108 ps coincidence time resolution through optimized scintillators, photodetectors, readout electronics, and DOI-based timing correction in orthogonally stacked detector configurations

Based on the abstract and conclusion sections of the paper, here is the summary:

Objective: Existing commercial time-of-flight positron emission tomography (TOF-PET) systems yield a coincidence time resolution (CTR) of 200 ps or less full width at half maximum (FWHM). Recently, there has been a challenge to achieve a CTR of 100 ps FWHM at the system level. However, current silicon photomultipliers (SiPMs) and 20-mm-thick scintillators in conventional single-ended readout scheme is difficult to achieve 100 ps CTR; the photon transport time spread (PTS) within the scintillator crystal is a major barrier. Differences in the interaction position result in variations in PTS on the order of several tens of ps, thereby degrading the CTR. A shorter scintillator can improve CTR; however, this can degrade detection efficiency.

Approach: To overcome this trade-off between the CTR and detection efficiency, we previously proposed xDetector, an orthogonally stacked configuration along the longitudinal axis of scintillator crystals. We investigated the CTR potential of the xDetector by improving the scintillator, photodetector, and readout electronics, and by applying CTR correction based on a three-dimensional interaction within the scintillator.

Main results: Based on error propagation, the CTR of the paired xDetector was calculated as 113.5 ± 2.7 ps FWHM. Furthermore, the CTR of the xDetector was measured at four positions along the longitudinal axis by manually sliding the xDetector, and the corrected achieved CTR was 108.6 ± 1.9 ps FWHM. Moreover, compared with the conventional single detector using a 20.0 mm scintillator, CTR improved by an average of 10.3%.

Significance: The xDetector offers potential as a PET detector concept to achieve a CTR of 100 ps FWHM. Such timing performance is expected to improve TOF-PET image quality and quantitative accuracy, contributing to more reliable disease detection and diagnosis than current PET detectors.

Conclusion: In this study, we investigated the achievable CTR of the proposed xDetector. This detector concept was designed to overcome the trade-off between the CTR and detection efficiency by improving the scintillators, SiPMs, and the readout electronics, and by applying DOI-based timing correction. Consequently, the CTR was 107.8 ± 1.2 ps FWHM in coincidence with the reference detector. Applying the DOI-based timing correction improved the CTR further to 105.3 ± 0.6 ps FWHM. Based on error propagation, the CTR for the two xDetectors in coincidence was estimated to be 108.6 ± 1.9 ps FWHM, which corresponded to an 10% improvement compared to that of the conventional single detector with a 20.0 mm scintillator. The xDetector improved CTR while maintaining detection efficiency, thereby providing a promising design concept for PET detector modules. Future work can investigate the methods for enhancing longitudinal resolution, mechanically maintaining xDetector configuration, and reducing the number of SiPMs used towards further performance improvement.

Improvements to AI Systems:

1. AI-Driven 3D Interaction-Position Correction for TOF-PET: Develop a deep learning model (e.g., CNN or transformer) that takes raw scintillation pulse waveforms from the xDetector’s multiple SiPMs and predicts the 3D photon interaction coordinates (x, y, z) within the crystal. This model would replace manual sliding calibration and heuristic DOI correction, enabling real-time, sub-millimeter interaction localization. The improved AI system can reduce CTR below 100 ps FWHM by learning nonlinear PTS variations across the entire crystal volume, directly improving image signal-to-noise ratio and lesion detectability in clinical PET.

2. AI-Optimized Detector Geometry and Readout Design: Use a generative AI or reinforcement learning agent to explore novel orthogonally stacked scintillator configurations (beyond the current xDetector) and SiPM placement patterns. The AI can simulate thousands of geometries, predict PTS and CTR via a fast surrogate model (trained on Monte Carlo data), and output optimal crystal lengths, stacking angles, and SiPM count. This improved system can automatically design PET detectors that achieve 100 ps CTR while maximizing detection efficiency, reducing the need for manual prototyping.

3. AI-Based Adaptive Timing Correction with Error Propagation: Implement a Bayesian neural network that ingests per-event timing data from both detectors in coincidence, along with predicted interaction depths, to output a corrected timestamp with uncertainty. The AI can dynamically weight the DOI correction based on signal quality and noise, minimizing CTR variance. This system can achieve sub-100 ps CTR in real-time, enabling more accurate kinetic modeling and lower radiation dose in PET scans.

4. AI-Powered Multi-Parameter Co-Optimization for PET Modules: Train a multi-objective optimization AI (e.g., Pareto-front learning) that simultaneously optimizes scintillator material properties (decay time, light yield), SiPM gain, readout electronics bandwidth, and DOI correction algorithms. The AI can predict the trade-off frontier between CTR and detection efficiency, then recommend the best operating point for specific clinical tasks (e.g., brain vs. whole-body imaging). This improved system can autonomously configure PET detectors for optimal performance across different diagnostic scenarios.

5. AI-Enhanced Longitudinal Resolution and SiPM Reduction: Use a sparse-sensing AI algorithm (e.g., compressed sensing or autoencoder) to reconstruct high-resolution longitudinal interaction positions from fewer SiPM signals. The AI can learn to infer missing spatial information from correlated waveforms, allowing a reduction in SiPM count (as noted in future work) without losing timing accuracy. This system can lower detector cost and complexity while maintaining 100 ps CTR, making high-performance TOF-PET more accessible.