Modeling Brain Aging With Explainable Triamese ViT: Towards Deeper Insights Into Autism Disorder.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Advanced imaging techniques like 3D MRI are crucial for medical diagnostics.
• Machine learning models are increasingly used to analyze complex medical data.
• Accurate brain age estimation and neurological disorder diagnosis remain challenging.

Purpose of the Study:

• Introduce Triamese-ViT, a novel Tri-structure of Vision Transformers (ViTs) for enhanced brain age estimation.
• Incorporate a built-in structure-aware explainability function for identifying key predictive regions.
• Improve diagnostic accuracy and interoperability with existing medical imaging techniques.

Main Methods:

• Developed Triamese-ViT, integrating information from three perspectives for brain age estimation.
• Utilized a built-in interpretability function for structure-aware explainability and visualization.
• Applied the model to analyze natural aging and diagnose Autism Spectrum Disorder (ASD).

Main Results:

• Triamese-ViT demonstrated superior performance in brain age estimation and diagnostic tasks.
• Generated insightful attention maps, validated by occlusion analysis.
• Identified key brain regions: Cingulum, Rolandic Operculum, Thalamus, and Vermis for normal aging; Thalamus and Caudate Nucleus for ASD.

Conclusions:

• Triamese-ViT offers a powerful and interpretable approach for medical diagnostics using 3D MRI.
• The model enhances understanding of neurodevelopmental and aging processes.
• Structure-aware explainability aids in identifying critical brain regions for specific conditions.