CNeuroMod-THINGS fMRI Dataset

• CNeuroMod-THINGS is a densely sampled fMRI dataset that captures neural responses to a broad array of 720 object concepts using a continuous recognition task.
• The dataset combines rich, standardized THINGS stimuli with extensive longitudinal imaging and behavioral metrics, enabling precise subject-specific neuro-AI modeling.
• High-quality annotations, multimodal physiological recordings, and robust preprocessing pipelines enhance cross-modal comparisons and reproducibility in visual neuroscience research.

Searching arXiv for the specified paper and closely related resources to ground the article in current literature. CNeuroMod-THINGS is a densely sampled, large-scale fMRI dataset for visual neuroscience designed to support data-hungry neuro-AI modelling by capturing neural representations for a wide set of semantic concepts using well-characterized stimuli (St-Laurent et al., 11 Jul 2025). It exploits synergies between the THINGS initiative and the Courtois Project on Neural Modelling (CNeuroMod): THINGS provides a common set of thoroughly annotated images broadly sampling natural and man-made objects, whereas CNeuroMod provides hundreds of hours of fMRI data from a core set of participants during controlled and naturalistic tasks, including visual tasks such as movie watching and videogame playing (St-Laurent et al., 11 Jul 2025). In CNeuroMod-THINGS, four deeply phenotyped CNeuroMod participants completed 33–36 sessions of a continuous recognition paradigm using approximately 4000 images from the THINGS stimulus set spanning 720 categories, thereby creating a resource intended to expand the capacity to model broad slices of the human visual experience (St-Laurent et al., 11 Jul 2025).

1. Dataset scope and conceptual positioning

CNeuroMod-THINGS is defined by the combination of dense longitudinal sampling, broad semantic coverage, and standardized stimulus annotation (St-Laurent et al., 11 Jul 2025). The dataset includes four deeply phenotyped participants, identified as “sub-01”, “sub-02”, “sub-03”, and “sub-06”, drawn from CNeuroMod. Each participant completed a continuous recognition paradigm across 33–36 fMRI sessions. All were healthy, right-handed adults aged 39–49, had normal or corrected vision, had no neurological history, and provided informed consent for open data sharing; sub-01, sub-02, and sub-03 were released under CC0 via CONP, whereas sub-06 was released under DTA (St-Laurent et al., 11 Jul 2025).

The central design objective is to bridge together two existing infrastructures. THINGS contributes a stimulus ecosystem centered on thoroughly annotated object images, and CNeuroMod contributes a dense within-subject imaging framework with repeated observation over many weeks (St-Laurent et al., 11 Jul 2025). This suggests that the dataset is particularly suited to subject-specific modelling regimes in which the limiting factor is not participant count but repeated sampling depth and trial-level stability. A plausible implication is that CNeuroMod-THINGS occupies a distinct niche relative to broader but more sparsely sampled datasets: it prioritizes repeated measures and individual cortical geometry over large-$N$ population inference.

The relation to the THINGS ecosystem is methodologically important. The THINGS stimulus database contains 1,854 categories and more than 26k images, and CNeuroMod-THINGS samples from that space to create a controlled subset with common annotations and potential interoperability with other THINGS-based neural datasets (St-Laurent et al., 11 Jul 2025). This interoperability is explicitly emphasized in the recommended use cases, which include linking the fMRI data to THINGS-data multimodal responses such as EEG and behavioral similarity judgments through the common stimulus set and annotations (St-Laurent et al., 11 Jul 2025). Related work using THINGS-derived EEG benchmarks illustrates this cross-modal potential, including large-scale EEG decoding on THINGS-EEG (Dixen et al., 20 Mar 2025) and EEG-informed image generation on ThingsEEG (Chen, 2024).

2. Experimental design and stimulus sampling

The experimental paradigm is a continuous recognition task deployed over many sessions (St-Laurent et al., 11 Jul 2025). The first session included 3 runs, and subsequent sessions included 6 runs each, totaling 213 runs for sub-01, sub-02, and sub-03, and 195 runs for sub-06. Each run lasted 283 s, acquired 190 volumes with $TR = 1.49$ s, and comprised 60 trials (St-Laurent et al., 11 Jul 2025).

Each trial presented a 900×900 px image subtending $10^\circ$ visual angle for 2.98 s, followed by a 1.49 s inter-stimulus interval time-locked to the scanner. A fixation marker subtending $2^\circ$ visual angle, described as a bull’s-eye plus crosshair, remained visible throughout (St-Laurent et al., 11 Jul 2025). The fixation constraint, together with eye-tracking acquisition, is intended to reduce ambiguity about retinal stimulation and attentional deployment.

The memory task was structured around three presentations of each image. A given image was first presented as “unseen” and then repeated twice as “seen”, with repetition patterns arranged such that images were repeated either first within the same session and then across sessions or vice versa (St-Laurent et al., 11 Jul 2025). Participants responded using four buttons to indicate “seen/unseen” with high or low confidence. Response-to-button mappings switched randomly on each trial in order to dissociate motor from memory signals (St-Laurent et al., 11 Jul 2025). This is a notable design decision because it makes motor preparation a less reliable surrogate for mnemonic state.

The sampled stimulus set covered approximately 4,320 unique images for three subjects and 3840 for sub-06, spanning 720 object concepts drawn from the full THINGS database (St-Laurent et al., 11 Jul 2025). The categories were chosen to represent visual and semantic diversity of American English–nameable objects while balancing man-made and natural entities. No category was presented more than once per session in order to minimize interference (St-Laurent et al., 11 Jul 2025).

The annotation schema is unusually rich. Images carry higher-order WordNet labels such as “animal”, “plant”, and “vehicle”, as well as THINGSplus concept ratings on 0–7 scales for attributes such as “size” and “naturalness” (St-Laurent et al., 11 Jul 2025). In addition, boolean flags were manually coded for each image, including face present, human versus non-human face, body parts, and scene versus lone object (St-Laurent et al., 11 Jul 2025). This annotation structure enables both categorical and continuous semantic analyses and supports representational models that combine visual and semantic feature spaces.

3. MRI acquisition and auxiliary measurements

All imaging was acquired on a Siemens PRISMA Fit 3 T scanner using a 64-channel head/neck coil (St-Laurent et al., 11 Jul 2025). Functional data used a multiband gradient EPI sequence in a Human Connectome Project–style configuration with slice acceleration $= 4$. The key parameters were $TR = 1.49$ s, $TE = 37$ ms, flip angle $= 52^\circ$, voxel size $= 2$ mm$^3$ isotropic, 60 slices, and a 96×96 matrix (St-Laurent et al., 11 Jul 2025).

Structural imaging was collected in periodic sessions. The T1-weighted MPRAGE sequence used $TR = 1.49$0 s, $TR = 1.49$1 ms, flip angle $TR = 1.49$2, 0.8 mm$TR = 1.49$3 resolution, and $TR = 1.49$4; the T2-weighted SPACE sequence used $TR = 1.49$5 s, $TR = 1.49$6 ms, 0.8 mm$TR = 1.49$7 resolution, and $TR = 1.49$8 (St-Laurent et al., 11 Jul 2025). Such anatomical resolution supports accurate cortical surface reconstruction, individual alignment, and the definition of fine-grained subject-specific ROIs.

Physiological monitoring extended beyond standard MRI acquisition. The dataset includes ECG sampled at 10000 Hz, plethysmography, EDA, respiration, and 250 Hz eye-tracking of the right eye using Pupil Core (St-Laurent et al., 11 Jul 2025). This multimodal physiological acquisition is relevant for both nuisance modelling and task compliance assessment. A plausible implication is that the data can support more aggressive denoising or state-dependent modelling than datasets limited to motion parameters alone.

The session counts were highly symmetric across most participants: sub-01, sub-02, and sub-03 each completed 36 sessions of the THINGS task, and sub-06 completed 33 sessions (St-Laurent et al., 11 Jul 2025). The resulting longitudinal density is central to the dataset’s intended use in single-subject computational modelling.

4. Preprocessing, derivatives, and data organization

All fMRI data adhere to BIDS 1.2.1 and are managed with DataLad submodules (St-Laurent et al., 11 Jul 2025). The organization of the release is explicit and modular, encompassing functional data, model-derived beta estimates, memory contrasts, localizer outputs, and anatomical surface products.

Functional preprocessing was performed with fMRIPrep v20.2.5 (St-Laurent et al., 11 Jul 2025). The pipeline includes motion correction by rigid-body realignment, slice-timing correction, susceptibility distortion correction, co-registration to T1w anatomy, and normalization to MNI space. Confound regressors are provided, including six motion parameters and CSF/WM signals. Outputs include preprocessed BOLD data in native T1w and MNI spaces, brain masks, and *_events.tsv files containing trial-wise timings and behavioural and eye-tracking metrics (St-Laurent et al., 11 Jul 2025).

Single-trial response estimation was performed with GLMsingle (St-Laurent et al., 11 Jul 2025). Timecourses were masked and z-scored per voxel, and the first two volumes were dropped for equilibrium. A customized 13-fold cross-validation scheme split runs so that no image’s repeats fell in the same fold. Internal GLMdenoise and ridge-regression regularization were then used to compute trial-specific beta maps, which were subsequently z-scored across all trials and voxels. The resulting derivatives include trial-wise beta maps, image-wise averaged beta maps, and voxelwise noise ceilings (St-Laurent et al., 11 Jul 2025).

Retinotopy and category localizers were also acquired and processed. The pRF task used three aperture shapes—ring, bar, and wedge—at $TR = 1.49$9 s, and was analyzed with analyzePRF to estimate eccentricity and receptive-field size. These estimates were combined with a Bayesian atlas via Neuropythy to define V1–V3, VO1/2, LO1/2, TO1/2, and V3a/b (St-Laurent et al., 11 Jul 2025). The functional localizer, fLoc, used blocks of faces, bodies, objects, characters, and places; analysis used a GLM in Nilearn with canonical HRF, drift regressors, AR1 noise, and smoothing with $10^\circ$0 mm. Category-selective ROIs including FFA, OFA, pSTS, EBA, PPA, OPA, and MPA were delineated via intersection with Kanwisher lab group parcels (St-Laurent et al., 11 Jul 2025).

The data structure is organized into named directories corresponding to these processing stages and task components.

Path	Contents
cneuromod-things/THINGS/fmriprep	raw & preprocessed BOLD, eye-tracking, *_events.tsv, stimuli, annotation TSVs
cneuromod-things/THINGS/glmsingle	beta maps, noise ceilings, quality metrics
cneuromod-things/THINGS/glm-memory	t-maps for memory contrasts
cneuromod-things/fLoc and cneuromod-things/retino	raw/preprocessed runs, ROI masks
cneuromod-things/anatomical	flat cortical map patches, patch application scripts, Pycortex surfaces

This structure is designed for reproducible data access and analysis. The recommended use cases specifically mention DataLad scripts, BIDS structure, Jupyter notebooks for figure reproduction and workflow adaptation, and Pycortex flat maps for visualization of model results on individual cortical surfaces (St-Laurent et al., 11 Jul 2025).

5. Behavioral performance and data quality

The reported behavioural and neuroimaging metrics are intended to showcase the quality of the dataset (St-Laurent et al., 11 Jul 2025). Behavioral performance was high. Response rates ranged from 91.1% for sub-01 to 99.8% for sub-03, and the median run-level response exceeded 98% (St-Laurent et al., 11 Jul 2025). Memory sensitivity, quantified as $10^\circ$1, was 1.744 for sub-01, 1.536 for sub-02, 1.623 for sub-03, and 1.898 for sub-06 (St-Laurent et al., 11 Jul 2025). Reaction times were faster for within-session hits than for between-session hits, consistently across subjects (St-Laurent et al., 11 Jul 2025). This suggests that the paradigm successfully induced memory conditions with measurable temporal differences rather than mere button-press compliance.

Signal quality was summarized using voxelwise noise ceilings computed following Allen et al. (2022) (St-Laurent et al., 11 Jul 2025). The signal-to-noise ratio was defined as

$10^\circ$2

with per-voxel noise variance estimated from across-repetition variance and signal variance estimated as $10^\circ$3 after standardization. The noise ceiling in percent was then defined as

$10^\circ$4

where $10^\circ$5 is the number of presentations per image (St-Laurent et al., 11 Jul 2025). Maximum observed ceilings were 56.5%, 66.4%, 73.9%, and 54.2% for sub-01 through sub-06, with the highest values in V1–V3 and category-selective patches such as FFA, PPA, and EBA (St-Laurent et al., 11 Jul 2025). These values indicate substantial explainable variance in visual and category-selective cortex.

Head motion was low. Framewise displacement had median values below 0.1 mm for most runs, mean FD below 0.15 mm in all subjects, and sub-03 below 0.1 mm mean FD (St-Laurent et al., 11 Jul 2025). Eye-tracking quality checks were passed by between 140/213 and 190/213 runs, and in most runs at least 75% of gaze samples fell within $10^\circ$6 of fixation (St-Laurent et al., 11 Jul 2025). These measures jointly support the interpretation that stimulus-locked visual responses were not severely compromised by overt movement or gaze instability.

The paper notes that full intraclass correlations were not explicitly reported, but proposes the repeated-measures ANOVA-based statistic

$10^\circ$7

where $10^\circ$8 and $10^\circ$9 are the mean squares across $2^\circ$0 sessions (St-Laurent et al., 11 Jul 2025). High noise ceilings and stable betas across weeks are described as supporting strong test–retest reliability (St-Laurent et al., 11 Jul 2025). Since explicit ICC values are not given, this should be interpreted as an inferential summary rather than a directly tabulated result.

6. Analytical affordances and recommended use cases

CNeuroMod-THINGS is explicitly positioned as a resource for encoding and decoding model pipelines (St-Laurent et al., 11 Jul 2025). The recommended workflow is to train voxelwise encoding models that map high-dimensional visual features—such as deep CNN activations or semantic embeddings derived from THINGSplus—to single-trial beta responses. Suggested estimation strategies include Bayesian ridge, regularized regression, and nonlinear methods such as kernel or neural models, taking advantage of the subject-specific data density for robust cross-validation (St-Laurent et al., 11 Jul 2025). The resource also supports decoding semantic categories or concept ratings from beta maps and testing generalization across within-session versus between-session repeats (St-Laurent et al., 11 Jul 2025).

The dataset’s value is amplified by its placement within two broader ecosystems. Through THINGS, it can be linked to multimodal responses including EEG and behavioral similarity judgments via the common stimulus set and shared annotations (St-Laurent et al., 11 Jul 2025). Through CNeuroMod, it can be integrated with movie watching, language, and resting-state tasks for multi-task modelling and cross-modal representational analyses (St-Laurent et al., 11 Jul 2025). This suggests an avenue for studying whether object representations measured under tightly controlled recognition paradigms transfer to more naturalistic perceptual and cognitive contexts.

The ROI infrastructure adds another layer of analytical specificity. Researchers can combine the THINGS task with retinotopy and fLoc ROIs to constrain models to known functional regions or to explore representations across the whole cortical surface (St-Laurent et al., 11 Jul 2025). Because the dataset also includes Pycortex surfaces and flat cortical map patches, it is naturally suited to analyses that require individual cortical topology rather than volumetric averaging (St-Laurent et al., 11 Jul 2025).

The dataset is also conducive to neuro-AI workflows in a practical sense. The release includes GLMsingle derivatives for direct use in machine-learning pipelines, DataLad scripts and BIDS organization for reproducible data handling, and Jupyter notebooks to reproduce figures and adapt workflows for custom analyses (St-Laurent et al., 11 Jul 2025). A plausible implication is that the resource lowers the engineering overhead typically associated with large single-subject fMRI datasets, particularly for groups whose primary expertise lies in computational modelling rather than MRI preprocessing.

7. Relation to broader THINGS-based multimodal research

CNeuroMod-THINGS belongs to a broader family of THINGS-based datasets that use a shared object stimulus ontology to study neural representation across modalities (St-Laurent et al., 11 Jul 2025). Within that ecosystem, THINGS-EEG has been used for deep-learning-based neural decoding of living versus non-living categories, showing that multiple nonlinear architectures can outperform a linear CSP + LDA baseline on large EEG object benchmarks (Dixen et al., 20 Mar 2025). ThingsEEG has also supported EEG-informed image generation and multimodal semantic alignment with CLIP-based image and text representations (Chen, 2024, Sun et al., 3 Sep 2025). These neighboring efforts underscore the importance of common stimuli and annotations for cross-modal comparison.

CNeuroMod-THINGS differs from these EEG-focused resources in modality, temporal granularity, and sampling strategy. Its distinctive contribution is densely sampled longitudinal fMRI within individual participants, enriched with retinotopy, functional localizers, trial-wise beta estimates, and voxelwise noise ceilings (St-Laurent et al., 11 Jul 2025). This suggests that it may function as a structural and spatial anchor for multimodal model validation: EEG-based decoding can characterize fast temporal dynamics, whereas CNeuroMod-THINGS can resolve the cortical distribution and reliability of object representations under repeated observation.

A common misconception about large visual-neuroscience datasets is that scale is exhausted by stimulus count or participant count. CNeuroMod-THINGS demonstrates a different axis of scale: repeated measurement depth within subjects across dozens of sessions (St-Laurent et al., 11 Jul 2025). Another possible misconception is that the resource is limited to category-level vision research. In fact, the continuous recognition design, confidence-based responses, and within- versus across-session repetition structure make the dataset simultaneously relevant to visual semantics and memory (St-Laurent et al., 11 Jul 2025).

By combining a continuous recognition paradigm, semantically diverse THINGS stimuli, dense longitudinal fMRI sampling, and interoperable annotations, CNeuroMod-THINGS provides an infrastructure for modelling visual semantics, category selectivity, mnemonic state, and cross-task representational structure in a unified single-subject framework (St-Laurent et al., 11 Jul 2025).