Explainable Normative Modeling for Brain Disorder Identification in Resting-State fMRI.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Accurate identification of brain disorders is crucial for timely intervention and improved patient outcomes.
• Existing AI models for rs-fMRI analysis often rely on supervised learning, requiring large annotated datasets and potentially missing subtle patterns.
• Unsupervised approaches like normative modeling offer an alternative by learning normality from healthy controls' data.

Purpose of the Study:

• To propose BRAINEXA, a novel framework for unsupervised brain disorder identification using rs-fMRI.
• To enhance normative modeling for rs-fMRI by improving normality construction and ensuring explainability.
• To identify clinically meaningful disruptions of brain function in an unsupervised setting.

Main Methods:

• BRAINEXA employs a novel training strategy predicting informative from less informative regions to construct accurate and stable normality models.
• Spatiotemporal mutual information regularization is incorporated to preserve distinctiveness in latent representations.
• Normality-defining (ND) subregions are extracted for interpretability, combined with anomaly scores for region- and connection-wise explanations.

Main Results:

• BRAINEXA effectively identifies brain disorders using unsupervised learning on rs-fMRI data.
• The framework demonstrates improved representation learning and prevents representational distortions.
• Region- and connection-wise explanations are generated, aiding in the identification of clinically relevant abnormalities.

Conclusions:

• BRAINEXA offers a powerful unsupervised approach for brain disorder identification from rs-fMRI data.
• The framework enhances normative modeling with improved normality construction and interpretability.
• BRAINEXA's ability to provide explainable insights facilitates clinical understanding of brain dysfunction.