Explainable Vision Transformer with Self-Supervised Learning to Predict Alzheimer's Disease Progression Using 18F-FDG...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Alzheimer's disease (AD) is a progressive neurodegenerative disorder impacting millions globally.
• Early prediction of AD progression is vital for intervention and treatment planning.
• Current methods for AD detection and staging require further development, particularly in leveraging neuroimaging biomarkers.

Purpose of the Study:

• To develop a self-supervised learning (SSL) method using vision transformers (ViT) for predicting the progression from mild cognitive impairment (MCI) to Alzheimer's disease (AD).
• To automatically extract meaningful AD characteristics from unlabeled 18F-FDG-PET images, reducing reliance on large labeled datasets.
• To enhance the interpretability of the prediction model by identifying key brain regions influencing MCI to AD transition.

Main Methods:

• Utilized a self-supervised learning approach with the vision transformer (ViT) architecture.
• Employed self-distillation with no labels (DINO) for pretraining the feature extractor and extreme learning machine (ELM) as a classifier.
• Applied the model to unlabeled 18F-FDG-PET images for predicting MCI to AD conversion.

Main Results:

• Achieved state-of-the-art classification performance with 92.31% accuracy, 90.21% specificity, and 95.50% sensitivity.
• Successfully identified brain regions significantly contributing to the prediction of MCI development.
• Demonstrated the model's ability to learn powerful representations from unlabeled neuroimaging data.

Conclusions:

• Introduced a novel Explainable SSL-ViT model for accurate prediction of AD progression using 18F-FDG-PET scans.
• Integrated SSL, attention, and ELM mechanisms for enhanced prediction accuracy and interpretability.
• The model offers a precise and efficient strategy for predicting the transition from MCI to AD, paving the way for future neurodegenerative disorder research and treatment development.