Speaker
Description
Accurate detector timing calibration is essential for time-of-flight positron emission tomography (TOF-PET), as coincidence time resolution (CTR) determines localization accuracy along the line of response (LOR) and affects image signal-to-noise ratio (SNR). Conventional calibration methods such as time-skew and time-walk correction rely on low-dimensional parametric models applied at the detector or channel level, which are often too restrictive to capture complex timing distortions arising from scintillation light transport, SiPM response, and readout electronics. Recent machine learning approaches address this limitation by regressing time differences directly from detector pair signals; however, they do not explicitly decompose timing errors into detector-wise contributions, limiting interpretability and preventing direct detector-level calibration.
We propose a structured learning framework that formulates timing calibration as detector-wise interaction-time offset prediction under coincidence-level supervision. A shared neural network estimates offsets for individual detector interactions from normalized timestamp matrices, photon-count distributions, local interaction position, deposited energy, and a learned detector embedding. Coincidence time differences are reconstructed as the difference of predicted offsets with reintroduced minimum timestamp terms, enabling parameter scaling linear in the number of detectors while sharing statistical strength across sensing elements. Training targets are derived geometrically from known annihilation positions with light-propagation corrections in LYSO.
The method was evaluated on a semi-monolithic prototype PET system comprising 24 detector modules arranged in three axial rings. Each module consists of two LYSO slab arrays read out by digital photon counters, providing $12\times12$ photon-count matrices and $6\times6$ timing channels. A total of 221 million coincidences were acquired from 450 $^{22}$Na point-source positions. Five-fold cross-validation with source-position-level splits (60/20/20) ensured evaluation on spatially unseen positions.
Across three energy windows (470--550 keV, 430--590 keV, 350--650 keV), an embedding-conditioned residual multilayer perceptron (1.83 M parameters) consistently outperformed classical calibration. For the 470--550 keV window, CTR improved from $475.16 \pm 0.57$ ps (time-skew) and $459.33 \pm 0.63$ ps (time-walk) to $422.52 \pm 0.54$ ps. Similar improvements were observed across wider energy windows, with expected degradation from inclusion of lower-energy events exhibiting higher intrinsic timing variance.
A lightweight residual MLP variant (0.89 M parameters) achieved comparable performance with only minor CTR degradation, enabling deployment on resource-constrained hardware. Feature ablation shows that, beyond timestamps, light-spread information provides the strongest auxiliary signal, positional features yield additional improvements, and deposited energy contributes only marginally. Excluding 10\% of detector pairs during training did not degrade performance on the held-out pairs at test time. This indicates that the proposed learning system effectively decomposes errors into consistent single-interaction offsets.
Our results show that the proposed learning system outperforms classical time-skew and time-walk calibration while decomposing coincidence timing errors into detector-wise interaction offsets under coincidence-level supervision, enabling statistically efficient learning, linear parameter scaling with detector count, and straightforward integration of additional signal features.
| Track | FTMI |
|---|---|
| Presentation type | Oral |