Reconstruction of genetically identified neurons imaged by serial-section electron microscopy

Abstract

Resolving patterns of synaptic connectivity in neural circuits currently requires serial section electron microscopy. However, complete circuit reconstruction is prohibitively slow and may not be necessary for many purposes such as comparing neuronal structure and connectivity among multiple animals. Here, we present an alternative strategy, targeted reconstruction of specific neuronal types. We used viral vectors to deliver peroxidase derivatives, which catalyze production of an electron-dense tracer, to genetically identify neurons, and developed a protocol that enhances the electron-density of the labeled cells while retaining the quality of the ultrastructure. The high contrast of the marked neurons enabled two innovations that speed data acquisition: targeted high-resolution reimaging of regions selected from rapidly-acquired lower resolution reconstruction, and an unsupervised segmentation algorithm. This pipeline reduces imaging and reconstruction times by two orders of magnitude, facilitating directed inquiry of circuit motifs.DOI: http://dx.doi.org/10.7554/eLife.15015.001

Current Status of Automatic Serial Sections for 3D Electron Microscopy

“…Researchers at the Harvard University used viral vectors to utilize electrondense peroxidases for staining specific neurons. They used this technique with 3DEM through ATUM (Joesch et al, 2016). In Australia, Filip Braet's group qualitatively and quantitatively reported the advantages and disadvantages of the five current SBF-SEM sample preparation protocol for liver tissue in 2016 and produced blocks using selected methods for correlative light and electron microscopy (CLEM), combining volume images with confocal laser scanning microscopy (CLSM) and EM.…”

Flexible Learning-Free Segmentation and Reconstruction for Sparse Neuronal Circuit Tracing

“…The predecessor to FLoRIN, introduced in Jösch et al 37 used Otsu's method 45 to binarize neural images. Otsu thresholding uses a global value to determine the threshold, making it vulnerable to grayscale shifts and low signal-to-noise ratios.…”

“…Whereas the goal of segmenting and reconstructing µCT X-ray volumes is to discover large structures and high-level information about the structure of the brain, sSEM images are of high enough resolution to follow individual processes (Figure 3 Panel A). The dataset introduced by Jösch et al 37 contains APEX2 positive processes with a sample preparation designed to enhance microstructures of interest for automatic sparse segmentation (Figure 3 Panel B). APEX2 labels starburst amacrine cells in the dataset, thus we attempt to reconstruct the associated dendritic branches of these cells (Figure 3 Panels C-F).…”

“…Learning-free computer vision methods incur less computational overhead than machine learning methods, allowing for real-time segmentation of neural volumes as they are imaged. We compared the running time of FLoRIN with that of a fully convolutional neural network, 3D U-Net 34 as well as previous work in EM sparse reconstruction 37 , and achieved high quality results in remarkably short time-frames.…”

“…Special tissue preparations can enhance the contrast of specific cell types 37 , and certain imaging techniques can selectively highlight specific anatomical structures 40 . These methods lead to a sparse reconstruction problem, which is much faster to solve than the exhaustive dense reconstruction problem, and provides useful results for multi-modal image reconstruction problems (e.g., co-registration of different types of microscopy) and cell-specific studies 10,11,41 .…”