A flexible image smoothing and segmenting prior to parametric estimation

PET images suffer from statistical fluctuations where pixels representing the same tissue present different amplitudes, especially in short frames acquired in dynamic studies. Another difficulty is the time consuming in fitting images pixel by pixel in order to obtain parametric images. In the present work, a flexible method based on spatial and temporal pixel variance to compute parametric images is reported. For fluorodeoxyglucose and [¹⁵O]-labeled water brain studies, a template image is obtained using: (1) summed frames, (2) thresholds to exclude background, (3) segmentation based on coefficients of variation and correlation coefficients of neighboring pixels, and (4) parameter estimation by dynamic fitting (DYN) or autoradiographic (ARG) method. Visually better images and images of parameters other than regional cerebral metabolic rates for glucose (rCMRGlc) or regional cerebral blood flow (rCBF) are obtained. For comparison, rCMRGlc and rCBF in both DYN and ARG methods estimated from segmented and usual images are compared. The maximal relative error is found to be 4% for ARG rCMRGlc, 10% for DYN rCMRGlc, and 17% for DYN rCBF, while the F-test shows no difference between values estimated from segmented and usual images. This technique allows more accurate parameter estimation in a reduced computation time