Neuronal connectivity can be reconstructed from a 3D electron microscopy (EM) image of a brain volume. Connectomics is a modern reincarnation of neuroanatomy, aiming at densely reconstructing neurons, exhaustively detecting synapses, and extracting a complete wiring diagram from a brain volume. Nanoscale resolution of EM brain images enables reliable tracing of densely packed neuronal branches as well as unambiguous detection of chemical synapses. For the past decade, convolutional networks, one of the workhorses of deep learning, have been extensively used for 3D reconstruction of neurons from EM brain images. More recently, the use of convolutional nets have been extended to other subproblems of connectomics including (1) image registration and alignment, (2) synapse detection and partner assignment, and (3) morphological error detection and correction.
In this talk, I will present a set of deep learning algorithms based on convolutional nets for automated reconstruction of neural circuits, with particular focus on highly anisotropic images of brain tissue acquired by serial section EM (ssEM). I will demonstrate the automated neuron reconstruction pipeline developed by our lab on multiple terascale and petascale datasets from the mouse and fly brain, and conclude with remarks on the future prospect and remaining challenges of deep nanoscale connectomics.