Deep-Learning-Based Separation of a Mixture of Dual-Tracer Single-Acquisition PET Signals With Equal Half-Lives: A Simulation Study

Abstract

Dual-tracer positron emission tomography (PET) can characterize various aspects of physiology and function by a single scan, both of which increase the accuracy of the diagnosis and benefit the patients significantly. However, the separation of dual-tracer PET images can be challenging as they yield indistinguishable 511-keV photon pairs. The existing methods rely on the differences in the kinetic behaviors of the compartment models and radioactive decay to differentiate dual tracers with staggered injection, thereby limiting the level of separation precision and increasing the dual-tracer scanning time. We herein propose a new method to separate the mixed dual-tracer signals with equal half-lives that allow the simultaneous injection of the two tracers. A deep belief network (DBN) framework with a supervised mechanism was constructed to map the complex relationship between dual-tracer signals and two individual signals. The DBN is composed of three restricted Boltzmnan machines (RBMs) and a supervised neural network with the same structure. Correspondingly, the whole training process includes two primary steps: 1) a forward pretraining by three RBMs and 2) a fine-tuning by the supervised neural network with the back-propagation (BP) algorithm. We validated the proposed DBN framework on four dual-tracer groups using the Monte Carlo simulation with different counts levels: 1) [18F] fluorodeoxyglucose–[18F] fluorothymidine on the Hoffman brain phantom; 2) [62Cu] ATSM–[62Cu] PTSM on the Zubal thorax phantom; 3) [11C] flumazenil (FMZ)–[11C] dihydrotetrabenazine (DTBZ) on the complex brain phantom; and 4) [11C] FMZ–[11C] DTBZ on the six-ball phantom. We demonstrate that our method can provide a satisfactory performance regarding the separation of dual tracers labeled by the same radionuclide with the order of $10^{6}$ to $10^{7}$ total counts.