X-ViTCNN: A Novel Network-Level Fusion of Transfer Learning and Customized Vision Transformer for Multi-Stage...

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Alzheimer's disease (AD) is a leading cause of dementia, necessitating accurate and rapid diagnostic tools.
• Current deep learning methods for AD diagnosis using MRI data lack generalization, are computationally intensive, and offer limited interpretability.

Purpose of the Study:

• To develop an interpretable deep learning framework for accurate multi-stage Alzheimer's disease diagnosis from MRI scans.
• To enhance diagnostic capabilities by combining local and global feature learning from MRI data.

Main Methods:

• A novel framework, eXplorative ViT-CNN (X-ViTCNN), was developed, integrating a Vision Transformer with DenseNet201 and MobileNetV2 CNNs.
• Contrast-enhanced preprocessing and Bayesian Optimization were employed for feature highlighting and hyperparameter tuning.
• Grad-CAM visualizations were utilized to ensure model interpretability.

Main Results:

• X-ViTCNN achieved high diagnostic accuracies of 97.98% on the ADNI dataset and 94.52% on the OASIS dataset.
• The model demonstrated superior performance compared to individual baseline models and other pre-trained architectures.
• Balanced sensitivity and specificity were observed across all stages of Alzheimer's disease.

Conclusions:

• The X-ViTCNN framework offers a powerful and interpretable solution for predicting Alzheimer's disease progression using MRI.
• Its combination of feature learning, hyperparameter optimization, and interpretability makes it a valuable tool for clinicians in early AD diagnosis and patient monitoring.