RT Journal Article SR Electronic T1 Revisiting noise models in small-animal PET JF Journal of Nuclear Medicine JO J Nucl Med FD Society of Nuclear Medicine SP 2601 OP 2601 VO 56 IS supplement 3 A1 Mansoor, Awais A1 Casas, Rafael YR 2015 UL http://jnm.snmjournals.org/content/56/supplement_3/2601.abstract AB 2601 Objectives Understand the acquisition process of PET in small-animal imaging and sources of potential noiseMethods Although considerable work has been done in the denoising of clinical PET data, the literature somewhat lacks similar studies in preclinical small-animal imaging. Under standard illumination, thermo-electronic fluctuations are much stronger than the photon-counting process that motivates the additive white Gaussian noise modeling assumption of the noise in most modalities. However, in low-photon-counting applications such as PET, noise is strongly signal dependent, motivating the Poisson assumption for noise. However, complete independence of the Poisson noise model does not bode well with the image data that posses strong depedence amongst neighboring voxels. Image intensity may therefore be modeled as Poisson distributed and potentially degraded by AWGN (Gaussian-Poisson distributed). Since Gaussian-Poisson noise model has the flexibilty to accommodate both neighborhood dependence as well as counting noise. We present an overview of the existing statistical model and the comparative analysis of different noise models in small-animal PET denoising including Gaussian noise model, Poisson noise model, and Gaussian-Poisson noise model.Results 20 PET scans are collected. A field-of-view of 4.8 cm and 6.7 cm in diameter is scanned with a time coincidence window of 10ns and energy windows of 250-700 keV. Noise statistics of the acquired PET data is empirically estimated. The noise models are evaluated using the performance of individual variance stabilizing tranform (VST) (Fig. 1). It has been demonstrated based on the results that the Gaussian-Poisson noise model fits best the noise present in the acquired data. Further, the clinical comparisons of the models is done in terms of the change in SUV as well as contrast and edge preservation.Conclusions In this work, we rigorously evaluated most commonly used noise models for small-animal PET data, we found that Poisson-Gaussian model fits best the noise characteristics in small animal PET.