An explainable three dimensional framework to uncover learning patterns: A unified look in variable sulci...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Three-dimensional (3D) global explanations are vital in neuroimaging for interpreting complex data.
• Existing methods often lack accuracy and comprehensibility for 3D neuroimaging data.
• A need exists for advanced explainable AI (XAI) to provide reliable 3D global explanations.

Purpose of the Study:

• To develop an accurate and low-complexity XAI 3D-Framework for neuroimaging.
• To generate reliable 3D global explanations for deep learning models in neuroscience.
• To investigate brain structure variations associated with psychosis using AI.

Main Methods:

• Developed an XAI 3D-Framework integrating statistical features (Shape) and XAI methods (GradCam, SHAP).
• Employed dimensionality reduction to ensure explanations reflect model learning and cohort variability.
• Evaluated the framework on 3D deep learning models trained on 596 structural MRIs for paracingulate sulcus (PCS) detection.

Main Results:

• The framework provides accurate, low-complexity 3D global explanations, reducing inter-method variability.
• Identified critical sub-regions (posterior temporal, internal parietal, cingulate, thalamus) linked to PCS presence/absence.
• Demonstrated the framework's ability to uncover developmental contexts of cortical features.

Conclusions:

• The XAI 3D-Framework enhances the faithfulness and reliability of global explanations in neuroimaging.
• This approach offers insights into normative brain development and atypical trajectories in mental illness.
• Advances AI interpretability in neuroimaging for more reliable applications.