Quantification of Cognitive Function in Alzheimer's Disease Based on Deep Learning

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Alzheimer's disease (AD) presents with progressive cognitive decline and neuropsychiatric symptoms, significantly impacting elderly quality of life and imposing societal burdens.
• Current diagnostic methods for AD can be invasive or lack comprehensive information for early and accurate quantification.

Purpose of the Study:

• To develop and validate a novel cascaded 3D neural network framework for distinguishing Alzheimer's disease (AD), Mild Cognitive Impairment (MCI), and normal controls using multimodal neuroimaging.
• To investigate the efficacy of graph theory metrics in identifying AD-related brain network alterations.
• To propose an efficient channel pruning method for depth-separable convolutional units within the neural network architecture.

Main Methods:

• Graph theory analysis was applied to brain networks, extracting parameters like node degree, efficiency, and betweenness centrality to identify significant differences between AD patients and healthy individuals.
• A cascaded 3D Convolutional Neural Network (CNN) framework was designed to process multimodal MRI and PET images, employing depth-separable convolutions for computational efficiency.
• A novel channel pruning method tailored for depth-separable convolutions was developed and implemented to optimize the network's performance.

Main Results:

• Significant differences in graph theory parameters were identified in specific brain regions between normal subjects and AD patients, highlighting potential biomarkers.
• The proposed cascaded 3D CNN framework demonstrated effectiveness in distinguishing AD and MCI from normal samples using combined MRI and PET data.
• The channel pruning method successfully reduced computational complexity while maintaining comparative performance, indicating efficient feature extraction.

Conclusions:

• Multimodal neuroimaging combined with advanced deep learning techniques offers a powerful approach for the accurate quantification and early detection of Alzheimer's disease.
• Graph theory analysis provides valuable insights into brain network alterations associated with AD.
• The developed channel pruning strategy enhances the efficiency of deep learning models for neuroimaging analysis in AD research.