Segmentation of neurons in EM images through the identification of a shared membrane component

We study computational methods for the automated or semi-automated segmentation of individual neurons in electron microscopic (EM) images, a task of crucial importance in the acquisition and analysis of connectomes, i.e. maps of the neural connections. The problem of isolating each neuron has been traditionally addressed by treating each neuron as a component on a graph which must be made distinct from every other. Because the neurons are tightly packed, adjacent cellular membranes can be conceived as a single continuous membrane component; finding this membrane component effectively yields a correct segmentation of the image into discrete neuron components. The images are pre-processed with spectral methods in such a way that the darker objects are accentuated and more easily collected into discrete components. The largest component can be assumed to include the membrane component, but organelles and clusters of vesicles can be found close enough to membranes to also be collected into this component. The main challenge in finding the membrane object is separating membranes from organelles and vesicles. By exploiting the morphological differences between membranes, whose cross sections are thin and spread out, and clusters of vesicles and organelles, which are more round and compact, we can filter out the latter to find the membrane component.