The plasma time-activity curve is often required as the input function for dynamic quantitative FDG PET studies to estimate the metabolic rate of glucose. The invasive gold standard arterial blood sampling has been suggested, however, it has many inconveniences and challenges in clinical and pre-clinical settings. Thus, the image-derived input function has been proposed to obtain the input function from dynamic images non-invasively. This method often needs a manual drawing of one or two regions of interest (ROIs), which is an operator-dependent and time-consuming task. The aim of the presented study was to capture the spatial and temporal patterns of dynamic PET images for automatic ROI extraction. Our proposed approach tries to overcome the main limitation of image clustering methods: the loss of temporal information for dynamic PET ROI definition. The experiments showed that the proposed automatic ROI method can be used for dynamic PET parameter estimation.