PURPOSE: Recent development in positron emission tomography (PET) dramatically increased the effective sensitivity by increasing the geometric coverage leading to total-body PET imaging. This encouraging breakthrough brings the hope of ultra-low dose PET imaging equivalent to transatlantic flight with the assistance of deep learning (DL)-based methods. However, conventional DL approaches face limitations in addressing the heterogeneous domain of PET imaging. This study aims to develop a wavelet-based DL method capable of restoring high-quality imaging from ultra-low-dose PET scans. MATERIALS AND METHODS: In contrast to conventional DL techniques that denoise images in the spatial domain, we introduce WaveNet, a novel approach that inputs wavelet-decomposed frequency components of PET imaging to perform denoising in the frequency domain. A dataset comprising total-body RESULTS: Our proposed WaveNet consistently outperforms the baseline UNet model across all levels of dose reduction factors (DRF), with greater improvements observed as image quality decreases. Statistical analysis (p <
0.05) and visual inspection validated the superiority of WaveNet. Moreover, WaveNet demonstrated superior generalizability when applied to two cross-scanner datasets (p <
0.05). CONCLUSION: WaveNet developed with total-body PET scanners may offer a computational-friendly and robust approach to recover image quality from ultra-low-dose PET imaging. Its adoption may enhance the reliability and clinical acceptance of DL-based dose reduction techniques.