Segmentation can be achieved by the detection of troughs of density separating the usually compact domains, or by taking advantage of differences in the electron scattering density of components (e.g., RNA versus protein), or differences in texture. The first kind of separation is achieved by some version of density thresholding (section 5.2.1); the other kind can be achieved by employing either of two image-processing techniques that have been recently introduced into the field of EM: the watershed transform and level set related techniques (Sethian, 1996). The third property of the density map that can be exploited for segmentation is texture. Although its applications may be mainly confined to low-resolution density maps obtained by electron tomography, a texture-sensitive method, employing an eigenvalue analysis of the so-called affinity matrix, is described below as well.
Low-resolution density maps are by nature fuzzy—exact definitions of boundaries are lacking, and it is only "after the fact,'' when an atomic structure becomes available, that the actual physical boundaries can be ascertained. Some newer approaches to segmentation, so far only tried in computerized axial tomography-scan or magnetic resonance imaging, start from a representation of a fuzzy 3D image, and the extent to which its elements hang together are expressed by a fuzzy connectedness (see Udupa and Samarasekera, 1996). It may be worth trying these newer methods in 3DEM, as well.
Was this article helpful?