Self-supervised Noise-aware Network for Dynamic Rubidium-82 Cardiac PET Image Denoising

PET is a functional imaging modality widely used in different clinical settings. Since PET scans are accompanied by substantial radiation exposure, reducing the radiation dose in PET scans is important. However, low-dose PET scans usually result in high image noise, which can negatively affect image quality and diagnostic performance. In addition, tracer decay in dynamic PET images results in continuously changing noise levels across different dynamic frames, especially for fastdecaying tracers such as Rubidium-82. Recent advances in deep learning showed great potential for low-dose PET denoising. However, most previously-proposed supervised methods cannot be adapted for this case since it is not possible to obtain paired labels for every single dynamic frame. Also, previously-proposed unsupervised methods were only developed for single noise-level images, and therefore cannot be adapted for dynamic cardiac PET scans due to changing noise levels across different frames. To address these challenges, we propose a self-supervised noise-aware method for dynamic Rubidium-82 cardiac PET image denoising. Validated on 10 patient studies obtained on a Siemens Biograph mCT PET/CT system, the proposed method showed promising denoising performance across all the dynamic frames.