One scan, many stories: deep learning for signal separation in multi-tracer PET imaging

Simultaneous imaging of multiple positron emission tomography (PET) tracers in a single acquisition promises richer multiparametric insight without prolonged protocols, addressing complex indications such as oncology or neurodegeneration. In these clinical scenarios, the concurrent assessment of metabolism, receptor binding, perfusion, or other biomarkers is valuable. Since all PET tracers produce identical 511 keV annihilation photons, separating overlapping signals is inherently challenging. Conventional approaches like compartmental modeling with staggered injections or blood sampling are often invasive, time-consuming and may be difficult to scale effectively beyond two or more tracers. Recent advances leverage deep learning (DL) to discriminate spatiotemporal patterns and incorporate kinetic priors, seeking unified frameworks that can robustly reconstruct and disentangle individual tracers.