In Silico Dynamic Dual-Tracer PET Image Separation for Prostate Cancer Diagnosis using Deep Learning

Dynamic positron emission tomography (PET) involves continuous imaging over a defined period after tracer injection, sequentially capturing PET images. For the diagnosis of prostate cancer, two radiotracers, PSMA and FDG, are usually necessary. In dual-tracer dynamic PET imaging, two series of radioactive tracer images could be obtained simultaneously within a single scan, reducing time-to-care. However, the PET acquisition cannot differentiate the two tracers since they emit photons of identical energy. The aim of this work is to study deep learning methods to separate two radiotracers from a dual tracer data. Given that research is presently limited to in silico studies before moving to clinical trials, we simulated time-activity curves (TACs) for both tracers to generate dynamic regions of interest (ROIs) around the prostate. We studied three deep learning methods to perform signal separation. The results demonstrate that the architecture taking into account temporal information shows better performance.