Deep-learning based separation method of a mixture of dual-tracer single-acquisition PET signals with equal half-lives

A deep belief network based method of separating a mixture of dual-tracer positron emission tomography (PET) signals from dual-tracers labelled with the same isotope, which comprises the following steps:

• (1) injecting the dual tracers, including tracer I and tracer II, labelled with the same isotope to a biological tissue and performing dynamic PET imaging on the biological tissue, then obtaining a coincidence counting vector corresponding to different moments, and then forming a dynamic coincidence counting sequence reflecting a concentration distribution of a mixture of the dual tracers, denoted as Sdual;
• (2) injecting individual tracer I and tracer II to the biological tissue sequentially and performing two separate dynamic PET image acquisitions on the biological tissue respectively, obtaining coincidence counting vectors of the two separate dynamic PET image acquisitions corresponding to different moments, and constituting the dynamic coincidence counting sequences that respectively reflect the distribution of tracer I and tracer II, denoted as SI and SII;
• (3) using a PET reconstruction algorithm to reconstruct dynamic PET images Xdual, XI and XII corresponding to the dynamic coincidence counting sequences Sdual, SI and SII;
• (4) repeating step (1)˜(3) multiple times to generate a plurality of dynamic PET image sequences Xdual, XI and XII and dividing them into training sets and testing sets;
• (5) extracting the time activity curve (TAC) of each pixel from the dynamic PET image sequences Xdual, XI and XII, taking the TACs of the training set Xdual as the input sample, and the TACs of the corresponding XI and XII as the ground truth, then training them by the deep belief network to obtain a dual-tracer PET reconstruction model;
• and
• (6) inputting every TAC of any Xdual in the test set into the PET reconstruction model one by one, and outputting the TACs corresponding to XI and XII, and finally reconstructing the TACs to obtain dynamic PET images XtestI and XtestII corresponding to tracer I and tracer II.