Generative synthesis of high-fidelity pathological brain MRI via label-conditioned latent diffusion

Abstract

The synthesis of high-fidelity brain Magnetic Resonance Imaging (MRI) is essential for addressing data scarcity and privacy constraints in medical research. In this study, we present a Latent Diffusion Model (LDM) for conditional MRI generation, operating in a compressed latent space to optimize computational efficiency while preserving anatomical integrity. The model is conditioned on both pathology (Healthy, Glioblastoma, Sclerosis, and Dementia) and acquisition modality (T1-weighted, T1ce, T2-weighted, FLAIR, and Proton Density). Trained on a diverse cohort of 1477 subjects from six public repositories, the framework learns disentangled representations of pathological features and imaging contrasts. To ensure methodological rigor, we comprehensively evaluated the framework across structural fidelity, latent disentanglement, and clinical utility. Quantitative analysis utilizing Fréchet Inception Distance (FID) with domain-specific MedicalNet features and Multi-Scale Structural Similarity (MS-SSIM) indicates that the synthetic images statistically align with real-world biological variance. This structural coherence is further validated by Two One-Sided Tests (TOST) for structural similarity, while deterministic label-swap experiments quantitatively support the meaningful disentanglement of the learned conditions. Furthermore, a blinded evaluation by expert radiologists confirmed that the synthetic volumes exhibit high perceptual realism, with quality scores comparable to real diagnostic scans. Crucially, the model demonstrates effective zero-shot extrapolation capabilities, successfully synthesizing anatomically plausible scans for combinations of pathology and modality that were completely absent from the training set. The functional utility of these extrapolated volumes is validated through a downstream 3D pathology classification task, where synthetic augmentation significantly stabilized decision boundaries and improved minority-class performance. By disentangling latent representations of anatomy, pathology, and modality, our approach facilitates scalable cross-condition data generation while reducing patient privacy risks, as no individual anatomy or source image is referenced during the inference phase.