We proposed an organ-by-organ segmentation method for whole-body rodent dynamic PET images based on our hierarchical Mumford-Shah model for vector-valued image segmentation (HMSMv). First, we extract the shape parameters of time-activity curves (TAC) of a volume of interest(VOI) in a dynamic PET image via a noise-normalized principal component analysis (PCA). Then we segment one organ at a time in two steps. At the first hierarchy, we segment a rough VOI blurred by partial volume, motion, and spillover effects from the background via a level set method. Then, we select a seed voxel that is farthest from the VOI boundary and is least affected by spillover and motion blur. Finally, we perform the second level set based segmentation inside the initial segmented VOI to refine the segmentation result by evolving a sphere centered at the seed voxel. Once an organ is segmented, it is removed from the PET image by assigning zero value to the voxels inside the VOI and the segmentation algorithm will move on to another organ. The segmentation algorithm has only one regularization parameter, which is easy to choose. Segmentation results of computer simulated data of the MOBY mouse phantom and real mouse dynamic PET images are shown.