The relevance voxel machine (RVoxM): a self-tuning Bayesian model for informative image-based prediction

Area of Science:

• Medical Imaging Analysis
• Machine Learning
• Bayesian Modeling

Background:

• Generic machine learning algorithms are often used for medical imaging prediction.
• These methods may lack clinical interpretability and dedicated design for imaging data.

Purpose of the Study:

• Introduce the relevance voxel machine (RVoxM), a specialized Bayesian model for medical imaging.
• To develop a model that uses clinically interpretable, spatially clustered voxels.
• To achieve high predictive accuracy with probabilistic outcomes.

Main Methods:

• RVoxM utilizes small, spatially clustered sets of voxels for analysis.
• The model automatically tunes its parameters during training.
• Probabilistic prediction outcomes are generated.

Main Results:

• RVoxM demonstrated state-of-the-art predictive accuracy in regression (age prediction from gray matter) and classification (Alzheimer's disease vs. healthy controls using cortical thickness).
• The models generated by RVoxM were found to be biologically meaningful.

Conclusions:

• RVoxM offers a powerful and interpretable approach for medical imaging prediction.
• The specialized Bayesian model provides accurate and biologically relevant insights.