Artificial Intelligence Across the PET Reconstruction Pipeline: An Update

PET reconstruction has evolved from analytic and iterative physics-based methods toward hybrid approaches that increasingly incorporate artificial intelligence (AI). As digital detectors, long-axial FOV systems, and time-of-flight technology increase data richness and computational demand, AI, particularly deep learning, is being used across the acquisition, correction, reconstruction, and postprocessing stages to stabilize low-count imaging, refine system modeling, and improve noise-resolution tradeoffs while preserving clinically relevant image interpretation. In this context, most AI methods function to augment-rather than replace-established physics-based reconstruction frameworks. Early clinical and multicenter studies have demonstrated that selected AI-based methods maintain noninferior diagnostic performance and key quantitative metrics within defined acquisition and reconstruction contexts. As these tools transition from research into routine practice, their implementation is shaped by intended-use validation, interoperability, traceability, and software lifecycle management requirements. This Review explores the application of AI across the PET reconstruction pipeline, discussing technical foundations, highlighting key clinical implications, and considering regulatory and other practical issues that govern safe and reproducible deployment. Reconstruction pipeline steps considered in the article include mathematical inversion of projection data into images, signal processing during acquisition, prereconstruction corrections, system modeling incorporated into reconstruction, and postreconstruction processing steps that directly influence reconstructed image properties.