Abstract
In Positron Emission Tomography (PET) reconstruction, utilizing Time of
Flight (TOF) information can significantly enhance the signal-to-noise
(SNR) ratio, posing a greater challenge for the precision of TOF. To
address this, we employed two distinct waveform datasets for training
our developed network. One dataset comprises simulated waveform data
obtained through a comprehensive simulation process established using
Geant4 and GosSip. The other dataset consists of real waveform data
collected from lutetium yttrium orthosilicate (LYSO) scintillators and
silicon photomultiplier (SiPM) detectors placed at various positions.
Our network, a combination of Transformer and Convolutional Neural
Network (CNN), was developed for predicting the TOF of coincidence
events based on waveform data from PET detectors. Our network achieved
average full width at half maximum (FWHM) of 189 ps, with reductions of
82 ps and 13 ps compared to constant fraction discriminator (CFD) and
CNN, across multiple positions. Additionally, there was an average bias
reduction of 10.3 ps compared to CNN. We visualized the attention map,
revealing the remarkable enhancement of Transformer on the rising edge
of waveforms. We also demonstrated the robustness of our proposed
network by including waveforms with scattered events in the real
training dataset. Data augmentation through translation and flip was
investigated and resulted in an improvement of 5 ps. Furthermore, we
analyzed the characteristic differences between real and simulated
waveform data, providing valuable insights for generating more realistic
simulated data in the future. Our network improved the average FWHM and
bias, leading to enhanced SNR and clearer imaging. Data augmentation
effectively expanded the dataset and facilitated the data collection
process.