A data-driven approach for spatial alignment of PET add-on detectors with sub-millimeter accuracy.

The integration of high-resolution add-on detectors with clinical PET scanners has the potential to significantly enhance imaging resolution and sensitivity. Precise geometric alignment of these auxiliary detectors is critical to fully realize their benefits. The objective of this work is to develop a novel data-driven methodology for the precise geometric alignment of high-resolution auxiliary detectors with clinical PET scanners, relying solely on point source measurements.We modeled the add-on detector as a rigid body with six degrees of freedom (three translations, three rotations). The alignment process involved estimating the unknown pose of the add-on detector by minimizing angular deviations in the lines of response, assuming collinearity among the estimated point source position, the known scanner crystal, and the unknown add-on detector crystal involved in each coincidence event. To improve interaction localization within the crystals, an effective depth-of-interaction model was applied. The method was validated in two stages. First, Monte Carlo simulations enabled direct comparison between the predefined and estimated geometries. Six detector configurations were evaluated-two in their ideal poses and four with realistic positional perturbations-to test the method's robustness. Second, experimental studies on a clinical PET scanner used a point source placed at 54 locations that follow a structured grid alongside multiple line-source setups (both parallel and arbitrary orientations). Images were reconstructed using both the known native scanner geometry, serving as a reference, and the estimated geometry, allowing quantitative evaluation of alignment accuracy by measuring centroid deviations for point sources as well as centroid and angular deviations for line sources.In simulation, the alignment error between true and estimated geometries was 0.23 ± 0.10 mm and 0.20 ± 0.07 mm for the two ideal configurations; across the four perturbed setups, mean deviations remained below 0.3 mm. In experimental validation, the centroids of all point sources in PET images reconstructed using coincidence events from the native scanner alone, between the add-on detectors alone, or between the native scanner and add-on detectors, exhibited deviations consistently below 0.5 mm across the field of view with the estimated system geometry. For line source data, the angular deviations were consistently less than 1 degree, and the centroid deviations remained below 1 mm across all slices.The results demonstrate sub-millimeter accuracy for geometric alignment of high-resolution auxiliary detectors with a clinical PET/CT scanner using our novel data-driven method. The accuracy and simplicity of this approach, combined with not relying on physical phantoms, hold significant promises for improving image quality and diagnostic accuracy of the PET systems with auxiliary devices.