A deep neural network for positioning and inter-crystal scatter identification in multiplexed PET detectors: a...

Area of Science:

• Medical Imaging Physics
• Nuclear Instrumentation
• Computational Imaging

Background:

• High-resolution positron emission tomography (PET) requires precise photon positioning for accurate image reconstruction.
• Conventional positioning algorithms in light-sharing PET detectors struggle with edge effects and inter-crystal scattering (ICS).
• Finely segmented and highly multiplexed PET detectors present unique challenges for event localization.

Purpose of the Study:

• To explore the feasibility of deep neural network (DNN) techniques for enhanced event positioning in light-sharing PET detectors.
• To investigate the impact of inter-crystal scattering (ICS) on positioning accuracy.
• To compare DNN-based positioning with conventional Anger logic.

Main Methods:

• Simulated PET detector events using Geant4 Application for Tomographic Emission (GATE) to study inter-crystal scatter (ICS).
• Developed and trained a DNN for crystal localization using simulated photoelectric (P) and Compton + photoelectric (CP) events.
• Integrated an energy-guided positioning algorithm into the DNN, leveraging ICS properties.
• Compared DNN performance against Anger logic using simulated 511 keV point sources.

Main Results:

• The DNN achieved crystal classification accuracies of 90% for P events and 82% for CP events.
• DNN-based positioning outperformed Anger logic by at least 34% for P events and 14% for CP events.
• The ratio of backward to forward gamma-ray scattering in the crystal array near 1 limits further improvement for CP events.

Conclusions:

• DNN-based event positioning shows significant potential to improve accuracy in light-sharing, multiplexed PET detectors.
• DNNs offer a substantial enhancement over Anger logic for 2D coincidence event positioning.
• Further advancements may require additional information, such as event timing, to overcome physical limitations.