Single-scan adaptive graph filtering for dynamic PET denoising by exploring intrinsic spatio-temporal structure

The performance of Dynamic Positron Emission Tomography (PET) is often degraded by high noise levels. A key challenge is the significant variability across scans, which makes fixed denoising models suboptimal. Furthermore, current denoising algorithms are often confined to a single data domain, limiting their ability to capture deeper structural information. To overcome these limitations, we introduce a novel single-scan adaptive spatio-temporal graph filtering (ST-GF) technique. The fundamental principle is to explore the latent structure of the data by representing it in a graph-signal space. Unlike deep learning approaches requiring external training data, our algorithm works directly on a single acquisition. It maps the noisy sinogram to a graph to reveal its underlying spatio-temporal structure-such as spatial similarities and temporal correlation-that is obscured by noise in the original domain. The core of the framework is an iterative process where a graph filter is adaptively constructed based on this latent structure. This ensures the denoising operation is precisely tailored to the unique characteristics of the single scan being processed, effectively separating the true signal from noise. Experiments on simulated and <i>in vivo</i> datasets show our approach delivers superior performance. By leveraging the latent structure found exclusively within each scan, our method operates without prior training and remains immune to potential biases or interference from irrelevant external data. This self-contained approach grants ST-GF high robustness and flexibility, highlighting its substantial potential for practical applications.