MLTiming: Pairwise-Constrained Learning for Improving Timing Resolution in TOF-PET

Accurate timing characterization of detected radiation events in Positron Emission Tomography (PET) provides Time-of-flight (TOF) information for detected coincidences, which improves the signal-to-noise ratio of the reconstructed images. In this work, we propose a method, called MLTiming, to train machine learning (ML) models to assign accurate timestamps to events measured in radiation detectors training using just the signals detected by the individual single-channel detectors, without the need of resorting to other detector signals nor simulations. We performed a proof-of-concept study using three different neural network (NN) architectures, including the recently proposed Kolmogorov-Arnold network (KAN), trained with event pairs constructed with the measured signal and a delayed version of it. The delay is used as a label in the training process. The proposed trained model was evaluated with coincidence data acquired with a22Na point source using a pair of LaBr3(Ce) crystals in the shape of truncated cones coupled to fast photomultiplier tubes (PMTs). The detected γ-ray pulses were sampled at 5 Gsamples/s. A 3% improvement in Coincidence Time Resolution (CTR) over an ultra optimized genetic algorithm–enhanced digital CFD (D-CFD) method (246±3 ps to 238±1 ps) was achieved. The possibility of training ML models for each individual detector using only the information contained in their own detected events may facilitate the use of ML models for large and medium-sized scanners, as well as other research areas. MLTiming is available at https://github.com/wjosefw/MLTiming.git.