Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session C: Friday, August 15, 2:00 – 5:00 pm, de Brug & E‑Hall

CNeuroMod Data Collection Complete: 200h of individual fMRI Across Diverse Naturalistic and Controlled Tasks to build NeuroAI models

Julie A. Boyle¹, Basile Pinsard, Marie St-Laurent², Pierre Bellec³, Courtois Neuromod Team; ¹Centre de Recherce de l' IUGM, ²Centre de Recherche de l'IUGM, ³Université de Montréal

Presenter: Julie A. Boyle

Abstract Dense fMRI datasets are gaining popularity as a powerful resource to build integrative NeuroAI models to understand the brain. The Courtois Project on Neuronal Modelling (CNeuroMod) has now completed a 5-years data collection process resulting in 200 hours of fMRI data per subject using diverse naturalistic and controlled cognitive tasks. CNeuroMod is the largest dense fMRI dataset currently available to support the development of individualized and generalizable AI models of complex brain processes. Keywords: dense fMRI dataset, multimodal datasets, NeuroAI, individual brain models, Algonauts 2025 Introduction Several large individual fMRI datasets have emerged to train artificial intelligence (AI) models on specific cognitive processes like natural image (NSD: Allen et al. 2021; BOLD5000: Chang, et al. 2019) and movies (Dr Who: Seeliger et al. 2019) viewing. However, a key feature of the brain is the capacity to integrate and switch between specialized processes and cognitive contexts.The CNeuroMod project, which just completed its data collection, is the largest individual fMRI dataset to date. We collected a rich neuroimaging dataset that probes numerous cognitive domains in the same subjects (N=6) using both carefully controlled and engaging naturalistic tasks in order to build versatile and complex Neuro-AI models. Additionally, the overlap between CNeuroMod tasks and other open data resources opens the possibility to test NeuroAI models transferability and generalization across subjects and datasets. Methods Participants (aged 31 to 47 at the time of recruitment in 2018); 3 women and 3 men; 3 native French & 1 native English speakers, 2 bilingual; all right-handed) consented to participate for at least 5 years of data collection. All participants had good general health and normal hearing for their age. Four subjects were scanned 80h+ / year and two were scanned 40h+ / year. FMRI data were acquired with a 3T scanner). Setup included physiological signal recording, headcases to minimize motion, and a custom-built fiber optic controller to play videogames (Harel et al., 2023). Data were preprocessed with the fMRIprep pipeline LTS (Esteban et al, 2019). The CNeuroMod Databank The databank includes 29 fMRI datasets targeting several cognitive domains and modalities. Per subject, there are 197h of fMRI data, 17h of anatomical MRI data (Boudreau et al, 2025), 29h of data collected outside the scanner, including longitudinal hearing measures (Fortier et al, 2025) and videogame (Shinobi) training. The following is a breakdown of the fMRI datasets by cognitive domains, per subject (Ss), Table 1). Vision. 22 datasets (175h) including movies (10h), 7 seasons of TV-show Friends (70h), & functional localizers (Stigliani et al, 2015; Kay et al. 2013). Language. 19 datasets (73h), including listening to Le Petit Prince (3h; Li et al, 2022) in 3 languages, reading a chapter of harrypotter (Wehbe et al 2014), triplets is a semantic association task involving triplets of words (7h), functional localizers (Scott et al, 2017 & Malik-Moralda et al 2021). Memory. 5 datasets (44h), including a memory task (18h) using THINGS (18h; Hebart et al. 2019), and multfs (8h), which is a working-memory task. Emotions. 13 datasets (147h), including gifs (5h) that evoke varying emotional dimensions (Cowen & Keltner, 2017). Auditory. 18 datasets (146h), including Mutemusic, which is an auditory imagery task, and narratives, a listening task (Nastase et al., 2018) with verbal recal inside the scanner. Videogames (48h). In-scan playing of videogames including Shinobi, SuperMario, Mariostars, and Mario3. OOD Algonauts (2h/Ss). A secret dataset acquired for the Algonauts 2025 challenge (Gifford et al., 2025). Data access & release Raw and preprocessed fMRI data, behavioral responses and physiological recordings are formatted in BIDS (Gorgolewski, et al, 2016) and available via DataLad version control. After processing and quality checks, each dataset will be released independently with an accompanying data paper in the coming years. Data for 4 subjects is accessible without any restrictions (CC0 license) on the Canadian Open Neuroscience Platform's portal (https://portal.conp.ca/), while data for all subjects is available via registered access at https://www.cneuromod.ca/. Conclusion CNeuroMod has assembled an unprecedented resource to model individual brain function using a wide range of controlled and naturalistic tasks. Data from the movie watching tasks are currently in use to assess the robustness of brain encoding models for the Algonauts 2025 challenge (Gifford et al., 2025). The wealth of additional data that will be released in the coming years will fuel novel insights into the ways human brains process complex stimuli. Acknowledgements The Courtois NeuroMod project was made possible by a grant from la Fondation Courtois given to LPB.

Topic Area: Brain Networks & Neural Dynamics

Extended Abstract: Full Text PDF