PhysiCal-PET: Physics-driven neural networks for the rapid position Calibration of PET Scanners

Abstract

Depth-of-interaction (DOI) information is essential for mitigating parallax error and preserving uniform spatial resolution in high-resolution and long-axial-field-of-view positron emission tomography (PET). Conventional DOI calibration relies on mechanically scanned collimated sources or detector-specific bench measurements, which become impractical as the number of detector modules grows. We present PhysiCal-PET, a physics-driven neural-network framework for fast, label-free DOI calibration of fully assembled PET scanners. Coincidences are acquired from an uncollimated point source placed at a few known positions, and a network mapping DOI-sensitive detector signals to interaction depth is trained by minimizing the distance between the known source position and the line of response (LOR) defined by the two predicted interaction positions, so that the annihilation geometry alone provides supervision, without measured DOI labels. In realistic simulations, the DOI mean absolute error decreased in parallel with the geometric loss and converged to 0.98 mm, confirming that the loss is a faithful surrogate for the unobservable DOI error. Applied to a preclinical scanner with 144 pixelated dual-ended-readout detector modules, the method calibrated the entire system from 11 manually placed source positions in 55 min of acquisition, reducing the trimmed source-to-LOR distance by 24%, improving the radial point-source FWHM from 4.99 to 2.90 mm (42%), and raising the Derenzo peak-to-valley ratio of the 1.7 mm rods from 1.28 to 1.88 while separating the 1.2 mm rods, which remain merged without calibration. Requiring no collimation, motion stages, or reference measurements, and scaling favorably with system sensitivity, PhysiCal-PET offers a practical route toward autonomous, routine DOI calibration of fully assembled PET scanners with many detectors.