Physiology-guided Self-supervised Learning for Simultaneous Dual-Tracer PET Separation

Simultaneous dual-tracer PET provides more comprehensive information for clinical diagnosis than standard PET imaging, but separating the hybrid dual-tracer signal remains challenging. Deep learning (DL) offers a promising solution. However, most DL methods rely on large datasets with spatiotemporal alignment between dual-tracer and two single-tracer scans. Precise alignment across different scans is difficult, making data collection costly and time-consuming, thereby limiting scalability. To avoid these difficulties and improve clinical robustness, we propose a simultaneous dual-tracer PET self-supervised separation method (DTPSS) guided by tracer kinetic prior information. By incorporating kinetic priors and integrating the ODEs residual constraints derived from the parallel two-compartment model into the loss function, DTPSS enables accurate and robust dual-tracer separation without access to single-tracer labels. Compared with data-driven supervised methods and self-supervised kinetics-aware method, DTPSS achieved improved quantitative and qualitative performance across two datasets (2,400 brain image samples with and without deformation). It also shows robustness across tracer combinations and sampling protocols. In animal study (13 rats with [¹⁸F]FDG / [¹⁸F]FET and 17 mice with [¹⁸F]FDG / [¹⁸F]FAPI), DTPSS achieved improved separation results under real-data settings with unavailable paired single-tracer labels and substantial physiological variability.