Leveraging Swin Transformer for enhanced diagnosis of Alzheimer's disease using multi-shell diffusion MRI | JoVE Visualize

Area of Science:

Neuroimaging
Artificial Intelligence
Biomedical Diagnostics

Background:

Early diagnosis of Alzheimer's disease (AD) and detection of amyloid pathology are crucial for effective treatment.
Multi-shell diffusion MRI (dMRI) offers rich microstructural information potentially useful for early AD detection.
Deep learning, particularly vision transformers, shows potential for analyzing complex medical imaging data.

Purpose of the Study:

To develop and evaluate a deep learning framework using multi-shell dMRI for early Alzheimer's disease diagnosis.
To leverage vision transformer models for detecting amyloid accumulation in the brain.
To support biomarker-driven diagnostics in data-limited neuroimaging settings.

Main Methods:

A Swin Transformer-based deep learning pipeline was developed for classifying Alzheimer's disease and amyloid presence using multi-shell dMRI.
Diffusion Tensor Imaging (DTI) and Neurite Orientation Dispersion and Density Imaging (NODDI) metrics were extracted and projected onto 2D planes.
Low-Rank Adaptation was integrated to efficiently train the transformer model on limited neuroimaging data.

Main Results:

The framework achieved high classification accuracy, with a balanced accuracy of 95.2% for distinguishing Alzheimer's disease dementia from cognitively normal individuals using NODDI metrics.
Amyloid detection yielded balanced accuracies of 77.2% (distinguishing amyloid-positive MCI/AD dementia from amyloid-negative controls) and 67.9% (identifying amyloid-positive individuals among cognitively normal subjects).
Explainability analysis highlighted key brain regions like the parahippocampal gyrus and hippocampus in model predictions.

Conclusions:

Diffusion MRI combined with transformer architectures shows significant potential for the early detection of Alzheimer's disease and amyloid pathology.
The proposed framework demonstrates feasibility and effectiveness in data-limited biomedical settings.
This approach supports the advancement of biomarker-driven diagnostics for neurodegenerative diseases.