Unsupervised waveform reshaping for enhanced time-of-flight precision in PET detectors

Area of Science:

• Medical Imaging
• Nuclear Physics
• Machine Learning

Background:

• Accurate Time-of-Flight (TOF) estimation is crucial for improving image quality in Positron Emission Tomography (PET).
• Real detector signals in TOF-PET are often distorted by noise and electronic factors, hindering precise TOF measurement.
• Existing methods for signal correction may require extensive labeled data or are hardware-dependent.

Purpose of the Study:

• To enhance TOF-PET image reconstruction by improving TOF estimation accuracy.
• To develop an unsupervised waveform reshaping method using Cycle-Consistent Generative Adversarial Networks (CycleGAN).
• To transform distorted real detector waveforms into clean, simulated waveforms for better signal-to-noise ratio (SNR).

Main Methods:

• Utilized a real event signal dataset from a 22Na gamma source and LYSO-SiPM detectors.
• Proposed an unsupervised CycleGAN framework to map distorted experimental waveforms to clean Geant4 simulated waveforms.
• Employed cycle-consistency constraints, eliminating the need for paired or labeled data.
• Applied Constant Fraction Discrimination (CFD) timing to reshaped waveforms for TOF estimation.

Main Results:

• Achieved a coincidence time resolution (CTR) improvement of 41 ps (15.1%) compared to conventional CFD on raw signals.
• Demonstrated an 81 ps (26.0%) CTR improvement over Leading Edge Discrimination (LED).
• Exhibited lower edge bias compared to supervised Convolutional Neural Network (CNN) approaches.

Conclusions:

• The unsupervised TOF enhancement framework reduces reliance on large labeled datasets and hardware.
• CycleGAN enables efficient and generalizable waveform reshaping for improved TOF estimation.
• This approach offers a practical and scalable strategy for advancing high-precision TOF-PET systems.