Transformer-CNN hybrid network for improving PET time of flight prediction

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Signal Processing

Background:

• Positron Emission Tomography (PET) reconstruction benefits from time-of-flight (TOF) information, enhancing image signal-to-noise ratio.
• Precise TOF estimation relies on accurate coincidence time resolution (CTR) and minimal bias.
• Existing methods for TOF estimation have limitations in fully utilizing detector waveform data.

Purpose of the Study:

• To develop and evaluate a novel deep learning network for improved TOF estimation in PET.
• To integrate Transformer and Convolutional Neural Network (CNN) architectures to leverage both global and local waveform features.
• To assess the network's performance against established methods using experimental data.

Main Methods:

• A hybrid network combining Transformer's self-attention with CNN's local receptive fields was proposed.
• The network utilized event waveform pairs from lutetium yttrium oxyorthosilicate (LYSO) scintillators and silicon photomultiplier (SiPM) detectors as input.
• Waveform datasets were cropped and used for training and testing the model.

Main Results:

• The proposed Transformer-CNN network achieved an average CTR of 189 ps, outperforming CFD, CNN, CNN with attention, LSTM, and Transformer.
• This represents a CTR reduction of up to 82 ps (>30%) compared to CFD.
• The network also demonstrated reduced bias by 10.3 ps compared to standard CNN.

Conclusions:

• The study validates the efficacy of Transformer-based models for PET TOF estimation with real experimental data.
• The integrated CNN and Transformer approach offers optimal performance by combining local and global waveform analysis.
• This research presents a promising new direction for advancing TOF-PET reconstruction.