Ensemble Vision Transformer for Dementia Diagnosis.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Neuroscience

Background:

• Deep learning is increasingly used for computer-aided Alzheimer's Disease (AD) diagnosis.
• Traditional methods often use multiple learners or focus on partial brain anatomy.

Purpose of the Study:

• Introduce a novel deep learning approach, Monte Carlo Ensemble Vision Transformer (MC-ViT), for enhanced AD diagnosis.
• Overcome limitations of existing methods in characterizing 3D brain anatomy and inter-feature correlations.

Main Methods:

• Developed MC-ViT, an ensemble approach using a single Vision Transformer (ViT) learner with Monte Carlo sampling.
• Evaluated the method on the Alzheimer's Disease Neuroimaging Initiative (ADNI) and Open Access Series of Imaging Studies-3 (OASIS-3) datasets.

Main Results:

• Achieved 90% accuracy in AD classification with minimal preprocessing.
• MC-ViT demonstrated superior performance compared to 2D-slice CNNs and 3D CNNs.
• The method effectively discerns 3D inter-feature correlations in brain anatomy.

Conclusions:

• MC-ViT offers a powerful and efficient deep learning solution for Alzheimer's Disease diagnosis.
• The novel ensemble strategy enhances diagnostic accuracy by leveraging a single, robust learner.
• This approach represents a significant advancement in neuroimaging analysis for AD.