LS-MAT, a generative framework integrating a variational autoencoder with GAN, latent diffusion model, and ControlNet, achieves strong performance in synthesizing personalized multimodal age-conditioned structural MRIs across ages 5-100 years, outperforming previous approaches in both modality conversion and age-conditioned synthesis tasks.
Key Findings
Methods
LS-MAT integrates three components for multimodal structural MRI synthesis: a variational autoencoder with GAN for latent encoding, a latent diffusion model for high-resolution conditional synthesis, and a ControlNet for modality-guided structural consistency.
The framework synthesizes personalized, multimodal, and age-conditioned structural MRIs
The model supports both T1-weighted and T2-weighted MRI modalities
ControlNet was used to maintain structural consistency across modalities
The generative framework is designed to support lifespan neuroimaging research
Methods
LS-MAT was trained and evaluated on large-scale publicly available datasets spanning ages 5 to 100 years.
The datasets are publicly available and cover the full lifespan range of 5-100 years
Training included both T1-weighted and T2-weighted MRI data
The model was designed to generalize across developmental and aging periods
Results
LS-MAT achieved strong performance in modality conversion tasks as measured by peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and mean squared error (MSE).
Three quantitative metrics were used: PSNR, SSIM, and MSE
Performance was evaluated for modality conversion between T1-weighted and T2-weighted MRI
The model outperformed previous approaches in modality conversion tasks
Results
LS-MAT outperformed existing methods in both modality conversion and age-conditioned synthesis tasks.
Comparisons were made against previous generative modeling approaches
The model demonstrated superiority in age-conditioned synthesis in addition to modality conversion
The framework was described as outperforming 'previous approaches' in both task categories
Results
The generated images captured established developmental and aging trends including ventricular enlargement, cortical thinning, and age-related trajectories of T1/T2-weighted ratio-based microstructural profiles.
Ventricular enlargement with age was reproduced in synthesized images
Cortical thinning trajectories were captured across the lifespan
Age-related trajectories of T1/T2-weighted ratio (T1w/T2w) microstructural profiles were reflected in the generated images
These trends align with established neuroscientific findings on brain aging
Conclusions
The LS-MAT pipeline supports longitudinal analyses and enables derivation of microstructural profile features from synthesized MRI data.
The pipeline can be used for microstructural covariance profile analysis
T1/T2-weighted ratio was used as a proxy for microstructural properties
The toolbox is openly available at https://github.com/hobacteria/LS-MAT
The framework is intended to overcome data scarcity in lifespan neuroimaging
Background
Generative modeling was proposed as a means to overcome resource intensity and data scarcity limitations of acquiring longitudinal multimodal MRI.
Acquiring longitudinal multimodal MRI was characterized as 'resource-intensive'
The framework synthesizes data conditioned on age and modality to supplement real acquisitions
The approach is intended to support characterization of age-related trajectories without requiring full longitudinal acquisitions
Kim H, Kim J, Kim M, Park H, Park B. (2026). LS-MAT: Lifespan structural magnetic resonance imaging synthesis for microstructural covariance profile analysis toolbox.. NeuroImage. https://doi.org/10.1016/j.neuroimage.2026.121840