Multimodal-Conditioned Latent Diffusion Models for Spatiotemporal Brain MRI Forecasting

MC-LDM performs patient-specific spatiotemporal brain MRI forecasting, reducing average regional volumetric MAE by 18% versus the second-best baseline on an independent test set of 664 visits. The latent diffusion model is trained with MRI, tau-PET, CDR and APOE4, but at inference uses only baseline MRI, APOE4 and patient metadata. It was trained on 1,355 visits from 852 subjects in ADNI.