Segmenting neural structures in electron microscopic (EM) images is a crucial step in the acquisition and analysis of connectomes, i.e. maps of neural connections. Segmentation of such images is a very laborious task for humans, and it appears to be  a very difficult computational problem for which algorithms  and methods are still in their infancy. 

 

The talk will discuss a spectral approach based on the 'random walker'algorithm, which reduces graph segmentation with prior information to a small number of linear systems. We will show that an adaptation of the algorithm with a fair amount of 'engineering' yields a semi-automated method that finds good segmentations in only a fraction of the time required by a human. 

 

The problem is a natural candidate for spectral approaches: other ideas and open discussion during the talk are particularly welcome.