Improved neural network with multi-task learning for Alzheimer's disease classification.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Neurology

Background:

• Alzheimer's disease (AD) presents a major health challenge with limited treatment options.
• Early detection of AD is critical for timely intervention and management.
• Current diagnostic methods can be invasive or lack sufficient sensitivity.

Purpose of the Study:

• To develop and evaluate an enhanced neural network, ADnet, for improved Alzheimer's disease detection using MRI scans.
• To investigate the efficacy of incorporating depthwise separable convolutions, ELU activation, and SE modules in a VGG16-based model for AD classification.
• To assess the impact of auxiliary regression tasks on the primary classification performance.

Main Methods:

• Utilized a VGG16-based neural network architecture, termed ADnet.
• Implemented depthwise separable convolutions to reduce model parameters.
• Replaced ReLU with ELU activation and integrated Squeeze-and-Excitation (SE) modules for enhanced feature extraction.
• Employed multi-task learning, including clinical dementia and mental state score regression alongside MRI classification.

Main Results:

• ADnet demonstrated a 4.18% accuracy improvement over baseline VGG16 for Alzheimer's disease (AD) vs. Cognitive Normal (CN) classification.
• Achieved a 6% accuracy improvement for Mild Cognitive Impairment (MCI) vs. CN classification.
• The integrated SE module enhanced feature extraction efficiency for improved diagnostic accuracy.

Conclusions:

• ADnet offers a promising advancement in AI-driven early detection of Alzheimer's disease from MRI data.
• The architectural enhancements and multi-task learning strategy significantly improve classification accuracy.
• This approach provides valuable support for medical professionals in the early diagnosis and intervention of AD and MCI.