Innovations in clinical PET image reconstruction: advances in Bayesian penalized likelihood algorithm and deep learning

Recent advances in PET image reconstruction have focused on achieving high image quality and quantitative accuracy. Bayesian penalized likelihood (BPL) algorithms, such as Q.Clear and HYPER Iterative that have been integrated into commercial PET systems offer robust image noise suppression and edge preservation through regularization. In parallel, methods based on deep learning such as SubtlePET, AiCE, uAI (R) HYPER DLR, and Precision DL have emerged primarily as post-processing techniques. They use trained convolutional neural networks to reduce image noise while preserving lesion contrast. These methods have reduced image acquisition times or reduced radiotracer doses while maintaining diagnostic confidence. uAI (R) HYPER DPR represents a hybrid approach by embedding deep learning in iterative reconstruction. This review summarizes the technical principles and the clinical performance of BPL and deep learning-based PET reconstruction algorithms, and discusses key considerations such as image quality and quantitative accuracy of PET images. This review should deepen understanding of advanced PET image reconstruction techniques and accelerate their clinical implementation across diverse PET imaging applications.