Med-DDPM model generates Alzheimer's MRI with 0.7244 Dice score

A team of researchers from the University of Colorado and affiliated institutions has adapted the Med-DDPM conditional diffusion model — originally built for brain tumor synthesis — to generate 3D structural MRIs specific to Alzheimer's disease. Published on arXiv and accepted at IEEE MIPR 2025, the work tackles a core challenge in neuroimaging: the subtle, region-specific anatomical changes of Alzheimer's are hard for generative models to capture. The team solved this by conditioning the diffusion process on anatomical segmentation masks derived from the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset, ensuring the synthesized images preserve key brain structures affected by neurodegeneration.

The results are striking: segmentation models trained exclusively on synthetic MRI data achieved a Dice score of 0.6532 — nearly identical to the 0.6513 from real-only training — and showed significantly higher recall. More importantly, models trained on hybrid datasets (mixing real and synthetic scans) reached a Dice score of 0.7244, outperforming both pure real and synthetic baselines. This demonstrates that conditional diffusion models can produce anatomically accurate, AD-specific synthetic MRIs that augment limited real-world datasets, improve diagnostic model performance, and enable privacy-preserving research in Alzheimer's neuroimaging.