
Automated large-scale reconstruction of synaptic-resolution neural wiring diagrams from volume EM data
Description
Synaptic-level dense mapping of neural circuits currently requires tracing neurons within volume EM datasets. When done manually, this process is laborious, error-prone, and hard to scale up to keep pace with the increasing size and number of volumes available for study. Automated methods to perform the tracing are therefore necessary. Within the Connectomics at Google team, we developed a segmentation technique called Flood-Filling Networks based on a recurrent convolutional neural network, which has established a new state of the art for segmentation of block-face volume EM data, reaching error-free path lengths of more than 1 mm. We have successfully used FFNs to reconstruct brain tissue of various organisms ranging from drosophila melanogaster to homo sapiens. I will discuss how the method works, provide an overview of the available reconstructions, and discuss the remaining challenges for automated analysis of volume EM data of brain tissue.