Chapter 49 Suppression of Noise Artifacts in Spectral Analysis of Dynamic PET Data1

The method of spectral analysis as originally implemented is sensitive to noise in data. This sensitivity is reflected in a distribution of peaks across the whole spectrum, but is particularly evident when the parameters to be estimated require that data to be extrapolated. An example of this is the estimation of the total volume of distribution of a tracer in the tissue. Although the method is effective at extracting the tissue impulse response function (IRF) as defined within the time course of observed data, estimates of the total volume of distribution require extrapolation of the IRF to infinite time. In some cases, although satisfactory images of the volume of distribution can be obtained at the voxel level, a “pile up” of peaks at the low frequency end of the spectrum frequently gives noisy images of the volume of distribution. The introduction of penalty functions into the non-negative least-squares (NNLS) algorithm can be used to suppress noisy artifacts, but as yet there is no systematic way of generating penalty functions suitable for routine analysis of dynamic PET data. Image quality can be improved by smoothing the dynamic images spatially or by setting the range of basis functions above the decay constant for the isotope.