Abstract
Brain aging synthesis is a critical task with broad applications in clinical and computational neuroscience. The ability to predict the future structural evolution of a subject's brain from an earlier MRI scan provides valuable insights into aging trajectories. Yet, the high-dimensionality of data, subtle changes of structure across ages, and subject-specific patterns constitute challenges in the synthesis of the aging brain. To overcome these challenges, we propose NeuroAR, a novel brain aging simulation model based on generative autoregressive transformers. NeuroAR synthesizes the aging brain by autoregressively estimating the discrete token maps of a future scan from a convenient space of concatenated token embeddings of a previous and future scan. To guide the generation, it concatenates into each scale the subject's previous scan, and uses its acquisition age and the target age at each block via cross-attention. We evaluate our approach on both the elderly population and adolescent subjects, demonstrating superior performance over state-of-the-art generative models, including latent diffusion models (LDM) and generative adversarial networks, in terms of image fidelity. Furthermore, we employ a pre-trained age predictor to further validate the consistency and realism of the synthesized images with respect to expected aging patterns. NeuroAR significantly outperforms key models, including LDM, demonstrating its ability to model subject-specific brain aging trajectories with high fidelity.
Introduction
Understanding how the brain changes over time is vital for diagnosing neurodegenerative diseases and understanding development. However, longitudinal studies—where the same person is scanned repeatedly over years—are incredibly expensive and time-consuming. While AI can synthesize aging by generating images, doing so for 3D brain MRIs is exceptionally difficult due to the high dimensionality of the data and the need to maintain subject-specific traits (the unique "fingerprint" of an individual's brain). Existing methods like Generative Adversarial Networks (GANs) often struggle with 3D consistency, and Latent Diffusion Models (LDMs) can be computationally expensive.
Methodology
NeuroAR (Neural Autoregressive Model) shifts the paradigm from pixel-wise denoising to scale-wise prediction using a hierarchical approach:
- Multi-Scale VQVAE (MS-VQVAE): The model compresses 3D MRI scans into discrete "tokens" across five different spatial scales. This allows the model to handle high dimensionality by working in a compressed latent space.
- Autoregressive Transformer: Instead of predicting all pixels at once, NeuroAR predicts the tokens of a future scan scale-by-scale. It uses the tokens from the subject’s previous scan as a foundation to ensure anatomical consistency.
- Age-Guided Conditioning: To guide the aging process, the model uses Adaptive Normalization Layers (AdaLN) and Multi-Head Cross-Attention to inject the subject's current age and the desired target age into every block of the transformer.
- Concatenated Latent Space: By concatenating the embeddings of the past and future scans, the transformer learns the specific morphological shift required to move from one age to another.
Experimental Results
The model was evaluated on three major neuroimaging datasets: ADNI and PPMI (elderly population, ages 31–95) and the ABCD Study (adolescents, ages 9–16). NeuroAR significantly outperformed state-of-the-art Latent Diffusion Models (LDM) and StarGAN across all metrics.
Quantitative Performance
NeuroAR achieved superior image fidelity and structural similarity:
- ADNI Dataset: NeuroAR reached a PSNR of 21.08, compared to 18.10 for LDM.
- ABCD Dataset: NeuroAR achieved an SSIM of 0.836, far exceeding LDM’s 0.674.
- Downstream Utility: Using NeuroAR to augment a training set for age regression reduced the Mean Absolute Error (MAE) from 0.685 to 0.582, proving the synthetic scans carry meaningful biological information.
Qualitative Realism
Visual inspections showed that NeuroAR preserved anatomical structures, such as the ventricles and cortical folding, much more accurately than competing models, which often produced blurry or distorted results.
Conclusion
NeuroAR demonstrates that scale-wise autoregressive modeling is a powerful approach for synthesizing subject-specific brain aging trajectories. By maintaining anatomical consistency and high image fidelity, it offers a valuable tool for longitudinal neuroscience research. Future work will focus on extending this model to pathological aging cases, such as simulating the progression of Alzheimer’s Disease.
Summary updated