PETNet -- Coincident Particle Event Detection using Spiking Neural Networks

TL;DR

PETNet applies spiking neural networks to coincidence photon detection in PET, introducing a multi-objective loss that jointly optimizes spike counts and timing and a Dirichlet-window geometry encoding to leverage detector layout. The approach achieves competitive accuracy and substantial inference-time savings on simulated clinical and preclinical PET data, outperforming a traditional SCW baseline on the clinical set and offering up to about $36\times$ faster inference on the larger SAFIR dataset. These results argue for the practicality of SNNs as real-time, low-energy level-two triggers in particle-detection contexts and point toward future neuromorphic hardware deployment. The work also highlights remaining challenges, such as hyperparameter sensitivity on large-scale data and the need for sparse, hardware-efficient implementations.

Abstract

Spiking neural networks (SNN) hold the promise of being a more biologically plausible, low-energy alternative to conventional artificial neural networks. Their time-variant nature makes them particularly suitable for processing time-resolved, sparse binary data. In this paper, we investigate the potential of leveraging SNNs for the detection of photon coincidences in positron emission tomography (PET) data. PET is a medical imaging technique based on injecting a patient with a radioactive tracer and detecting the emitted photons. One central post-processing task for inferring an image of the tracer distribution is the filtering of invalid hits occurring due to e.g. absorption or scattering processes. Our approach, coined PETNet, interprets the detector hits as a binary-valued spike train and learns to identify photon coincidence pairs in a supervised manner. We introduce a dedicated multi-objective loss function and demonstrate the effects of explicitly modeling the detector geometry on simulation data for two use-cases. Our results show that PETNet can outperform the state-of-the-art classical algorithm with a maximal coincidence detection $F_1$ of 95.2%. At the same time, PETNet is able to predict photon coincidences up to 36 times faster than the classical approach, highlighting the great potential of SNNs in particle physics applications.

Figures (4)

• Figure 1: Schematic representation of the SCW sorting algorithm, modified from oliver2008comparison. Three event classes are displayed: (Type II) are two hits within the defined coincidence window frame (gray dotted), resulting in an accepted coincidence. (Type I) is rejected as only one hit is detected, (Type III) is rejected as more than two hits are registered.
• Figure 2: Overview of PETNet's coincidence detection process. Left: functional principle of PET -- detection of photons emitted through annihilation of the $\beta^+$ tracer particles with tissue electrons using a ring of scintillating crystals. Center: input spike trains of detector hits and corresponding coincidences (gray dotted box). Right: PETNet, a supervised denoising spiking neural network with LIF neurons.
• Figure 3: Visualization of the explicit geometry modeling. Left: Dirichlet window (orange) for a detector with $C=8$ crystals, a hit at crystal $c=2$ (blue) and window size $w=1$. Center: time-resolved detector hits. Right: corresponding time-resolved geometry spikes.
• Figure 4: Loss and evaluation metrics $TP$, $FP$ and $FN$ compared to the ground truth over the course of training for PETNet and the SAFIR dataset (top) and for the LSTM baseline model and the Clinical dataset (bottom).