BRAIN CONNECTS: A Scalable Automated Proofreading Framework for Connectomics

Summary Generating synapse-resolution maps or connectomes of the brain are crucial to understanding the neural basis of behavior, and can provide key insights into the onset, progression, and treatment of neurological disease and injury. Towards this goal, major advances in electron microscopy imaging and automated image segmentation have enabled researchers to produce millimeter-scale connectomic datasets, and forge a path towards an even larger whole mouse brain volume. Despite the high quality of automated segmentation at this scale, the enormous extent of axon and dendrite “wiring” in the brain unavoidably leads to errors in neuronal connectivity that require correction with post-hoc proofreading. Although a variety of approaches have been developed to enable faster manual proofreading, the number of human hours needed to correct errors are prohibitive, and prevent us from realizing the full potential of valuable datasets such as the cubic millimeter MICrONS (mouse cortex) and H01 (human cortex) volumes. To enable even larger connectomes, we must develop cost-effective and time-saving automated methods to replace labor-intensive human proofreading where possible and allow human resources to focus on other connectomic tasks that include generating training data and validating automated correction. The goal of this proposal is to build capabilities for scalable automated proofreading, leveraging and extending software tools built during our previous IARPA MICrONS activities: NEURD (short for NEURal Decomposition), an automated error detection and correction framework built by Baylor College of Medicine, and NeuVue, a scalable manual proofreading platform built by the Johns Hopkins University Applied Physics Laboratory. Both tools are deeply integrated and complementary to the existing ecosystem of open- source connectomics tools and resources from the community such as Neuroglancer, PyChunkedGraph (PCG), and Connectomics Annotation Versioning Engine (CAVE). Building on the foundation of these tools and our existing collaboration, we will add capabilities for machine learning enabled error detection of a wider range of error types including both merge and split errors. We will implement an active learning approach that focuses valuable human validation effort on the most informative error examples, with the goal of statistically validating entire classes of edits that can be applied in automated batches to the segmentation. Finally, we will develop a workflow for applying these automated edits to the segmentation in an optimized way that also does not conflict with existing manual proofreading and retains a confidence metric for each edit that can be used for downstream analysis. The successful completion of this project, “Connects-Proof: A Scalable Automated Proofreading Framework for Connectomics” will yield a mature workflow that is validated across multiple data sets and that can support existing and future work in the BRAIN-CONNECTS program.