Unbiased TOF estimation using leading-edge discriminator and convolutional neural network trained by...

Area of Science:

• Medical Imaging
• Nuclear Physics
• Machine Learning

Background:

• Convolutional neural networks (CNNs) can improve coincidence time resolution (CTR) in time-of-flight positron emission tomography (TOF-PET).
• Current CNN training requires extensive data from multiple source positions, posing challenges for TOF estimation bias and variance.
• Existing methods face limitations in TOF estimation accuracy, particularly near the edges of the training data space.

Purpose of the Study:

• To develop a novel, unbiased method for TOF estimation in TOF-PET.
• To enable CNN training using data from a single source position, simplifying the process.
• To improve the CTR and signal-to-noise ratio in TOF-PET imaging.

Main Methods:

• A hybrid approach combining a conventional leading-edge discriminator (LED) with a CNN was proposed.
• The method focuses on estimating and correcting time difference errors from the LED, rather than absolute time differences.
• The CNN is trained using signal waveforms from a single source position, creating a continuous, symmetric data distribution.

Main Results:

• The proposed method successfully estimated all source positions without bias, using data from a single source location.
• Evaluation using scintillation detector waveforms confirmed the elimination of TOF estimation bias.
• The method demonstrated an improvement in CTR compared to the conventional LED approach.

Conclusions:

• The developed method offers an unbiased approach to TOF estimation in TOF-PET.
• Single-source training data requirement simplifies CNN application in TOF-PET.
• Improved CTR enhances signal-to-noise ratio and contributes to direct positron emission imaging advancements.