Conditional Latent Diffusion for Synthetic Brain MRI in Alzheimer’s Disease: A Preprocessing-Focused Pipeline

Source abstract

Alzheimer’s disease (AD) detection from structural magnetic resonance imaging (MRI) using deep learning depends on large, labelled datasets. However, many cohorts contain only a few hundred subjects, for which standard augmentation adds very limited diversity. Latent diffusion models (LDMs) offer an alternative. The current study presents a reproducible two-stage pipeline that generates 2D coronal brain MRI conditioned on diagnosis and trained on 295 subjects from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). In the first stage, a variational autoencoder with an adversarial objective compressed 256 × 256 slices into a 32 × 32 × 8 latent space. Then, a class-conditional LDM with classifier-free guidance generated AD and cognitively normal (CN) images. A baseline and a redesigned pipeline adding MNI152 registration, anatomically guided slice selection, and increased latent capacity were compared on the same subjects. The redesigned pipeline reached a Kernel Inception Distance of 0.030 ± 0.002 and a bias-corrected Fréchet Inception Distance (FID∞) of 42.14. A classifier trained only on synthetic data achieved an area under the curve (AUC) of 0.754 (95% CI, 0.60 to 0.91), compared with 0.810 for real data. No instance memorisation was detected across 880 samples, and a control analysis exposed a 32.7-percentage-point inflation in the standard memorisation metric under unequal reference sets. The redesigned preprocessing pipeline improves synthesis quality at a small-cohort scale and offers an approach that may be transferable to other privacy-restricted medical imaging domains.