Efficient correction for EM connectomics with skeletal representation

Machine vision techniques for automatic neuron reconstruction from electron microscopy (EM) volumes have made tremendous advances in recent years. Nonetheless, large-scale reconstruction from teravoxels of EM volumes retains both under- and over-segmentation errors. In this paper, we present an efficient correction algorithm for EM neuron reconstruction. Each region in a 3D segmentation is represented by its skeleton. We employ deep convolutional networks to detect and correct false merge and split errors at the joints and endpoints of the skeletal representation. Our algorithm can achieve the same or close accuracy of the state-of-the-art error correction algorithm by querying only at a tiny fraction of the volume. A reduction of the search space by several orders of magnitude enables our approach to be scalable for terabyte or petabyte scale neuron reconstruction.