Dynamic PET (positron emission tomography) involves continuous imaging over a predefined period immediately after tracer injection, acquiring PET images over time. With dual-tracer dynamic PET imaging, 2 image series of the radioactive tracers (radiopharmaceuticals) could be obtained simultaneously with a single scan over time, reducing the time-to-care and patient pain. However, the separation of the 2 tracers for the observation of regions of interest (ROIs) relative to each tracer is challenging, as the tracers cannot be individually identified by the PET system (the two tracers emit the exact same photon energy). In this paper, we propose a new image separation method based on neural ordinary differential equations (ODEs) to separate two tracers, FDG and PSMA, from a single dynamic acquisition. As research can currently only be conducted in silico, prior to clinical studies, we simulated time activity curves (TACs) of both tracers to generate dynamic ROIs to evaluate our methods. A comparison with a state of the art method for signal separation confirms the better performance of our approach.