Research Highlights

Big Data Analysis: Applying Electron Microscopy to Semi-automated Neuron Boundary Detection and Non-branching Process Segmentation

box
Screen capture of NeRV displaying the automatic segmentation results on the C. elegans worm ventral nerve cord for a portion of the data.

January 2014 La Jolla

Reconstructing neural circuits is important to studying neural circuit connectivity and its behavioral implications. Understanding the differences between neuronal classes, patterns, and connections is critical to enable a more general understanding of how neural circuits process information.

Electron microscopy (EM) is a useful method to determine the anatomy of individual neurons and their connectivity because it has a resolution that is high enough to identify features such as synaptic contacts and gap junctions, which define connectivity and, therefore, are required to reconstruct the neural circuit. But because the complexity and size – often approaching tens of terabytes -- of this data make for very difficult and labor-intensive interpretation, new segmentation techniques for identifying neurons in these data sets are needed.

A team of researchers from the University of Utah and UCSD recently published their results describing a new method for segmentation and extraction of neurons from EM images. They developed a way to detect neuron boundaries and segment non-branching processes in the images and visualize them in three dimensions. This method combines automated segmentation with a custom graphical user interface to correct mistakes. The automated process uses machine-learning and image-processing techniques to identify neuron membranes that delineate the cells in each 2D section. To segment non-branching processes, the cell regions in each 2D section are connected in 3D by correlating regions across sections.

The combined process enables users to quickly segment cellular processes in large volumes. In this study, images were acquired using serial-section transmission electron microscopy and serial block-face scanning electron microscopy. These methods provide much higher resolution than other methods and make it possible to resolve the 3D structure and connectivity of neurons.

The method devised by NCMIR and Utah scientists to reconstruct non-branching neuron cell processes consisted of two steps. First, neuron membranes were segmented in 2D and neuronal cross-sections identified. Second, the regions were linked across all sections to form 3D renderings of parallel processes.

The initial neuron segmentation used for each 2D section builds on previous work that uses a series of artificial neural networks (ANNs) to detect neuron membranes. To improve membrane detection, that method was extended in this study by incorporating learned membranes from sequential sections into another ANN. Examining the detected membranes in sequential sections, above and below the current section, helps the classifier learn to detect membranes that are grazed or complicated by internal cellular structures. Tensor voting, a method for closing remaining gaps, was applied in a post-processing step. Also drawing from previous work, the team incorporated an optimal path algorithm to connect similar regions through the volume to form complete 3D segmentations. The result is significantly improved segmentation by comparison with the original method.

The automatic methods described above all work fairly well on their own but, in the end, require the ability to view and edit the segmentation results. The Neuron Reconstruction Viewer (NeRV), developed in this study, bridges these two requirements by providing a visual interface to large volumes of EM images and neuron segmentations, and the option of making corrections that improve segmentation.

In essence, NeRV is an interface to view raw image data and the 3D reconstruction. Interacting with the image data and the rendered neuron provides insight for the scientist on the arrangement of the neurons within the data. Users can view the membrane detection, the region segmentation, and the raw data all in one viewer. They can correct segmentations to close gaps with a simple drawing tool, then recompute the regions and correlations to improve the optimal path calculation. They can also manually select regions in slices and create their own 3D renderings with the automatic path calculation. For precomputed and segmented neurons, a separate window allows them to select different neurons for viewing, deleting, or joining.

Two EM data sets were segmented using the methods described above. The first was a stack of 400 sections from the ventral nerve cord of the C. elegans worm. The second was a stack of 400 sections from the mouse neuropil. These data sets contained very different types of neural cells and had different resolutions.

The team’s results were generated using two computers. The first was a desktop computer containing 8 x 2.8-GHz Intel CPUs and 8 gigabytes of memory. The second was a 32-node, 2.93-GHz, shared-memory computer containing 200 gigabytes of memory. The raw data, if loaded entirely into memory at one time, required 4.2 gigabytes of memory for C. elegans, vs. 25 gigabytes for the mouse neuropil. Because of these requirements, processing was distributed across computers, in parallel, to achieve the most efficient computation.


Citation: Elizabeth Jurrus, Shigeki Watanabe, Richard J. Giuly, Antonio R. C. Paiva, Mark H. Ellisman, Erik M. Jorgensen, and Tolga Tasdizen, Semi-automated Neuron Boundary Detection and Nonbranching Process Segmentation in Electron Microcopy Images. Neuroinformatics, 2013 1: 5-29 DOI 10.1007/s12021-012-9149-y. PMID: 22644867

Link to PDF of Article