The properties of functional brain networks strongly depend on how their nodes are chosen. Commonly, nodes are defined by Regions of Interest (ROIs), predetermined groupings of fMRI measurement voxels. Earlier, we demonstrated that the functional homogeneity of ROIs, captured by their spatial consistency , varies widely across ROIs in commonly used brain atlases. Here, we ask how ROIs behave as nodes of dynamic brain networks. To this end, we use two measures: spatiotemporal consistency measures changes in spatial consistency across time and network turnover quantifies the changes in the local network structure around an ROI. We find that spatial consistency varies non-uniformly in space and time, which is reflected in the variation of spatiotemporal consistency across ROIs. Furthermore, we see time-dependent changes in the network neighborhoods of the ROIs, reflected in high network turnover. Network turnover is nonuniformly distributed across ROIs: ROIs with high spatiotemporal consistency have low network turnover. Finally, we reveal that there is rich voxel-level correlation structure inside ROIs. Because the internal structure and the connectivity of ROIs vary in time, the common approach of using static node definitions may be surprisingly inaccurate. Therefore, network neuroscience would greatly benefit from node definition strategies tailored for dynamical networks. Author Summary Regions of Interest (ROIs) are often used as the nodes of functional brain networks. ROIs consist of several fMRI measurement voxels that are assumed to be functionally homogeneous, that is, behave similarly. Earlier, we showed that the assumption of similar voxel dynamics is not always true: functional homogeneity varies widely across ROIs. In this paper, we demonstrate that functional homogeneity changes in time. These changes are connected to changes in local network structure around ROIs, which suggests that an ROI’s functional homogeneity may reflect its role in the network. Finally, we show that there is rich, time-dependent structure of voxel-level connectivity inside ROIs. This leads us to ask if the dynamic brain networks can be described by any set of static ROIs.