Accurate brain-age models for routine clinical MRI examinations

This repository contains scripts to enable readers to run the trained models presented in Accurate brain-age models for routine clinical MRI examinations, (Wood et al., 2022, NeuroImage) and Optimising brain age estimation through transfer learning: a suite of pre-trained foundation models for improved performance and generalisability in a clinical setting, (Wood et al., 2024, Human Brain Mapping). The models were trained on > 23,000 axial T2-weighted head MRI scans from two large UK hospitals, and demonstrate accurate, generalisable brain-age prediction:

Requirements

The code requires the data to be in Nifti file format (.nii or .nii.gz extension) and makes use of the Project Monai library. This repository is compatible with python 3.8. See requirements.txt for all prerequisites; you can also install them using the following command:

pip install -r requirements.txt

Usage

Running our models is straightforward. All that is needed is a .csv file with the following two columns:

• 'ID' (which is a unique identifier for each participant/scanning session e.g., 'pat119' etc.).

• 'file_name' (which gives the absolute path to the Nifti file for each participant).

Brain-age prediction can then be performed using the following basic command:

python run_inference.py --project_name NAME --csv_file /PATH/TO/CSV/FILE

This will save a .csv file within the local cloned repository (./NAME_output.csv) with the brain-predicted ages for each subject. Optionally, users can also provide a third column named 'Age', which gives the chronological age of each participant in years, in order to generate performance metrics (e.g., mean absolute error [MAE] and scatter plots). To do this, simply add the argument --return_metrics to run_inference.py.

By default, our model will run on a cpu. If a GPU is available, run.py should be called with the additional argument --gpu (in this case, run time is <2 seconds per scan).

To run our skull-stripped T2 model, simply add the argument --skull_strip to run_inference.py. Likewise, to use our volumetric T1-weighted ensemble model (which also relies on skull-stripping), then run_inference.py should be called with the following additional arguments:

--sequence t1 --ensemble

Note skull-stripping is performed using HD-BET - a deep-learning based brain extraction tool which takes ~10 seconds per scan with a gpu and ~ 1 minute without.

To run our axial diffusion-weighted, axial susceptibility-weighted, coronal flair or sagittal T1-weighted models, then run_inference.py should be called with the following additional arguments:

-- dwi

-- swi

-- cor_flair

-- sag_T1

Please note that our model only provides meaningful brian-age predictions for scans that are oriented in the 'LPS' coordinate system (i.e., right to Left, anterior to Posterior, inferior to Superior). For this reason, run_inference.py automatically reorients scans to this coordinate system.

Citation

If you found this repository useful, please consider citing our work:

Wood, D. A., Townend, M., Guilhem, E., Kafiabadi, S., Hammam, A., Wei, Y., ... & Booth, T. C. (2024). Optimising brain age estimation through transfer learning: A suite of pre‐trained foundation models for improved performance and generalisability in a clinical setting. Human Brain Mapping, 45(4), e26625.

Wood, D. A., Kafiabadi, S., Al Busaidi, A., Guilhem, E., Montvila, A., Lynch, J., ... & Booth, T. C. (2022). Accurate brain-age models for routine clinical MRI examinations. NeuroImage, 118871. https://doi.org/10.1016/j.neuroimage.2022.118871