Feature Space-Guided Denoising of Noisy 4D Data: Applications to Dynamic PET Imaging and Dual-Calibrated Functional MRI

4D neuroimaging methods, including dynamic PET and functional MRI, capture the spatiotemporal behaviour of physiological processes. Subsequent analysis, such as kinetic modeling of dynamic PET data, can provide parametric images related to unique aspects of physiology. However, 4D data is often exceptionally noisy, thus requiring post-processing denoising for reliable quantitative results. Several proposed 4D denoising algorithms reduce noise via signal averaging of physiologically similar voxels, but often have the trade-off of reduced accuracy. Furthermore, many do not fully exploit non-local voxel similarity, due to computational constraints and/or a suboptimal heuristic for identifying physiological similarity. In this work, we propose a denoising framework that uses a low-dimensional feature space representation of the data to identify similar voxels. This permits targeted non-local denoising to produce an initial denoised product, which is then passed to the HighlY constrained backPRojection (HYPR) algorithm to ensure data consistency. We optimize the feature space for data from different PET tracers and demonstrate cross-modality applicability, using dual-calibrated fMRI as a proof-of-concept. Additionally, we show our proposed method is superior to comparable denoising algorithms in terms of quantitative accuracy and precision of parametric images computed from the denoised data—thus demonstrating improvements relevant for research and clinical applications.