶

Each approximately five-minute run has two blocks from each of six categories. A block contains 16 image presentations over roughly 16 seconds, with fixation between blocks. Category order varies between runs. The colors describe the stimulus schedule, not inferred brain states.

We retain the publisher's individual image timings. A second public release summarizes events as nominal 16-second blocks; that representation is not silently substituted here. The block envelopes below span the measured first image onset through the final image offset. Small timing differences are therefore expected.

Are stimulus and image times aligned?¶

The saved BOLD scans contain 156 volumes at a verified TR of approximately two seconds. Every volume is retained, and each run keeps its own time origin. Image onsets are interpreted relative to the first saved volume, following the event-file convention; no extra discarded-volume or hemodynamic shift is invented. This checks file consistency, not an independently measured scanner-display clock.

The BOLD response follows neural activity with a delay and temporal spread. The second panel shows an illustrative SPM hemodynamic response model driven by the actual image events. It is a design visualization, not a fitted activation result.

Is movement likely to dominate these recordings?¶

The publisher corrected head movement with MCFLIRT and aligned scans to a participant-specific BOLD template. Here we estimate framewise displacement from its six motion parameters, using a 50 mm radius to convert rotations to millimeters. Frames above 0.5 mm are flagged and retained. The first displacement is undefined because there is no preceding volume.

Low displacement is useful evidence about movement, but it does not establish complete signal quality or correct anatomical alignment. No claim is made that these derivatives include all fMRIPrep steps.

Where are the inspected signals sampled?¶

The native-space example uses published visual masks: face-, place-, body-, and object-related areas, plus early visual cortex. Multiple clusters with the same label are united; different ROI masks remain independent and may overlap. Missing areas are not invented. These nine masks are not a whole-cortex atlas.

The masks were defined using these same four localizer runs. Their apparent category responses therefore cannot be treated as an independent validation of our method. They provide a useful acquisition and timing check while a separate whole-cortex extraction is prepared.

How do the signals vary through the stimulus blocks?¶

Choose a visual region to inspect its four runs. The lines are detrended, motion-regressed, high-pass-filtered ROI means, expressed as within-run standard scores for display only. Colored bands indicate the stimulus blocks; responses need not begin or end at their boundaries because the BOLD response is delayed.

The separately saved model-input series uses the selected 0–1 min/max scaling. Unscaled, uncleaned ROI means and cleaned amplitudes are preserved alongside it, so changing normalization later does not require downloading the scans again. No low-pass filter or stimulus regression is applied; evoked responses remain.

Can these scans be normalized to MNI space?¶

Yes. The publisher supplies a rigid BOLD-to-T1 registration and a 12-parameter T1-to-MNI152 affine registration. We compose these mappings and interpolate each volume once onto a 3 mm MNI grid. The temporal axis is unchanged. Native scans and their ROI representation are retained as a separate version.

This is affine normalization, not nonlinear registration. It adjusts position, orientation, scale, and shear but does not match individual cortical folding. Before applying it, we check that the coordinate conversion reproduces the publisher's own checksum-verified resampled BOLD and anatomical images. The mirror and GitHub HEAD expose different versions, so the mirror's exact commit is pinned and all used files are checked against that snapshot. This is a numerical implementation check, not proof of precise cortical correspondence.

The figure compares the registered participant anatomy and BOLD coverage with an MNI template. Schaefer extraction and group inference should follow visual registration and parcel-coverage review. FSL coordinate conventions.

What is ready, and what remains a design choice?¶

The immediate result is a traceable stimulus-and-signal dataset, with native scans, event timings, motion parameters, explicit preprocessing, and versioned derivatives. It gives us a basis for a modest task-linked demonstration before cohort expansion.

Spatial normalization does not replace temporal nuisance processing. Slice timing, distortion correction, and a fuller confound policy remain to be reviewed before a larger study.

To reproduce acquisition/preparation: python scripts/prepare_studyforrest_localizer.py. To create the separate MNI derivative: python scripts/normalize_studyforrest_mni.py. Change the saved configuration to select subjects, runs, filtering, or normalization; new parameter combinations receive their own version directory.