Technical Validation

[Demonstrate the quality and reliability of the data through quantitative quality metrics, reproducibility analyses, and comparisons against established benchmarks.]

fMRI data quality¶

Functional image quality was assessed with MRIQC Esteban et al., 2017, computed per run on the raw BOLD data. We report head-motion and temporal signal-to-noise ratio (tSNR) image-quality metrics (IQMs) across the 4,537 functional runs with released MRIQC derivatives, spanning 17 datasets, all six participants, and 880 unique dataset × subject × session combinations. Four datasets (anat, emotion-videos, langlocalizer, mario) have no BOLD MRIQC derivatives available yet and are excluded from this analysis; the numbers below therefore describe the datasets with released MRIQC derivatives, not the full collection. No pass/fail threshold is applied anywhere in this analysis and no run is excluded — the distributions below describe the released data as-is.

Head motion, summarized as mean framewise displacement (FD) per run Power et al., 2012, was low overall: median 0.145 mm (mean 0.147, SD 0.062, range 0.046–0.480 mm across runs). No run exceeded a mean FD of 0.5 mm, and 787 runs (17.3%) exceeded 0.2 mm. At the volume level (3,582 runs with MRIQC framewise-displacement timeseries available), a median of 21.7% of volumes per run exceeded 0.2 mm and a median of 0.5% exceeded 0.5 mm (means 23.2% and 2.1%). Motion varied about 2.5-fold across participants — sub-03 and sub-04 showed the lowest motion (median FD 0.074 and 0.081 mm), sub-02 and sub-06 the highest (0.178 and 0.187 mm) — while sub-01 and sub-05 fell in between (0.122 and 0.132 mm). Motion also varied systematically with task demands: passive paradigms such as floc, movie10, and hcptrt had the lowest median FD (0.111–0.118 mm), while the video-game datasets mario3, shinobi, and mariostars had the highest (0.177–0.206 mm), consistent with the additional head movement associated with active gameplay.

Temporal SNR followed the inverse pattern: median 29.8 across runs (mean 29.6, SD 5.3, range 11.5–40.1), with a strong negative correlation between per-run mean FD and tSNR (Pearson r = −0.76, Spearman ρ = −0.77) — runs and participants with more motion have correspondingly lower tSNR. Per-subject median tSNR ranged from 26.5 (sub-06) to 35.2 (sub-03), and per-dataset median tSNR from 22.3 (shinobi) to 33.0 (hcptrt).

Regional tSNR was further characterized by averaging per-run tSNR maps within a combined Schaefer-1000/7-network cortical Schaefer et al., 2018Yeo et al., 2011, Tian-S3 subcortical Tian et al., 2020, and Nettekoven cerebellar Nettekoven et al., 2024 atlas, collapsed to 11 region groups. This analysis covers 936 runs from the floc, retinotopy, and things datasets only, as the per-run tSNR maps for the remaining datasets sit on credentialed data remotes that were not retrievable for this pass; the regional pattern below should be read as illustrative of the acquisition’s spatial signal-quality profile rather than as a comprehensive summary. Median tSNR was lowest in the Limbic network (18.7 — orbitofrontal and ventral-temporal cortex) and in subcortical structures (thalamus 27.4, caudate 28.6, putamen 29.0) and cerebellum (30.7), and highest in dorsal cortical networks (Dorsal Attention 46.5, Control 43.6, Somatomotor 43.4, Salience/Ventral Attention 42.6, Default 40.7, Visual 37.7). This ordering reflects the expected susceptibility-dropout pattern of gradient-echo EPI near air-tissue interfaces, a property of the acquisition geometry rather than of CNeuroMod specifically; users of ventral-temporal, orbitofrontal, or deep subcortical signal should budget for reduced tSNR in these regions.

Taken together, these metrics indicate low and stable head motion and adequate temporal signal quality across the released functional runs, with expected, interpretable variation across participants, tasks, and brain regions. The main limitations of this first-pass analysis are its restriction to datasets with released MRIQC and tSNR derivatives, and the regional breakdown’s reliance on a three-dataset subset; both will be extended as further derivatives become publicly available.

sMRI data quality¶

Structural image quality was assessed separately across the anatomical acquisitions of the same six participants and is reported in a companion publication [MISSING REF: companion CNeuroMod structural data quality paper — citation to be supplied].

Longitudinal stability and state-dependence of fMRI measures¶

Session-level within-network functional connectomes were computed from the parcellated BOLD timeseries of cneuromod.all, using the cneuromod2026 parcellation — 1,134 parcels combining a Schaefer cortical parcellation Schaefer et al., 2018, grouped into the 7 Yeo cortical networks Yeo et al., 2011, a Tian subcortical parcellation Tian et al., 2020, and a Nettekoven cerebellar parcellation Nettekoven et al., 2024. Runs were z-scored individually and concatenated within a session, and connectomes were estimated independently within each network (Pearson correlation of parcel timeseries, Fisher-z transformed). Session-pair similarity is the Pearson correlation between two sessions’ Fisher-z edge vectors within a network, and bins of session pairs are summarized by their median similarity. Connectomes were computed for all 829 available sessions across 10 datasets; the analyses below use the 559 sessions from 7 datasets (friends, harrypotter, hcptrt, mario, movie10, petit-prince, shinobi) carrying at least 30 minutes of usable data, covering all six participants. This 30-minute gate removes floc, retinotopy, and things entirely.

Within-subject connectome similarity in friends — the most task-homogeneous dataset — declines gently and monotonically with the number of seasons separating two sessions, the only time axis available since sessions carry no acquisition dates (Figure 2, panel A). The decline over a five-season lag ranges from 0.019 (cerebellum) to 0.043 (Limbic network) — e.g., 0.956 to 0.935 in the Visual network — and every network remains far above the between-subject floor (0.564–0.572). Drift over years of scanning is therefore small relative to the gap between individuals.

Connectome similarity is also sensitive to cognitive context. Across four session-pair types, the ordering within-subject/within-dataset > within-subject/between-dataset > between-subject/within-dataset > between-subject/between-dataset holds in all 9 networks (e.g., Visual 0.95/0.80/0.69/0.63; Default 0.94/0.72/0.56/0.45; Figure 2, panel B). This contrast is not confounded by acquisition duration: similarity increases with session duration, so the four bins were matched by construction, with median pair minimum duration ranging only 2,669–2,784 s (within 4%) across bins. The between-dataset drop in similarity therefore reflects a genuine effect of cognitive state rather than a duration artifact or measurement noise.

Similarity also varies by network quality. The Limbic network has both the lowest median tSNR (18.6) and the lowest within-subject similarity (0.859), and the cerebellum and subcortex sit below the cortical networks on both measures (Figure 2, panel C). This comparison is descriptive only: the per-network tSNR values are available only for the floc, retinotopy, and things datasets (182 sessions) — precisely the three datasets removed by the 30-minute gate — while similarity is computed over the disjoint set of 559 gated sessions. With nine network-level points and no shared sessions between the two axes, this panel establishes an ordering, not a quantitative tSNR–similarity relationship.

As a robustness check on the state-dependence result, restricting the “different task” comparison to a swap within a single naturalistic stimulus domain — movies (friends and movie10, 333 sessions), video games (mario, mario3, mariostars, shinobi, 138 sessions), and stories (harrypotter, petit-prince, 19 sessions) — still yields within-subject/within-task similarity exceeding within-subject/between-task similarity in all 9 networks for all three domains (Figure 2, panels D–F), with median gaps of 0.022 (movies), 0.047 (video games), and 0.077 (stories). The effect is smallest for movies, where “different task” means a different film rather than a different kind of activity; the stories domain, resting on only 19 sessions, is suggestive rather than conclusive. Stratifying session pairs by head motion or by tSNR does not change any of these orderings (not shown).

Preprocessing Pipeline Validation¶

[Describe any validation steps applied to preprocessed derivatives.]