Disentangled Representation Learning for Capturing Individualized Brain Atrophy via Pseudo-Healthy Synthesis

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Brain atrophy is a key indicator in neurodegenerative diseases, correlating with cognitive decline.
• Individualized atrophy patterns are crucial for precision medicine but difficult to study longitudinally.
• Current methods struggle to accurately model disease-specific changes while preserving subject identity.

Purpose of the Study:

• To develop a deep disentangled generative model (DDGM) for capturing individualized brain atrophy patterns.
• To generate realistic pseudo-healthy counterfactual brain images and identify disease-specific residual maps.
• To evaluate the model's effectiveness in preserving healthiness and subject identity in synthetic images.

Main Methods:

• A deep disentangled generative model (DDGM) with four modules: normal MRI synthesis, residual map synthesis, input reconstruction, and mutual information neural estimator (MINE).
• Adversarial learning and MINE were used to ensure independence between disease-related and shared features.
• Comprehensive evaluation of synthetic pseudo-healthy images for healthiness and subject identity.

Main Results:

• The DDGM successfully generated pseudo-healthy images that preserved healthiness and subject identity, outperforming existing methods.
• The model demonstrated robust generalization across diverse datasets (different races and sites).
• Analysis of residual maps identified specific atrophy patterns in Alzheimer's disease (AD) patients, particularly in the hippocampus and amygdala.

Conclusions:

• The DDGM effectively captures individualized atrophy patterns, aiding in understanding neurodegenerative disease progression.
• The generated pseudo-healthy images are valuable for research and clinical applications.
• The method significantly improved Alzheimer's disease classification accuracy to 92.50 ± 2.70%.