Speaker
Description
In dynamic brain positron emission tomography (PET) studies, recovering the images in the missing time frame is often required in order to reduce the scanning protocol, or to perform kinetic modelling with sparse dynamic information. Likewise, the rapid dual-tracer studies, which aim to administer two tracers in a single scan staggered in time, will largely be benefitted if the later frames of the first administered tracer can be predicted where both tracers overlap, as it will allow to single out the signal from the second tracer as well by simple subtraction from the total measurement. This indicates that each individual tracer information would be recovered, which has been a challenge owing to the fact that all tracers give rise to indistinguishable 511KeV annihilation photon pairs as a signal to PET scanner. Traditionally, this was done by setting a kinetic model that consists of sets of ordinary differential equations (ODE), such as parallel compartment model, and fitting the measured time-activity curve (TAC). Here, we introduce the novel deep learning model, neural ODE which shares the same concept in data-driven manner. Simply put, the neural ODE solves sets of ODE and converges into the functional shapes that best describe the underlying pharmacokinetic processes. We customized the neural ODE and applied the proposed model to 60-minute dynamic 18F-PI-2620 brain PET images such that it will predict the late 30-min kinetics, given the early 30-min frame images as an input.
