A system-level DOI discrimination method based on SSDA for a brain-dedicated DOI-PET scanner

OBJECTIVE: Depth-of-interaction (DOI) information is essential for improving the spatial resolution in positron emission tomography (PET) imaging. Existing deep learning-based DOI discrimination methods rely on side-irradiation measurements for each detector to obtain labeled data, making them impractical for system-level calibration. APPROACH: This study proposes a system-level DOI discrimination method based on semi-supervised domain adaptation (SSDA) for a brain-dedicated DOI-PET scanner. The method requires only side-irradiation data from a small number of detectors (even one) as labeled data, greatly reducing calibration time and labor. It was evaluated using a two-layer lutetium yttrium oxyorthosilicate (LYSO) detector array with different pixel sizes: the top layer is a 16 × 16 array of 1.53 × 1.53 × 5 mm³ crystals, and the bottom layer is an 8 × 8 array of 3 × 3 × 15 mm³ crystals. Performance was further evaluated on a PET prototype scanner comprising 72 detectors, through assessments of flood image quality and image spatial resolution of a 22Na point source positioned at different radial offsets. MAIN RESULTS: The method achieved an accuracy of 98.21% on a benchmark established using fully supervised learning, with flood image quality k of 4.80 ± 0.47 and 6.84 ± 0.53 for the top and bottom layers, and MAEk of 0.04 ± 0.03 and 0.07 ± 0.04, respectively. In the flood image quality evaluation of the 72-detector PET prototype scanner, it maintained stable and high performance comparable to fully supervised learning while using side-irradiation data from only one detector as labeled data, achieving k of 4.82 ± 0.55 and 6.65 ± 0.65 and identifying merely 0.80% misclassified crystal clusters. In the spatial resolution evaluation, the method achieved an average radial spatial resolution of 1.83 ± 0.45 mm for the two-level DOI configuration and 1.39 ± 0.20 mm for the four-level DOI configuration. SIGNIFICANCE: With its low cost, high accuracy, and strong scalability, the proposed method provides an efficient and practical solution for system-level DOI calibration in next-generation high-performance brain-dedicated PET scanners.