Multi-GPU parallel framework for DOI-enabled long axial field-of-view PET image reconstruction

Objective.Depth-of-interaction (DOI)-enabled long axial field-of-view (LAFOV) PET can mitigate parallax-induced resolution degradation, but DOI modeling substantially expands the lines-of-response (LOR) space and increases computational cost. This study aims to develop a scalable reconstruction framework for DOI-enabled LAFOV-PET.Approach.We present an optimized multi-GPU reconstruction framework based on list-mode ordered-subsets expectation maximization (LM-OSEM). The framework targets the major computational bottlenecks of DOI-enabled LAFOV-PET reconstruction, including sensitivity image calculation and Monte Carlo-based scatter estimation.Main results.The extended IQ phantom demonstrated near-linear scalability across multiple GPUs for sensitivity image calculation and Monte Carlo-based scatter estimation. Additional Derenzo and hybrid anthropomorphic phantom studies validated the imaging benefits of DOI modeling, including improved spatial resolution and performance under realistic imaging conditions.Significance.The proposed framework improves the scalability of DOI-enabled LAFOV-PET reconstruction by enabling its dominant computational components to efficiently utilize available GPU resources, establishing a foundation for high-performance reconstruction in next-generation PET systems.