OpenBHB: a Large-Scale Multi-Site Brain MRI Data-set for Age Prediction and Debiasing

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Neuroscience

Background:

• Predicting chronological age from neuroimaging data, particularly Magnetic Resonance Imaging (MRI), serves as a proxy for biological age and can indicate deviations towards brain disorders.
• Current machine learning (ML) models for brain age prediction lack a standardized benchmark, hindering consensus on optimal model selection.
• Large neuroimaging datasets often exhibit site-specific biases, impairing the generalization capabilities of ML models, especially Deep Learning (DL) algorithms prone to simplicity bias.

Purpose of the Study:

• To introduce Open Big Healthy Brains (OpenBHB), a novel, large-scale, multi-site public benchmarking resource for neuroimaging.
• To establish a permanent challenge focused on brain age prediction and site-effect removal using a representation learning framework.
• To facilitate the development and comparison of ML models for accurate brain age estimation and bias mitigation.

Main Methods:

• The OpenBHB resource comprises over 5,000 3D T1 brain MRI scans from healthy controls across more than 60 global centers and 10 studies.
• Datasets are uniformly preprocessed and include 3D Voxel-Based Morphometry (VBM) maps, quasi-raw aligned images, and Surface-Based Morphometry (SBM) indices.
• A representation learning framework is employed for the challenge, enabling participants to benchmark their models against each other on a public leaderboard.

Main Results:

• The OpenBHB dataset provides a standardized and diverse collection of neuroimaging data for robust model training and validation.
• The associated challenge promotes the development of advanced ML techniques capable of accurate brain age prediction and effective removal of acquisition site biases.
• Uniform preprocessing and quality control ensure data consistency and comparability across different sites and studies.

Conclusions:

• OpenBHB offers a valuable, scalable resource for advancing brain age prediction research and understanding neurodevelopmental trajectories.
• The challenge framework encourages innovation in ML algorithms, particularly in addressing data heterogeneity and improving model generalizability in neuroimaging.
• This initiative aims to foster collaboration and accelerate discoveries in the field of computational neuroscience and brain health monitoring.