Related Experiment Video
Updated: Feb 18, 2026
11:54
Microfluidic Platform with Multiplexed Electronic Detection for Spatial Tracking of Particles
Published on: March 13, 2017
9.9K
Using convolutional neural networks to estimate time-of-flight from PET detector waveforms
Eric Berg1, Simon R Cherry1,2,3
1Department of Biomedical Engineering, University of California-Davis, Davis, CA, United States of America.
Physics in Medicine and Biology
|November 29, 2017
Summary
Deep convolutional neural networks (CNNs) enhance time-of-flight (TOF) estimation in positron emission tomography. This machine learning approach improves timing resolution by over 20% compared to traditional methods.
Area of Science:
- Medical Imaging
- Nuclear Physics
- Machine Learning
Background:
- Current time-of-flight (TOF) positron emission tomography (PET) detectors primarily use simple signal processing like leading edge discrimination (LED) or constant fraction discrimination (CFD).
- These traditional methods, rooted in analog electronics, do not fully exploit the rich timing information available in digitized detector waveforms.
- Advancements in waveform digitizer technology enable the application of sophisticated signal processing techniques for improved TOF accuracy.
Purpose of the Study:
- To apply deep convolutional neural networks (CNNs), a machine learning technique, for direct estimation of TOF from digitized detector waveforms in PET.
- To evaluate the performance of CNN-based TOF estimation against conventional methods like LED and CFD.
- To investigate the impact of CNN architecture on timing resolution.
Main Methods:
- Utilized a machine learning approach employing deep convolutional neural networks (CNNs) to process digitized detector waveforms from coincident events.
- Generated ground-truth TOF data experimentally by controlling the path length difference between positron emission and detector pairs.
- Employed two photomultiplier tube-based scintillation detectors and a point source moved in 5 mm increments over a 15 cm range, with waveforms digitized at 10 GS/s.
Main Results:
- CNN-based TOF estimation achieved a timing resolution of 185 ps.
- This represents a significant improvement: 20% better than LED (231 ps) and 23% better than CFD (242 ps).
- CNN depth was identified as the most critical factor influencing timing resolution, while parameters like filter size had minimal impact.
Conclusions:
- Deep convolutional neural networks offer a superior method for TOF estimation in PET compared to traditional signal processing techniques.
- The developed CNN approach significantly enhances timing resolution, paving the way for more precise PET imaging.
- Future research should focus on optimizing CNN architectures, particularly depth, for further advancements in PET detector signal processing.
