Rapid and Cost‐effective Connectomics with Intelligent Image Acquisition, Reconstruction, and Querying

High-throughput connectomics is needed to generate the TB-, PB- and EB-scale wiring diagrams of mammalian brains, but is limited to the few research institutes (e.g., Janelia, Allen, Max Planck) with sufficient infrastructure. As resource-rich as these institutes are, none are able to do a whole brain at nanometer scale on their own. The failure to broaden participation to a larger community is an obstacle to scaling connectomics. We propose a new and more affordable imaging strategy that will allow many more teams to engage in connectomics. High-speed electron microscopes for connectomics – e.g., multibeam SEMs – are rare and prohibitively ex- pensive. More common single-beam SEMs have sufficiently high spatial resolution, but are prohibitively slow for connectomics. We plan to increase the speed of single-beam SEM systems to the speed of multibeam SEMs without substantially increasing cost. Our strategy adds artificial intelligence to SEM architecture to re- duce the number and dwell time of pixels that need to be imaged at high-resolution without adversely affecting “segmentability”. With new software and standard computer hardware, we can turn single-beam SEMs into intel- ligent, powerful devices at negligible cost. We demonstrated a proof-of-concept of a smart scanning system that we engineered into a single-beam SEM. The modified SEM acquires a low-resolution/low-dwell time image of a brain slice at high speed. It then uses ultrafast ML algorithms to extract most of the wiring from these images, while at the same time identifying in real time those salient pixels that should be rescanned to improve signal-to noise in the final wiring diagram. We have achieved >10-fold speedup in image acquisition, and plan to increase the rate significantly more. A significant scale-up in the rate of connectomics demands comparable improvements in image processing (stitching, alignment, and segmentation). We have built computationally more efficient methods for aligning and segmenting connectome datasets. We will integrate these methods into a cloud-based platform that will allow researchers without significant computational infrastructure or expertise to process connectomics datasets. All data products and capabilities will be publicly accessible through BossDB. In summary, this integrated research program will scale connectomics to a much larger neuroscience community.