Compton PET: a simulation study for a PET module with novel geometry and machine learning for position decoding

Abstract

This paper describes a simulation study of a positron emission tomography (PET) detector module that can reconstruct the kinematics of Compton scattering within the scintillator. We used a layer structure, with which we could recover the positions and energies for the multiple interactions of a gamma ray in the different layers. Using the Compton scattering formalism, the sequence of interactions can be estimated. The true first interaction position extracted in the Compton scattering will help minimize the degradation of the reconstructed image resolution caused by intercrystal scatter events. Because of the layer structure, this module also has readily available user-defined resolution for the depth of interaction. The semi-monolithic crystals enable high light collection efficiency and an energy resolution of ∼10% has been achieved in the simulation. We used machine learning to decode the gamma ray interaction locations, achieving an average spatial resolution of 0.40 mm. Our proposed detector design provides a pathway to increase the sensitivity of a system without affecting other key performance features.

A Comparison of Neural Networks and Center of Gravity in Muon Hit Position Estimation

“…We call them Fully Connected Network (FCN) and CNN. In previous studies, it has been proven that the neural networks can achieve good spatial resolution on particle hit position estimation on positron emission tomography (PET) systems [ 8 , 9 , 10 ]. Since we have a similar task, we decided to utilize neural networks in our methods as well.…”

“…They trained 7 neural networks for x and 7 neural networks for y as sub-interval networks. In addition to MPL architecture, [ 10 ] used a convolutional neural network (CNN) to decode the gamma-ray interaction locations, achieving an average spatial resolution of 0.40 mm. They collected the light distribution as a 16-pixel image using the sensors placed at the edges of the plate.…”

The promise of artificial intelligence and deep learning in PET and SPECT imaging

“…In comparison with other positioning algorithms, such as Anger logic and correlated signal enhancement, which rely on determination of the centre of gravity, machine learning algorithms led to a better position estimation particularly at the crystal edges [16]. In this regard, Peng et al trained a CNN classifier that was fed with signals from each Silicon photomultiplier's channel to the coordinates of the scintillation point for a quasi-monolithic crystal [17]. Another study applied a multi-layer perceptron to predict the 3D coordinates of the interaction position inside a monolithic crystal and compared the performance of this positioning algorithm with anger logic for a preclinical PET scanner based on NEMA NU4 2008 standards [18].…”

Applications of artificial intelligence and deep learning in molecular imaging and radiotherapy

“…Moreover, implementation of deep learning-based positioning would not add computational burden compared to other algorithms, such as maximum likelihood estimation methods, and benefits from lower complexity for execution on graphical processing units (GPUs). In this regard, a convolutional neural network (CNN) was trained to map the light distribution over SiPMs (charge collected from each channel of the SiPMs) as input to 2D position-of-interaction as classification output for a quasi-monolithic detector (Peng et al 2018 ). Regarding other position estimators, such as COG, machine learning techniques resulted in superior spatial resolution owing to the reduced positioning bias, particularly at the edges of the PET detectors (Müller et al 2018 ).…”