An Attention-Based 3D CNN With Multi-Scale Integration Block for Alzheimer's Disease Classification

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Alzheimer's Disease (AD) diagnosis faces challenges with existing models processing single-scale brain atrophy and high computational costs.
• Convolutional Neural Networks (CNNs) show promise but require enhancement for comprehensive AD analysis.

Purpose of the Study:

• To introduce a novel Attention-based 3D Multi-scale CNN model (AMSNet) for improved Alzheimer's Disease diagnosis.
• To enhance the capture and integration of multi-scale spatial features of AD with a concise model structure.

Main Methods:

• Developed AMSNet, an Attention-based 3D Multi-scale CNN, to process structural MRI (sMRI) scans.
• Evaluated AMSNet on binary classification (AD vs. Cognitively Normal - CN) and three-way classification (AD-MCI-CN).
• Compared AMSNet's performance against seven other models regarding accuracy, sensitivity, specificity, parameters, and computational load.

Main Results:

• AMSNet achieved high performance in binary classification (91.3% accuracy, 88.3% sensitivity, 94.2% specificity).
• The model demonstrated superior performance with fewer parameters and lower computational load compared to existing models.
• AMSNet showed good generalization capabilities in three-way AD-related classification tasks.

Conclusions:

• AMSNet effectively integrates multi-scale spatial features and utilizes an attention mechanism for efficient AD classification.
• The proposed model offers a feasible and efficient approach for AD diagnosis using sMRI.
• AMSNet's findings suggest potential biomarkers for exploring the neuropathological underpinnings of Alzheimer's Disease.