Project lead: Connor Lane
Discord channel: #fmri-fm
GitHub: https://github.com/MedARC-AI/fmri-fm
MedARC Meetings Calendar: public link | iCal link
Our goal is to train a foundation model for functional MRI (fMRI) data. The basic strategy is to leverage large-scale publicly available fMRI data to train models that can "decode" noisy fMRI data into structured, interpretable "neuro embeddings". In turn, we hope that these learned neuro embeddings will enable downstream tasks such as identifying early biomarkers of neurodegenerative disease, distinguishing subtypes of neuropsychiatric conditions, predicting an individual's response to specific mental health treatment, and decoding the contents of a person's perception.
Our current approach is based on a novel flat map representation of the fMRI data. This approach strikes a balance between parcellation based representations, which are straightforward to model but lose a lot of information, and native 4D volume based approaches, which preserve the full fidelity fMRI data but require the model to learn the complex structure of fMRI from scratch.
