Speaker
Description
The recent advent of Total-Body PET (TB-PET) represents a transformative leap in molecular imaging, offering an effective sensitivity gain of over 40-fold and enabling full-body dynamic tracking [1, 2]. To maintain uniform spatial resolution across such extended fields of view, precise Depth-of-Interaction (DOI) capability is crucial to mitigate severe parallax errors. Advanced light-sharing detector designs ranging from multi-layer finely segmented arrays to fully monolithic scintillation crystals provide excellent intrinsic 3D positioning capabilities to meet this demand [3]. However, the translation of these high-performance architectures into large-scale clinical TB-PET systems is severely hindered by a critical bottleneck: the prohibitive complexity of their positioning calibration. Traditional 3D calibration relies on meticulous, labor-intensive benchtop setups using mechanically scanned collimated gamma-ray sources [4].
While recent in-system "virtual collimation" methods have shown promise in scaling this process for fully assembled scanners [5, 6], they fundamentally rely on pre-existing, accurate position estimates in a partner 'collimating' detector to define a precise Line-of-Response (LOR). In a newly assembled system where all detectors are uncalibrated, this creates a circular dependency. Relying on uncalibrated analytical methods for initial estimates propagates significant spatial uncertainties, preventing true de novo calibration [7].
To overcome this fundamental barrier, we present a novel, autonomous in-system calibration methodology driven by Physics-Informed Neural Networks (PINNs) [8]. Rather than relying on physical collimators or imprecise analytical pre-positioning, our approach leverages the fundamental physics of positron annihilation as an implicit, powerful training constraint . By placing a simple, uncollimated source at various locations within the scanner’s field of view, the neural network learns to directly map raw electronic detector signals to 3D interaction coordinates based purely on the physical laws dictating that the coincident gamma-rays must geometrically intersect the known source position.
In this work, we present the evolution of this methodology, beginning with an ideal proof-of-concept case. Using simplified Monte Carlo simulations restricted to single-point photoelectric absorptions, we demonstrate the baseline capability of the physics-constrained network to achieve highly accurate 3D positioning (sub mm) from scratch, without any prior analytical pre-positioning or ground-truth label generation.
Following this ideal case, we detail the progressive introduction of realistic detector phenomena and the necessary adaptations to our AI framework. We transition to high-fidelity Geant4 simulations of light-sharing scintillators, introducing multi-interaction events such as Compton scattering.
By detailing this progression from idealized simulations to handling the noisy, multi-variate reality of physical detectors, we demonstrate a clear pathway toward experimental validation. Ultimately, this physics-based deep learning framework promises to replace weeks of mechanical calibration with rapid, autonomous de novo calibration and routine fine-tuning, unlocking the cost-effective scaling required for the future of TB-PET and advanced multimodality integrated systems.
References:
[1] Vandenberghe., et al. EJNMMI Phys 7 (1) (2020): 35.
[2] Spencer, Benjamin A., et al. J Nucl Med 62 (6) (2021): 861-870.
[3] Gonzalez-Montoro, Andrea, et al. IEEE TRPMS 5 (3) (2021).
[4] Berg, Eric, and Simon R. Cherry. Semin Nucl Med 48 (4) (2018).
[5] Bruyndonckx, Peter, et al. NIM-A 571 (1-2) (2007).
[6] Gonzalez-Montoro, Andrea, et al. IEEE TRPMS 5 (6) (2020).
[7] Kuhl, Yannick, et al. Med Phys 52 (1) (2025): 232-245.
[8] Cai, Shengze, et al. Acta Mech Sin 37 (12) (2021): 1727-1738.
| Track | TBPET |
|---|---|
| Presentation type | Oral |