LS-MAT: Lifespan structural magnetic resonance imaging synthesis for microstructural covariance profile analysis...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Developmental Neuroscience

Background:

• Human brain structure changes dynamically throughout life, impacting cognitive function.
• Magnetic Resonance Imaging (MRI), specifically T1- and T2-weighted sequences, are crucial for assessing brain anatomy and microstructure.
• Acquiring longitudinal multimodal MRI data is costly and time-consuming, hindering comprehensive lifespan studies.

Purpose of the Study:

• To introduce LS-MAT, a generative framework for creating personalized, multimodal, and age-conditioned structural MRIs.
• To facilitate large-scale lifespan neuroimaging research by overcoming data acquisition challenges.
• To enable detailed characterization of age-related brain structural trajectories.

Main Methods:

• Integration of a variational autoencoder with a generative adversarial network (VAE-GAN) for latent space encoding.
• Utilizing a latent diffusion model for high-resolution, conditional MRI synthesis.
• Employing ControlNet to ensure structural consistency across different MRI modalities.

Main Results:

• LS-MAT demonstrated high performance in modality conversion and age-conditioned synthesis tasks.
• Generated MRIs accurately reflected known developmental and aging brain changes, such as cortical thinning and ventricular enlargement.
• The framework successfully captured age-related microstructural profiles based on T1/T2-weighted ratios.

Conclusions:

• Generative modeling, as exemplified by LS-MAT, can effectively address data scarcity in lifespan neuroimaging.
• LS-MAT provides a valuable tool for studying dynamic structural brain changes across the entire human lifespan.
• The open-source pipeline supports longitudinal analyses and microstructural feature derivation.