Static Segmentations in Dynamic PET Images: The need for a new method

The Task Group 211 report of the American Association of Physicists in Medicine (AAPM) reviewed static segmentation techniques in nuclear positronemission tomography (PET) imaging used in nuclear medicine. These methods, when applied to a dynamic image, such as the ones obtained in pharmacokinetic analyses, fail to take into account the dynamic nature of the acquisitions. In this article, the leading hypothesis was that a static segmentation was not adequate in even the simplest dynamic PET images. To put this idea forward, a simple dynamic PET phantom was devised. Many dynamic acquisitions were obtained using FDG. To analyze them, different static segmentations were performed on each timeframe. These were followed by quantitative analyses to determine whether the segmentations were consistant between various timeframes of reference. The quantitative analytical tools used were the S{\o}rensen-Dice coefficients, the overlapping of the time-activity curves (TACs), and the pharmacokinetic parameters extracted from the images using the Dynesty Python package. In order to perform some of the analyses, an uncertainty had to be added to the TACs themselves: to do so, the individual segmentations were spatially displaced to estimate the sensibility of the TAC to the underlying segmentation. Using these analytical tools, we propose that static segmentations are not sufficient tools for segmenting dynamic images in a nuclear medicine context. The specific case of pharmacokinetic modelling is used to exemplify this idea. Further work could include a method of estimating uncertainties on segmentations or a novel method for dynamic segmentations in dynamic PET images.
Comment: 15 pages, 5 figures