Unraveling The Connectome: Visualizing and Abstracting Large-Scale Connectomics Data

Abstract

We explore visualization and abstraction approaches to represent neuronal data. Neuroscientists acquire electron microscopy volumes to reconstruct a complete wiring diagram of the neurons in the brain, called the connectome. This will be crucial to understanding brains and their development. However, the resulting data is complex and large, posing a big challenge to existing visualization techniques in terms of clarity and scalability. We describe solutions to tackle the problems of scalability and cluttered presentation. We first show how a query-guided interactive approach to visual exploration can reduce the clutter and help neuroscientists explore their data dynamically. We use a knowledge-based query algebra that facilitates the interactive creation of queries. This allows neuroscientists to pose domain-specific questions related to their research. Simple queries can be combined to form complex queries to answer more sophisticated questions. We then show how visual abstractions from 3D to 2D can significantly reduce the visual clutter and add clarity to the visualization so that scientists can focus more on the analysis. We abstract the topology of 3D neurons into a multi-scale, relative distance-preserving subway map visualization that allows scientists to interactively explore the morphological and connectivity features of neuronal cells. We then focus on the process of acquisition, where neuroscientists segment electron microscopy images to reconstruct neurons. The segmentation process of such data is tedious, time-intensive, and usually performed using a diverse set of tools. We present a novel web-based visualization system for tracking the state, progress, and evolution of segmentation data in neuroscience. Our multi-user system seamlessly integrates a diverse set of tools. Our system provides support for the management, provenance, accountability, and auditing of large-scale segmentations. Finally, we present a novel architecture to render very large volumes interactively. We focus on two aspects: (1) Segmented objects are often toggled on and off by an interactive query, which makes it unfeasible to pre-compute a well-adapted space subdivision. (2) To scale to large data, culling and empty-space skipping must scale with the output size instead of the input volume. Our approach combines the advantages of object- and image-order stages of the empty-space skipping process.