Whole body dynamic PET kernel reconstruction using nonnegative matrix factorization features

Abstract

The kernel reconstruction is a method that reduces noise in dynamic positron emission tomography (PET) by exploiting spatial correlations in the PET image. Although this method works well for large anatomical regions with relatively slow kinetics, whole body PET reconstruction with the kernel method can produce suboptimal results in regions with fast kinetics and high contrast. In this work we propose a new design of the spatial kernel matrix to improve reconstruction in fast and slow kinetics body regions. We calculate voxels features using nonnegative matrix factorization (NMF) with optimal rank selection. These features are then used to calculate similarities between voxels considering relative differences between features to adapt to a wide range of activity levels. Simulations and whole body mouse scans of high temporal resolution [18F]SynVesT-1, low dose [11C]raclopride, and [18F]Fallypride were performed to assess the performance of the method in different settings. In simulations, bias vs variance tradeoff and contrast was improved using the NMF kernel matrix, compared with the original kernel method. In real data, fast kinetic regions such as the heart, veins and kidneys presented oversmoothing or artifacts with the original kernel method. Our proposed method did not present these effects, while reducing noise. Brain kinetic modeling parametric maps with image derived input function ([18F]SynVesT-1) and with reference region ([11C]raclopride and [18F]Fallypride) also had lower standard error using the proposed kernel matrix compared with other methods. The NMF kernel reconstruction reduces noise and maintains high contrast in whole body PET imaging, outperforming the traditional kernel method.