Optimal filtering strategies for task-specific functional PET imaging

Abstract

Functional Positron Emission Tomography (fPET) is an effective tool for studying dynamic processes in glucose metabolism and neurotransmitter action, providing insights into brain function and disease progression. However, optimizing signal processing to extract stimulation-specific information remains challenging. This study systematically evaluates state-of-the-art filtering techniques for fPET imaging. Forty healthy participants performed a cognitive task (Tetris®) during [18F]FDG PET/MR scans. Seven filtering techniques and multiple hyperparameters were tested: including 3D and 4D Gaussian smoothing, highly constrained backprojection (HYPR), iterative HYPR (IHYPR4D), MRI-Markov Random Field (MRI-MRF) filters, and dynamic/extended dynamic Non-Local Means (dNLM/edNLM). Filters were assessed based on test-retest reliability, task signal identifiability (temporal signal-to-noise ratio, tSNR), spatial task-based activation, and sample size calculations were assessed. Compared to 3D Gaussian smoothing, edNLM, dNLM, MRI-MRF L = 10, and IHYPR4D filters improved tSNR, while edNLM and HYPR enhanced test-retest reliability. Spatial task-based activation was enhanced by NLM filters and MRI-MRF approaches. The edNLM filter reduced the required sample size by 15.4%. Simulations supported these findings. This study highlights the strengths and limitations of fPET filtering techniques, emphasizing how hyperparamter adjustments affect outcome parameters. The edNLM filter shows promise with improved performance across all metrics, but filter selection should consider specific study objectives and resource constraints.