Normalized group activations based feature extraction technique using heterogeneous data for Alzheimer's disease...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Deep learning networks are crucial for identifying Alzheimer's disease (AD) patterns in neuroimaging.
• While individual activation functions are studied, group activations in neuroimaging lack comprehensive exploration.
• Existing methods may not fully capture complex atrophic patterns in AD.

Purpose of the Study:

• To propose a novel feature extraction technique using normalized group activations for Alzheimer's disease detection.
• To develop an automated diagnosis system applicable to both structural MRI and resting-state fMRI (rs-fMRI).
• To evaluate the system's performance across multiple AD stages and heterogeneous imaging features.

Main Methods:

• A two-phase approach: multi-trait condensed feature extraction and regional association networks.
• Utilizing multi-layered convolutional networks for feature extraction from brain regions.
• Training regional association networks with normalized group activations and feeding outputs to a classifier.
• Testing on diverse features (curvelets, wavelets, textures, etc.) and multi-cohort ADNI data.

Main Results:

• The proposed system demonstrated effective classification of multiple Alzheimer's disease stages.
• Achieved a 1-4% performance increase in mild cognitive impairment (MCI) classifications.
• Showcased discriminatory power and efficiency on both rs-fMRI time-series and MRI data.

Conclusions:

• The normalized group activation feature extraction method is effective for AD diagnosis.
• The automated system shows promise for classifying various stages of Alzheimer's disease using neuroimaging data.
• This technique offers a valuable tool for advancing neuroimaging analysis in AD research.