OmniMIRA

An anatomically indexed foundation model across structural, molecular and metabolic neuroimaging

OmniMIRA learns brain-region representations from T1-weighted MRI, amyloid PET, FDG PET and CT. It combines local image features within atlas-defined regions, allowing the same embeddings to be used for downstream prediction and regional analysis.

Release status

Version 0.2 provides the OmniMIRA v9 inference architecture, its multi-atlas resources and an epoch-1000 checkpoint. It produces anatomically indexed features and is intended for research use, not clinical diagnosis.

Installation

Requires Python 3.9 or later and PyTorch 2.0 or later.

pip install omnimira

Usage

The wheel contains the checkpoint, model-space templates and atlas definitions. Input NIfTI images must already be spatially normalized to MNI space. The package resamples them to the committed model grid and applies modality-specific intensity normalization.

from omnimira import from_pretrained

model = from_pretrained(device="cpu")
features = model.extract("scan_mni.nii.gz", modality="t1")
print(features["aal3"].shape)  # (166, 128)

The command-line interface writes an NPZ with ROI identifiers, names and features:

omnimira scan_mni.nii.gz features.npz --modality t1

The output contains 128-dimensional embeddings for each region: 166 AAL3 regions, 69 Harvard–Oxford regions and 7 Yeo networks.

The modality options are t1 (T1-weighted MRI), av45 (amyloid PET), fdg (FDG PET) and ct (CT).

Documentation

• Input preparation
• Atlas setup
• Output format
• Model architecture
• Pretraining

Tests

Model and loss tests use synthetic inputs:

pip install pytest
python -m pytest -q tests/test_model_forward.py tests/test_losses.py

The full test suite also requires the atlas and template files.

Questions

For questions about the code, please open an issue.

License

Source code is available under the Apache 2.0 license. The bundled pretrained model weights are licensed separately under CC BY-NC 4.0, which does not permit commercial use. Third-party resources have separate terms.

For research use only; not for clinical diagnosis.