A Conjugate Bayesian Framework for Fast 3D Positronium Lifetime Estimation with a Partial System Matrix
Berkin Uluutku, Giulianno Gasparato, Manish Das, Jarosław Choiński, Anand Pandey, Sushil Sharma, Paweł Moskal, Ewa Stępień, Chien-Min Kao, Hsin-Hsiung Huang · 2026
Read A Conjugate Bayesian Framework for Fast 3D Positronium Lifetime Estimation with a Partial System Matrix by Berkin Uluutku, Giulianno Gasparato, Manish Das, Jarosław Choiński, Anand Pandey, Sushil Sharma, Paweł Moskal, Ewa Stępień, Chien-Min Kao, Hsin-Hsiung Huang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-09804.txt.
Introduction
Uluutku and nine co-authors develop a scalable framework for three-dimensional positronium lifetime estimation from positron emission tomography data. Voxel-wise estimation is hard because detector-time channels multiply while only a small subset is observed. The team restricts the forward model to observed channels, using a partial system matrix, posterior weighting, and a conjugate Gamma-Exponential update. Restricting the model cuts memory and compute needs. In a J-PET scanner dataset with a NEMA phantom, a 234,375-voxel rate map was estimated in about 3 seconds from about 364,000 retained events.
Catalog record
| CREATOR | Berkin Uluutku, Giulianno Gasparato, Manish Das, Jarosław Choiński, Anand Pandey, Sushil Sharma, Paweł Moskal, Ewa Stępień, Chien-Min Kao, Hsin-Hsiung Huang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 51515 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.09804
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
A Conjugate Bayesian Framework for Fast 3D Positronium Lifetime Estimation with a Partial System Matrix Berkin Uluutku, Giulianno Gasparato, Manish Das, Jarosław Choiński, Anand Pandey, Sushil Sharma, Paweł Moskal, Ewa Stępień, Chien-Min Kao, Hsin-Hsiung Huang Abstract Background: Positronium lifetime imaging extends conventional positron emission tomography by using the time interval between positron emission and annihilation as an additional contrast mechanism. Voxel-wise lifetime estimation in fully three-dimensional settings is computationally difficult because the number of feasible detector-time channels grows rapidly, whereas only a small subset is observed in practice. We developed a scalable statistical framework for three-dimensional positronium lifetime estimation based on a time-of-flight-aware partial system matrix restricted to observed detector-time channels, combined with posterior event-to-voxel weighting and a conjugate Gamma–Exponential update for closed-form voxel-wise effective-rate estimation. Results: Restricting the forward model to observed detector-time channels reduced memory and computational requirements while preserving the Poisson data model for retained detected triple coincidences. In simulated data with 4056 voxels, the analytic Bayesian estimator required 2.76 s versus 74.46 s for 10 L-BFGS-B iterations on the same CPU while accurately recovering the effective-rate map. In a triple-coincidence dataset acquired with a J-PET prototype scanner and a NEMA image-quality phantom, a 234 375-voxel effective-rate map was estimated in approximately 3 s from about 3.64 × 10 5 3.64\times 10^{5} retained events. Conclusions: Restricting the system matrix to observed detector-time channels makes fully three-dimensional positronium lifetime estimation computationally practical for sparse triple-coincidence data. The proposed posterior-weighted conjugate update provides a fast and stable single-component surrogate estimator of voxel-wise effective lifetime for large-scale three-dimensional positronium lifetime imaging. keywords Bayesian inference, positronium lifetime imaging, time-of-flight PET 1 Background Positronium lifetime imaging (PLI) extends conventional positron emission tomography (PET) by exploiting the time interval between positron emission and annihilation as an additional source of image contrast ( 13 ) . Because positronium properties depend on the surrounding microenvironment, the measured lifetime carries information that complements conventional activity imaging ( 14 ) . Over the past several years, experimental demonstrations of positronium imaging have progressed rapidly, including dedicated detector systems ( 13 ; 5 ; 17 ; 20 ) , high-resolution phantom studies ( 5 ; 3 ; 4 ; 8 ) , and recent in vivo measurements on clinical and prototype scanners ( 12 ; 11 ) . These developments motivate reconstruction methods capable of estimating voxel-wise lifetime parameters in fully three-dimensional settings. Reconstruction methods for positronium imaging now include direct statistical formulations for time-of-flight (TOF) PET 15 , surrogate likelihood methods such as SPLIT 6 , and fast high-resolution reconstruction strategies such as SIMPLE 7 . At the same time, recent experimental studies on commercial scanners have emphasized that realistic lifetime analysis is influenced by multi-component decay behavior, detector timing response, and random triple coincidences 19 ; 10 . The main challenge addressed in this work is efficient three-dimensional estimation when only a small fraction of feasible detector–time channels is actually observed in the data. We consider a two-stage estimation strategy. First, we estimate a detected activity image representing the expected number of retained detected triple coincidences contributed by each voxel. Second, conditional on this activity estimate, we estimate a voxel-wise effective annihilation-rate map using a single-component exponential model. The resulting rate should be interpreted as a compact representation of the observed timing signal rather than a complete multi-component model of positronium decay. The paper makes three contributions. First, we formulate a TOF-aware partial system matrix defined only on observed detector–time channels, substantially reducing memory and computational cost while preserving the Poisson data model. Second, we introduce an event-to-voxel weighting formulation that interprets the unknown source voxel of each event as a latent variable, enabling principled fractional attribution of events across voxels. Third, we introduce a conjugate Bayesian estimator that yields closed-form voxel-wise annihilation-rate estimates, enabling fast three-dimensional estimation while providing a compact mean representation of the observed positronium lifetime signal. Finally, the proposed framework is evaluated using both simulated data and an experimental triple-coincidence dataset acquired with a J-PET prototype scanner and a NEMA Image-Quality phantom ( 3 ) . The results demonstrate that the partial system matrix formulation enables computationally tractable three-dimensional reconstruction while the conjugate Bayesian estimator provides stable lifetime estimates and voxel-wise uncertainty information in realistic experimental conditions. 2 Methods This section presents the 3D estimation framework and simulation setup. To make volumetric estimation computationally feasible, we replace the full system matrix with a partial system matrix defined only on observed detector–time channels. We then estimate voxel-wise annihilation rates using both maximum likelihood and a conjugate Bayesian update, the latter providing fast inference and voxel-wise uncertainty estimates. Figure 1 summarizes the proposed workflow. Figure 1: Proposed estimation workflow 2.1 Event Representation and Preprocessing Each recorded decay is treated as a triple coincidence consisting of two annihilation-photon interactions and one prompt-gamma interaction. In list mode, event k is stored as e_{k}=\{d_{1,k},d_{2,k},t_{1,k},t_{2,k},d_{p,k},t_{p,k}\}, (1) where d_{1,k} and d_{2,k} are the detectors for the annihilation photons in arrival order, d_{p,k} is the detector for the prompt gamma ray, and t_{1,k},t_{2,k},t_{p,k} are the corresponding detection times. The annihilation-pair TOF observable is \Delta t_{k}=t_{2,k}-t_{1,k}. (2) Because the annihilation photons are indexed in arrival order, \Delta t_{k}\geq 0 . With TOF-bin width \Delta\tau_{\mathrm{bin}} and TOF window \Delta t_{\max} , the TOF bin index is \xi_{k}=\min\!\left(\left\lfloor\frac{\Delta t_{k}}{\Delta\tau_{\mathrm{bin}}}\right\rfloor,\,N_{\xi}-1\right),\qquad N_{\xi}=\left\lceil\frac{\Delta t_{\max}}{\Delta\tau_{\mathrm{bin}}}\right\rceil. (3) We denote the ordered annihilation-detector pair by \eta_{k}=(d_{1,k},d_{2,k}) and the detector-time channel by c_{k}=(\eta_{k},\xi_{k}) . The event-wise prompt-to-annihilation time difference is \tau_{k}=t_{a,k}-t_{p,k},\qquad t_{a,k}=\frac{t_{1,k}+t_{2,k}}{2}. (4) Because the voxel of origin is unknown, \tau_{k} contains both the desired lifetime information and an unmodeled flight-path mismatch. In the present single-component exponential surrogate model, we retain only events with \tau_{k}\geq 0 . This filtering removed 3.6% of simulated detected triples. In the experimental dataset it removed 20.6% of all recorded triples and 8.3% of events within the estimated 44 Sc regions. The larger fraction among all recorded triples is consistent with enrichment of background or mismatched events among negative- \tau_{k} observations. More flexible exponentially modified Gaussian type models should absorb rather than discard such events in future work 19 ; 11 . In what follows, N_{\mathrm{ev}} denotes the number of retained events after this \tau_{k}\geq 0 filter. 2.2 Partial System Matrix and Detected Activity Let \mathcal{C} denote the feasible set of detector-time channels implied by the scanner geometry, coincidence window, and TOF binning. Because the annihilation detectors are indexed in arrival order, the total number of such channels grows on the order of N_{\mathrm{sensor}}(N_{\mathrm{sensor}}-1)N_{\xi} , which becomes computationally expensive even in sparse form. In finite acquisitions, however, only a small subset of these channels is observed. We therefore restrict the forward model to the observed channels, \mathcal{C}_{+}=\{c_{k}:k=1,\dots,N_{\mathrm{ev}}\}=\{c\in\mathcal{C}:y_{c}>0\}, (5) where y_{c} is the count in channel c . The partial system matrix is defined as the conditional distribution H_{c,j}=P(C=c\mid J=j,\;C\in\mathcal{C}_{+},J\in\mathcal{V}),\qquad c\in\mathcal{C}_{+}, (6) where \mathcal{V}=\{1,\ldots,N_{\mathrm{vox}}\} is the voxel domain. Under this definition, \sum_{c\in\mathcal{C}_{+}}H_{c,j}=1,\qquad j\in\mathcal{V}, (7) so H is normalized over observed detector-time channels for each voxel. Under this parameterization, \tilde{f}_{j} denotes the expected number of retained detected triple coincidences contributed by voxel j during the acquisition. The observed channel counts satisfy the Poisson model y_{c}\sim\mathrm{Poisson}(\hat{y}_{c}),\qquad\hat{y}_{c}=\sum_{j=1}^{N_{\mathrm{vox}}}H_{c,j}\,\tilde{f}_{j},\qquad c\in\mathcal{C}_{+}. (8) Thus, activity reconstruction in this manuscript targets the distribution of retained detected triples, not the latent total number of all