CNeuroMod: Neuroimaging & Brain Modeling

• CNeuroMod Project is an interdisciplinary neuroimaging initiative providing deeply phenotyped, large-scale fMRI datasets for modeling neural responses.
• It employs a repeated-measures design with long-form naturalistic stimuli, ensuring high-quality data for both cognitive neuroscience and neuro-AI research.
• The project underpins benchmarking efforts through challenges like the Algonauts Project Challenge, fostering development of robust encoding models.

The CNeuroMod Project is an interdisciplinary neuroimaging initiative designed to provide deeply-phenotyped, large-scale brain data for modeling and understanding neural responses to complex naturalistic and controlled stimuli. By deploying a densely sampled approach—repeatedly measuring the same core set of participants—the project generates high-quality datasets to support both cognitive neuroscience and neuro-AI research. CNeuroMod directly supplies foundational resources for benchmarking, training, and evaluating computational models of human brain function, including state-of-the-art encoding models tasked with predicting neural activity under ecologically valid conditions.

1. Project Mission and Approach

The CNeuroMod Project’s primary aim is to advance brain modeling through rich, longitudinal neuroimaging data acquisition centered on deep phenotyping. Unlike conventional population-level studies, CNeuroMod intensively samples a small number of participants, collecting up to hundreds of hours of fMRI per individual under varied task protocols—including movie watching, videogame play, and controlled cognitive/visual paradigms (St-Laurent et al., 11 Jul 2025). This approach maximizes both data reliability and within-subject variance estimation, enabling fine-grained characterization of neural dynamics, idiosyncrasies, and adaptation effects.

CNeuroMod’s methodology emphasizes:

• Long-form, naturalistic stimulation (e.g. nearly 80 hours of movie watching per subject).
• Dense sampling within subjects via repeated sessions spanning multiple task types.
• Integration with external multimodal stimulus sets (e.g. THINGS) for semantic diversity (St-Laurent et al., 11 Jul 2025).

2. Principal Datasets and Data Features

CNeuroMod delivers several flagship datasets, two of which are highlighted below:

Dataset	Subjects	Stimulus Type	Scope
Movie-watching (Algonauts)	6+	Long-form TV episodes, films	~80 hours/subject; multimodal (visual, audio, text)
CNeuroMod-THINGS	4	Object image recognition	33–36 sessions/subject; ~4000 images, 720 categories

The project’s collaboration with initiatives like THINGS allows acquisition of fMRI responses to a systematically annotated set of object images, bridging categorical and semantic domains with continuous behavioral and imaging measures (St-Laurent et al., 11 Jul 2025). For naturalistic data, extensive movie and videogame recordings support development of encoding models that span both familiar and out-of-distribution conditions (Gifford et al., 31 Dec 2024).

Quality metrics routinely reported include:

• High behavioral response rates (typically >91%)
• Low framewise displacement (most sessions <0.15 mm average)
• Voxelwise noise ceilings (e.g., up to ~73% explainable variance in select ROIs)
• Eye-tracking compliance in tasks requiring fixation

3. Benchmarking and the Algonauts Project Challenge

A major application of CNeuroMod data is benchmarking brain encoding models via open competitions. The 2025 Algonauts Project Challenge leverages the CNeuroMod movie-watching corpus to require prediction of individual-level fMRI responses to multimodal movies (Gifford et al., 31 Dec 2024).

• Models must integrate audiovisual and linguistic stimuli.
• Evaluation uses Pearson’s correlation coefficient ($r$) between predicted and actual fMRI responses; baseline models achieve $r ≈ 0.20$ (in-distribution) and $r ≈ 0.09$ (out-of-distribution).
• Public leaderboards (hosted on Codabench) foster transparent, reproducible assessments and rapid iterative model development.

Challenges are structured around both in-distribution and OOD generalization, with capped submissions in the final selection phase to promote model reliability. The challenge culminates in presentations at Cognitive Computational Neuroscience conferences.

4. Design and Analysis Methods

Data acquisition and analysis in CNeuroMod emphasize robust single-trial estimation and reliability:

• GLMsingle toolbox combines GLMdenoise and ridge regression to estimate single-trial beta values with optimal regularization.
• Noise ceiling calculation for voxel $i$ with $n$ repeats and $ncsnr = \frac{\sigma_{signal}}{\sigma_{noise}}$:

$$\text{Noise Ceiling} (\%) = \frac{100 \times ncsnr^2}{ncsnr^2 + \frac{1}{n}}$$

• Dimensionality reduction (e.g., t-SNE) visualizes semantic organization in ROI activation space.
• Memory and repetition effects are contrasted using statistical tests (e.g., t-tests on beta weights).

These procedures facilitate both region-level and whole-brain analyses while enabling modelers to relate neural representations to categorical, semantic, and behavioral variables.

5. Integration and Synergy with External Resources

CNeuroMod explicitly orchestrates data acquisition in concert with external projects such as THINGS (St-Laurent et al., 11 Jul 2025), resulting in:

• Dense, well-annotated sampling across semantic categories, supporting multi-level modeling from object to experiential timescales.
• Capacity for cross-task generalization analyses, including visual, movie-based, and multimodal paradigms.
• Expansion of available multimodal neural data for machine learning and AI benchmarking, including rigorous OOD model assessment.

This strategic integration has broadened both the scope and the scientific utility of the datasets, enabling research into individual variability, semantic encoding, and memory mechanisms.

6. Scientific and Methodological Impact

The CNeuroMod Project delivers several high-value outcomes:

• Establishes a new standard for neuroimaging dataset depth and reliability, enabling detailed modeling of brain function at the individual level.
• Enables large-scale, public benchmarking of brain encoding models for both in-distribution and OOD generalization (Gifford et al., 31 Dec 2024).
• Facilitates systematic paper of phenomena such as repetition suppression, semantic clustering, and cross-modal neural integration.
• Advances neuro-AI modeling efforts by supplying data volumes sufficient for training state-of-the-art architectures and validating computational hypotheses about neural representations.

This suggests that the project’s design can serve as a template for future dataset-driven research in both theoretical and applied computational neuroscience.

7. Prospective Directions

Future directions for CNeuroMod include:

• Further expansion of subject cohorts and stimulus diversity.
• Inclusion of additional behavioral and physiological measurements (e.g., eye tracking, memory assessment).
• Application of datasets in adaptive neuro-AI frameworks, especially architectures modeling flexible neural structures, neuromodulation, and biological learning dynamics.
• Continued participation in public benchmarking initiatives and collaborative challenge design to accelerate methodological progress in neural encoding and decoding.

Plausible implications are the further integration of CNeuroMod data with multimodal AI pipelines and the establishment of persistent public neuroimaging benchmarks, sustaining open competition and methodological rigor.