%0 Journal Article %F FNC12:GraphTracking %A Choromanska, Anna %A Chang, Shih-Fu %A Yuste, Rafael %T Automatic reconstruction of neural morphologies with multi-scale graph-based tracking %J Frontiers in Neural Circuits %V 6 %X Neurons have complicated axonal and dendritic morphologies and their peculiar structures probably reflect functional differences and thus have been traditionally used to classify neurons into different classes. Because of this, reconstruction of neural morphologies is an important step towards understanding the structure of the brain circuits. Manual reconstructions of 3D neural structure from image stacks obtained using confocal or bright-field microscopy are time-consuming and partly subjective, and, also given the large number and variety of neuronal cell types, it appears essential to develop automatic or semi-automatic reconstruction algorithms. Nevertheless, despite the fast development of new techniques in data acquisition and image processing, automatic reconstructions still remain a challenge. In this paper we present a novel and fast method for tracking neural morphologies in 3D space with simultaneous detection of branching processes. The method exploits some existing procedures and adds to them the machine vision technique of multiscaling. Specifically, the algorithm starts from a seed point and tracks the structure using a ball of a variable radius. In each step the algorithm moves the ball center to the new point on the ball's surface with the shortest Dijkstra path. It detects the presence of the branching point by examining the spatial spread of points on the surface of the ball. The algorithm scales the ball size until branches are well separated and then continues tracking each branch. We evaluate the performance of our algorithm on synthetic data stacks obtained by manual reconstructions of neural cells, corrupted with different levels of noise. Additionally, we report results on real data sets. Our proposed algorithm is able to reconstruct 3D neural morphology that is highly similar to the ground truth and simultaneously achieves 90% average precision and 81% average recall in branching region detection %U http://www.ee.columbia.edu/ln/dvmm/publications/12/FNCIR_GraphTracking.pdf %D 2012