3D Brain MRI Generation with a Clinically-Conditioned VAE-GAN and Diffusion-Driven Feature Sampling

Najmeh Mashhadi, Emmanouil Nikolakakis, Razvan Marinescu

Keywords: 3D VAE-GAN, latent diffusion sampling, Conditional Synthetic Brain MRI

TL;DR: The paper proposes a 3D VAE-GAN framework to generate clinically consistent brain MRI volumes conditioned on key attributes, enabling controllable and high-quality neuroimaging data synthesis for Alzheimer’s disease research and related tasks.

Abstract:

We introduce a 3D VAE-GAN framework that synthesizes brain MRI volumes con-
ditioned on seven clinical attributes, such as Alzheimer’s disease (AD) diagnostic
labels and key volumetric measures, including the hippocampus, amygdala, and
lateral ventricle, which are known to correlate with AD. Leveraging a 3D encoder-
decoder with depthwise-separable convolutions and a style-based modulation, our
model efficiently captures critical biomarkers and injects clinical information di-
rectly into the generation process. During the training, two pre-trained auxiliary
heads, Alzheimer’s Disease and Cognitively Normal (AD/CN) classification and
brain volume vector regression, provide additional cross entropy and regression
losses, ensuring that generated scans remain anatomically plausible and clinically
consistent. To sample realistic clinical vectors during inference, we additionally
train a diffusion model on clinical vectors, enabling flexible sampling of disease
states without the need for manual feature engineering. Experimental results
demonstrate high-quality 3D MRI generations. Additionally, adjusting disease
labels or specific brain volumes demonstrates a feasible level of conditional control,
suggesting that this approach could benefit data augmentation and support clinically
relevant neuro-imaging tasks.

Submission Number: 62