Modeling of 2D PET noise autocovariance function applied to individual activation studies

Functional brain mapping involves comparing cerebral blood flow (CBF) images corresponding to different states of activation. Due to the low sensitivity of 2D PET cameras, the weak signal-to-noise ratio (SNR) of individual “activated minus control” images hampers the detection of significant CBF changes. In the case of filtered backprojection (FBP), it is possible to include spatial filtering in the image reconstruction to decrease noise level, i.e. to choose a low reconstruction filter cutoff frequency. The authors investigated the consequences of such a choice on the detection according to a recent analysis method (J.B. Poline and B. Mazoyer, J. Cerebral Blood Flow Metab., vol. 13, p. 425-437, 1993), which focuses on the size of pixel clusters defined by thresholding a difference image. The authors modelled the noise autocovariance function (ACF) and produced sets of simulated noise images. They then estimated the probability of observing at least one supra-threshold pixel cluster of size greater than a given one, in a noise-only image of correlation structure corresponding to the authors´ model. This estimation was performed for various thresholds, and the results were compared to those obtained for a 2D gaussian autocorrelation of identical width. For low thresholds, estimated probability values according to the authors´ model were markedly inferior to those corresponding to the gaussian one; this noise modelisation might therefore improve activation detection for images yet low-frequency reconstructed