CoCoLIT: ControlNet-Conditioned Latent Image Translation for MRI to Amyloid PET Synthesis

Daniele Ravi

Abstract: Medical imaging synthesis between modalities has become an important task in medical imaging. While cross-modality synthesis has enabled many downstream applications, generating high-fidelity medical images remains challenging. We propose CoCoLIT, a ControlNet-conditioned latent image translation approach for synthesizing Amyloid PET from structural MRI. Our approach leverages a pre-trained latent diffusion model conditioned with a ControlNet to enable anatomically-guided cross-modal synthesis. We demonstrate the effectiveness of our approach on the ADNI dataset for MRI to Amyloid PET synthesis, achieving high-quality results that preserve anatomical structures while accurately modeling the distribution of amyloid pathology.