Joint a priori kinetic modeling with neural ODE for dual-tracer dynamic PET separation

Simultaneous dual-tracer dynamic PET imaging acquires a time series of images and could streamline oncology workflows by enabling the quantification of complementary biological processes in a single session. However, separating the superimposed tracer signals from the resulting mixed image sequence remains a key challenge. We introduce a physics-informed framework that combines Neural Ordinary Differential Equations (NODEs) with a priori compartmental kinetic models to achieve robust signal separation and generate individual tracer image sequences. By embedding the tracer dynamic models directly into the learning procedure, our method ensures the decomposed time-activity curves (TACs) are physically consistent. A superpixel-based preprocessing step clusters voxels in the dynamic data into homogeneous regions, substantially improving noise robustness and computational efficiency. Experiments on clinical dynamic PET data confirm our approach accurately reconstructs the individual tracer image sequences, outperforming existing methods. This work provides a generalizable solution for dual-tracer dynamic PET, with direct implications for the diagnosis of prostate tumors.