Many questions remain on the detailed morphology of convection patterns in the mantle. In particular, the existence of deep mantle plumes has been the subject of debate ever since they were proposed to explain the presence of hotspot volcanoes. With the advent of numerical methods for accurate seismic wavefield computations, it is now possible to apply the tools of waveform tomography to better detect the presence, throughout the mantle, of slow velocity anomalies, previously "hidden" by wavefront healing effects not captured by approximate wave propagation methods. Using waveform tomography based on the spectral element method (SEM) for the first time, we have recently constructed a global, radially anisotropic, shear velocity whole mantle model, which shows better focused, finer scale low velocity structures both in the upper and in the lower mantle than in any previous global tomographic models. In particular, the lower mantle structure is dominated by vertically elongated structures that form discrete "columns" rooted at the base of the mantle, positioned in the vicinity of major hotspots lying over the large lower mantle low shear velocity provinces. The vertical conduits are quite straight from the base of the mantle to 1000 km depth, but wider (500-1000 km) than expected from the standard "plume" model. Their character changes above this depth, as they seem to become narrower and meander across the upper mantle, where they appear to interact with secondary scale convection set off by plate motions. I will also discuss the significance of the apparent rheological boundary around 1000 km depth. The roots of several of these broad plumes contain large ultra-low velocity zones (ULVZs). In particular, the ULVZ under Iceland has a regular, axi-symmetric shape, which suggests the presence of partial melt in the hot center of the plume roots.