Neural Autoregressive Modeling of Brain Aging

Area of Science:

• Computational neuroscience
• Medical imaging analysis
• Artificial intelligence in healthcare

Background:

• Brain aging simulation is crucial for understanding neurological disorders and personalized medicine.
• Existing generative models face challenges with high-dimensional neuroimaging data and capturing subtle, subject-specific aging patterns.
• Predicting future brain structure from early MRI scans offers valuable insights into individual aging trajectories.

Purpose of the Study:

• To develop a novel model, NeuroAR, for high-fidelity brain aging synthesis using generative autoregressive transformers.
• To accurately simulate the structural evolution of the brain over time from earlier magnetic resonance imaging (MRI) scans.
• To overcome the limitations of current generative models in capturing subject-specific aging patterns and data complexity.

Main Methods:

• Proposed NeuroAR, a generative autoregressive transformer model for brain aging simulation.
• Synthesized future brain scans by autoregressively estimating discrete token maps from concatenated embeddings of past and future scans.
• Incorporated subject's previous scan, acquisition age, and target age via cross-attention to guide the generation process.

Main Results:

• NeuroAR demonstrated superior performance in image fidelity compared to state-of-the-art models like latent diffusion models (LDM) and generative adversarial networks.
• Evaluations on elderly and adolescent subjects confirmed NeuroAR's ability to model subject-specific brain aging trajectories.
• A pre-trained age predictor validated the consistency and realism of synthesized images with expected aging patterns.

Conclusions:

• NeuroAR effectively models subject-specific brain aging trajectories with high fidelity, outperforming existing generative approaches.
• The model's autoregressive transformer architecture provides a robust framework for complex neuroimaging data synthesis.
• NeuroAR shows significant potential for applications in clinical neuroscience, computational modeling, and personalized health monitoring.