MedARC structural MRI foundation model

Open research project building a structural MRI foundation model, targeting the FOMO26 challenge.

For the final FOMO26 push, we are working exclusively in src/fomo_tune/.

Installation

git clone --recurse-submodules https://github.com/MedARC-AI/smri-fm
cd smri-fm
uv sync
uv run pre-commit install --install-hooks

If you are on the medarc cluster, use the shared huggingface cache

export HF_HOME="/data/smri-datasets/huggingface"

This will avoid re-downloading the model checkpoints and datasets.

Train the default probe head for Task 1.

uv run python -m fomo_tune.main_task1 train

Path
`src/fomo_tune/`	The active work. One `main_task<k>.py` per task, plus the frozen backbone, data loading and container build. See the `README.md` for more info.
`src/smri_mae/`	MAE pretraining code. Currently stable, not accepting changes.
`experiments/<name>/`	Self-contained experiments: config, launch script, analysis, figures.
`third_party/`	Third party code. Has the challenge's container validator, and its submission spec.
`CODING_STANDARDS.md`	Read before writing code.
`AGENTS.md`	Orientation specifically for coding agents.