Synthetic Brain MRI Generation Using Class-Conditional Latent Diffusion Models

Abstract

The lack of annotated training data and the significant distributional skew across diagnostic categories still represent major challenges to automated medical image analysis, especially in developing robust deep learning models for detection and classification of brain MRI tumours. Generative modelling provides a technically sound way to augment limited clinical collections with realistic synthetic imagery. However, pixel-level adversarial architectures are known to be unstable during optimisation, and they often sacrifice fine structural details. The paper proposes a class-guided latent diffusion framework for targeted synthesis of brain MRI volumes. The methodology combines a Variational Autoencoder (VAE) to learn a dense, geometrically consistent latent embedding of MRI scans with a conditioningaware diffusion network to learn the category-stratified latent distribution for normal and tumour tissue independently. Grey scale brain MRIs at full resolution are projected to a compact low dimensional representation without loss of essential morphological cues. A self-attention enhanced conditional latent diffusion model then produces category-aligned latent codes which are then reconstructed to the image domain via the frozen VAE decoder. The experimental results show that the framework generates synthetic volumes that are structurally coherent and contain recognisable anatomical signatures. This method enables a computationally-feasible, adversary-free synthesis pathway that is directly applicable to enhancing training data and fortifying downstream classifiers in clinical MRI pipelines.