Lightweight Vision Transformer with transfer learning for interpretable Alzheimer's disease severity assessment

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Alzheimer's disease (AD) diagnosis relies on early detection for effective management.
• Current diagnostic methods can be invasive or lack accessibility.
• Developing reliable, non-invasive tools is crucial for slowing AD progression.

Purpose of the Study:

• To introduce ViTTL, a lightweight deep learning framework for Alzheimer's disease severity assessment using MRI data.
• To evaluate the performance and interpretability of ViTTL.
• To demonstrate the potential for clinical translation and improved patient outcomes.

Main Methods:

• ViTTL integrates Vision Transformers (ViT) with pre-trained convolutional neural networks (CNNs) for feature extraction from 2D MRI slices.
• The ViT-DenseNet201 model combined with an artificial neural network (ANN) classifier was evaluated.
• Interpretability was achieved using LIME and GRAD-CAM methods.

Main Results:

• The ViT-DenseNet201-ANN model achieved 99.89% classification accuracy on the OASIS dataset.
• Interpretability methods consistently highlighted cortical and hippocampal regions associated with AD pathology (Dice score: 0.85 ± 0.03).
• ViTTL demonstrated significant model size reduction (83.0 MB to 6.47 MB) and robust performance on independent datasets.

Conclusions:

• ViTTL provides an accurate, interpretable, and resource-efficient solution for Alzheimer's disease diagnosis.
• The framework's reduced size and strong performance suggest potential for deployment in resource-limited clinical settings.
• ViTTL shows promise for clinical translation, potentially improving early AD detection and patient management.