Full mapping of neural connections in brains ("connectomics") will reveal fundamental principles of organization and facilitate advances in neuroscience and neural computation. These wiring diagrams and cellular maps of circuits suggest novel hypotheses about how neurons generate behavior and provide fundamental explanations for specific neural computations that cannot be inferred in any other way. However, until recently, exhaustive anatomical brain reconstructions were limited to the tiniest volumes and to the few labs capable of the extraordinary investment of human effort required. Recent technologies are dramatically improving the size, speed and integrity of brain imaging but require substantial resources in both hardware and software. Existing large scale brain mapping efforts are producing data on standard laboratory animals (mice, flies, and worms) limiting neuroscientists’ ability to explore nature’s range of choices for brain function. Comparative studies across diverse phylogeny will require numerous mappings, across species, during development, and/or, across genetic variations, with each mapping requiring petabytes or terabytes of data to be imaged, analyzed, and stored. To realize these studies, we need an exascale computational pipeline to analyze nanometer-resolution serial-section electron microscopy (EM) data and generate large scale mappings for multiple specimens.

An exascale pipeline requires scalable computational tools for analysis of EM data,

including machine vision and machine learning, for stitching, alignment, segmentation

and tracing of neuronal structures in the extremely large 3D data sets. Real-time feedback from analysis of the brain datasets would increase efficiency of data collection and may improve quality. Current attempts at reconstructing brain wiring have already led to valuable insights, and integration of exascale computing to these efforts will advance significantly programs established by the NIH, NSF, and other federal agencies.

This project will establish and refine the computational methods needed for fast routine brain mapping, building an end-to-end processing pipeline for extracting large scale connectomic information from nanometer-resolution serial-section EM image data sets of brain samples, providing huge benefits to the brain science and neuroimaging communities.