Mapping the Brain

APL is part of a collaborative effort to build and validate a process for accelerating the acquisition of high-fidelity electron microscopy data using in-process artificial intelligence. Traditional workflows for generating a high-resolution connectome first involve generating a three-dimensional image of a portion of the brain. The generation of these images typically involves spending equal time on every part of the image, regardless of where the most important structures are. Our collaborative team is working on a more efficient approach, where a quick, low-resolution scan can identify areas of the image that are most important to point the microscope—by more intelligently targeting the time-intensive high-resolution scanning to the parts of the image that are most essential for generating a brain map.

Our team is focused on operationalization of connectomics data generation, processing, validation, and analysis. We have invested in the development of scalable data pipeline capabilities to allow for rapid data processing and analysis, as well as capabilities to enable integration of community tools and ensure FAIR (findable, accessible, interoperable, reusable) data, including new standards for image data, metadata, and annotations.

In partnership with researchers from Baylor College of Medicine, Allen Institute for Brain Science, and Princeton University, a team from APL created NEURal Decomposition (NEURD), a software package that automates the proofreading of neural datasets.

Brain maps are generated by identifying, or segmenting, the individual neurons within high-resolution scans of neural tissue. This process, often performed by artificial intelligence-based algorithms, can lead to errors because AI can mistakenly merge or split neurons. To date, correcting these errors has largely been performed by hand by neuroscience experts. NEURD enables those corrections to happen more quickly and more consistently than is possible with manual processes, making the overall process of generating detailed brain maps much faster and cost-efficient.

To efficiently review and confirm the accuracy of the edits generated by NEURD, researchers at APL developed NeuVue. NeuVue provides an intuitive visualization and queueing system that bridges the gap between automated and human proofreading in connectomics. Ultimately, the goal is to scale the technology and expand capabilities to study a full human brain. This future step will not only bring deeper understanding of how brains are wired but also allow for studying brain development and disease.

In addition to building and integrating tools to facilitate the construction of connectomes, our team is leveraging connectomic data to better understand neural circuits and their function. In one recent study, our team characterized a new feature of neural circuits, a “convergent motif,” where many neurons synapse with a target neuron in very close proximity to each other. This newly understood circuit feature could help explain how learning and memory occurs in the fruit fly brain. This analysis was powered by APL’s DotMotif tool for efficiently searching for patterns in brain maps.

Even after the completion of the MICrONS program, APL remains interested in uncovering how biological brains are able to reason and learn so effectively and efficiently. By analyzing state-of-the-art maps of the fruit fly brain, our team is uncovering new model architectures that can enable efficient continual learning.