RelA-Diffusion: Relativistic Adversarial Diffusion for Multi-Tracer PET Synthesis from Multi-Sequence MRI

A research team from the University of North Carolina at Chapel Hill has introduced RelA-Diffusion, a novel AI framework that synthesizes multi-tracer positron emission tomography (PET) scans from standard magnetic resonance imaging (MRI) sequences. The model addresses critical limitations in current neurological imaging where comprehensive PET assessment requires multiple radioactive tracers—each exposing patients to radiation, costing thousands of dollars, and facing availability constraints. By leveraging both T1-weighted and T2-FLAIR MRI scans as complementary inputs, RelA-Diffusion captures richer structural information to guide the generation of synthetic PET images that reveal tau accumulation, neuroinflammation, and β-amyloid deposition.

The technical breakthrough lies in integrating a gradient-penalized relativistic adversarial loss into the intermediate predictions of a diffusion model. This approach compares real and generated images in a relative manner, encouraging more realistic local structures while stabilizing training through adversarial feedback at each diffusion timestep. Extensive experiments on two clinical datasets demonstrate that RelA-Diffusion outperforms existing methods in both visual fidelity and quantitative metrics, producing synthetic PET scans with significantly reduced artifacts. The framework's ability to generate multiple tracer types from routine MRI could transform neurological diagnostics, making comprehensive brain pathology assessment more accessible and safer for patients worldwide.