A design and implementation of an efficient, parallel watershed algorithm for affinity graphs

Abstract

In this thesis, I designed and implemented an efficient, parallel, generalized watershed algorithm for hierarchical segmentation of affinity graphs. By introducing four variable parameters the algorithm enables us to use previous knowledge about the input graph in order to achieve better results. The algorithm is very suitable for hierarchical segmentintation of large scale 3D images of the brain tissue obtained by electron microscopy making it an essential tool for reconstructing the brain's neural-networks called connectomes. The algorithm was fully implemented in C++ and tested on a currently largest available affinity graph of size 90GB on which no existent watershed implementation could be applied.