Volume electron microscopy for neuronal circuit reconstruction

The last decade has seen a rapid increase in the number of tools to acquire volume electron microscopy (EM) data. Several new scanning EM (SEM) imaging methods have emerged, and classical transmission EM (TEM) methods are being scaled up and automated. Here we summarize the new methods for acquiring large EM volumes, and discuss the tradeoffs in terms of resolution, acquisition speed, and reliability. We then assess each method’s applicability to the problem of reconstructing anatomical connectivity between neurons, considering both the current capabilities and future prospects of the method. Finally, we argue that neuronal ‘wiring diagrams’ are likely necessary, but not sufﬁcient, to understand the operation of most neuronal circuits: volume EM imaging will likely ﬁnd its best application in combination with other methods in neuroscience, such as molecular biology, optogenetics


Introduction
Vertebrate and invertebrate nervous systems are densely packed with intertwining neuronal axons and dendrites and the synapses between them.The small physical size of these structures, as thin as 40-50 nm in diameter [1,2], requires imaging by electron microscopy (EM), particularly when the goal is the dense reconstruction of neuronal circuits.By imaging volumes of brain using 3dimensional EM, the details of neuronal shape and connectivity can be reconstructed.Importantly, and in contrast to fluorescence-based labeling approaches that require sparse labeling [3] or super-resolution optical imaging [4,5] to resolve densely packed neurites, standard EM stains result in a relatively unbiased staining of all membranes and synapses in the neuropil [6].This means that EM volumes can, in principle, be used to reconstruct the complete connectivity of a neuron with all its presynaptic and postsynaptic partners.Furthermore, this operation can be repeated for all the neurons in the volume, such that the connectivity of the neurons comprising a circuit -its wiring diagram or 'connectome'can be extracted.
The main challenge in volume EM imaging is to acquire a data set of sufficient size, resolution, and completeness that the tortuous trajectories of axons and dendrites can be followed, and the chemical (and, ideally, electrical) synaptic connections identified.The necessary volume depends on the anatomical extent of the circuit to be characterized.For example, the volume of an entire adult nematode, Caenorhabditis elegans (approximately 50 mm Â 50 mm Â 1000 mm, [7]), is about 1% that of a single mouse cortical column (400 mm Â 400 mm Â 1000 mm, [8]).The required 3D voxel resolution depends on how fine the processes are that must be traced: the finer the process, the greater the required resolution to reliably follow it over a long distance.For example, dendritic spine necks in the mammalian central nervous system can be as fine as 40 nm in diameter [2], and the fine neurites of the fruit fly Drosophila melanogaster can be as thin as 50 nm in diameter [1].The required completeness of the EM-imaged volume is related to required resolution: when image data are lost due to staining artifacts, a missed section, or some other glitch in the imaging process, the probability of ambiguities in the dataset increases, resulting in fine processes becoming lost or mixed up during tracing.
There is currently no 'best' volume EM imaging method.Rather, each of the available methods involves tradeoffs in size, resolution and completeness, and which method is most appropriate depends on the scientific questions under investigation.The field, however, is changing rapidly.Here we summarize current volume EM methods, focusing on those designed to answer questions about neuronal circuit structure, and offer suggestions about the types of circuit questions that each method is currently well suited to answer.

Acquisition techniques
The primary dichotomy between modern volume EM methods lays in the choice of widefield transmission electron microscopy (TEM)-based or scanning electron microscopy (SEM)-based techniques.Because TEMbased approaches rely on the imaging of those electrons that pass through a specimen, a requirement is the use of thin sections cut before imaging.SEM, in contrast, is typically used to image electrons backscattered from the surface of samples, allowing the surfaces of both thin sections and block-faces to be imaged.Both types of instrument are capable of high lateral (x-y) resolutions down to a few nanometers, a resolution sufficient for circuit reconstruction.
However, the z-resolution of all techniques (with the exception of electron tomography [9]) depends on the ability to remove thin sections from a block of plastic embedded tissue.The majority of the new technical developments we focus on are aimed at improving this minimum z-resolution or, at least, improving the reliability of thin sectioning.Since the lateral resolution is typically several-fold finer than the section thickness, zresolution imposes one important limit on the usefulness of datasets.

TEM-based methods
In modern serial section TEM (ssTEM [10 ]), individual thin sections, typically between 40 and 90 nm in thickness (e.g.[11 ,12,13]) are cut in an ultramicrotome with a diamond knife (Figure 1a).Sections are picked up by hand onto a 3 mm diameter metal support grid.The grid typically has a slot of length 2 mm, and width 0.2-1 mm.An electron-transparent support film is suspended across the slot, and the thin sections are picked up onto the film.To enhance contrast, the sections are usually post-stained with heavy metals.The grid is then manually inserted into a TEM, and a small (typically 10-20 mm) region of interest is selected for imaging.Electrons are accelerated using an 80-120 kV potential, permitting them to pass through the sample and underlying support film.Some electrons are scattered by the heavy metals in the sample, generating contrast.After passing through a series of electromagnetic enlarger lenses, the transmitted electron image reaches a layer of phosphor, and a digital camera is used to acquire EM images (Figure 1a).TEM imaging is inherently parallel, with each pixel on the camera corresponding to a location in the section.With digital TEM cameras, large image mosaics can be conveniently collected through the use of a motorized x-y stage to translate the sample between image acquisitions, allowing the entire section area to be imaged.
The upper limit for section size in routine ssTEM is 1 mm Â 2 mm, the dimensions of the slot in the support grid.Although custom grids with slots up to 25 mm long have been used (e.g.[14], p. 60), the prospects for increased section width in ssTEM are limited by the fragility of the support film underlying the cut sections and the narrowness of the gap in the objective lens pole piece which admits the sample holder.

SEM-based methods
In a SEM, a finely focused electron beam is raster scanned across the surface of a sample and backscattered electrons are collected with a detector positioned above the sample (Figure 1b-d).Unlike TEM-based imaging, SEM-based imaging is therefore inherently serial.The use of low electron energies (typically 1-3 kV) limits the depth of the back-scattered electron signal to the upper tens of nanometers of the sample [15].Therefore, both thin sections and block faces can be imaged with high zresolution.As in TEM-based methods, following the acquisition of one field of view (typically tens of microns on a side), an x-y stage translates the sample within the SEM, allowing large millimeter-sized mosaics to be imaged.
An automated tape-collecting ultramicrotome (ATUM) has been recently developed to automate the pickup of sections onto a spool of support tape (Figure 1b) [16 ].This technique obviates the specialized and error-prone process of manual sectioning, and provides the additional benefit of cutting sections thinner (30 nm, K. Hayworth, personal communication) than what is routinely possible in conventional ssTEM.Additionally, much wider and longer sections than normal -up to 2.5 mm Â 6 mmcan be cut (K.Hayworth, personal communication), permitting unusually large brain volumes to be thin-sectioned.The use of an electron opaque support tape, however, means that sections collected with ATUM cannot be imaged within a TEM, but rather must use the secondary or backscattered electron signal in a SEM (Figure 1b) [17 ].
Alternatively, the surface of tissue blocks can be imaged in SEMs and sectioned in situ within the SEM vacuum chamber [18 ,19].Ideally, the energy-dependent electron sampling depth is matched to the thickness of tissue that is then removed.Tissue can be removed either mechanically using a diamond knife, as with serial block-face SEM (SBEM, Figure 1c [18 ]) or by milling with a focused ion beam (FIB-SEM, Figure 1d [20 ,21 ]).The cycle of obtaining block-face images and then cutting/milling is fully automated, with no interaction from the experimenter.Unlike ssTEM and ATUM-SEM, block-face SEM methods are destructive; the sections are lost as soon as they are removed from the block face.

Resolution and reliability
The lateral resolutions obtainable in TEMs remain unparalleled, with sub-nanometer resolutions easily achieved in modern TEMs.In practice, sub-nanometer imaging is overkill for the purposes of circuit reconstruction; a common pixel resolution in TEM is $2-4 nm (e.g.[11 ,12,13]).Such resolutions are also achievable in SEMs where the size of the electron probe ultimately limits resolutions to 1-2 nm [22].What resolution do we need for circuit reconstruction?If the minimum diameter of a neurite, such as a spine neck, can be expected to be as thin as 40-50 nm [1,2], the maximum voxel size in any dimension should ideally be at least half as thick (i.e.20-25 nm) to reliably follow every neurite.However, much higher resolutions of 3-5 nm are in principle required for the reliable detection of some types of subcellular structures such as gap junctions [23].
The new techniques are distinguished by improved zresolutions (Figure 2).FIB-SEM [20 ] currently offers the highest z-resolution of 5 nm (Figure 3).Automated serial sectioning using diamond knives, either with SBEM or ATUM-SEM, can lead to repeatable sectioning of 20-30 nm [24 ].Manual sectioning using an ultramicrotome as in ssTEM is typically limited to sections of 40-50 nm [10 ]. z-Resolution in TEM can be improved to a few nanometers through tomographic reconstruction, in which the same field of view is imaged at multiple small tilt increments [9].The number of tilts typically used for EM tomography would be prohibitively slow for volume EM imaging of circuits.However, a recent method of sparse tomographic reconstruction [25 ], in which z-resolution is moderately improved using only a few tilt angles, may permit ssTEM methods to achieve z-resolution comparable to that of SBEM or ATUM-SEM.Sparse tilts in combination with re-imaging at high magnification (0.2 nm x-y resolution) can also be used to disambiguate gap junctions from chemical synapses and other membrane appositions [13].
There is currently a spectrum of opinion regarding the importance of high z-resolution.Some researchers have suggested that a high enough x-y resolution allows most neurites to be followed unambiguously even with modest z resolutions of 45-50 nm [11 ,12,26].Others stress the importance of achieving nearly isotropic voxels, not only for the ability to follow every neurite within a dataset, but also for the development of automated image segmentation algorithms [27].To date, a comprehensive assessment of the degree to which neuron traceability is compromised by the limited z resolution of all techniques has not been performed.
A limitation of manual sectioning remains that some sections inevitably suffer from folds or warping during cutting or post-staining.These errors in the final dataset can be minor (many neurite continuities are traceable across a single missed section [11 ]), although a rigorous analysis of the effects of missed or damaged sections has not been published.Manual sectioning methods also generally suffer from variability in section thickness, making it difficult to directly compare average z-resolutions.The automated sectioning methods (ATUM-SEM and SBEM) were designed with an increase in the repeatability of section thickness in mind.Again, however, a quantitative comparison of section thickness variability between the various methods has not yet been performed.

Acquisition speed
Because local neuronal circuits can span 3D volumes of at least hundreds of microns on a side, acquisition speed has become an increasingly important parameter (Figure 3).A TEM camera array (TEMCA, Figure 1a) enabled high acquisition rates through the use of multiple high frame rate cameras in combination with optimized stage motion and on-line image processing software [11 ].System throughput is largely determined by how quickly image frames can be read out from the cameras in the array, and how quickly the sample can be moved between image frame acquisitions.Saturating the camera sensors in the short period of a single frame exposure can require a large electron dose (electrons/nm 2 ).To avoid rupturing the thin section and its underlying support film with the intense electron beam, various optimizations in sample preparation are important, including carbon coating of the support film, minimization of post-staining artifact, and elimination of any inhomogeneity in the support film.Electron dose must also be monitored with SBEM and FIB-SEM, to prevent too much electron-beam induced damage to the block-face before section removal (although the tolerable dose may be different for SBEM and FIB-SEM).The imaging speed of SEM-based approaches is limited by the bandwidth of electron detectors and the available electron beam current leading to a tradeoff between the acquisition rate and the signal-tonoise ratio of images.In addition to raw image acquisition speed, the duration of an experiment is extended by time spent sectioning, loading and unloading specimens from vacuum, and translating samples under the electron beam to construct image mosaics (Figure 3).For large volumes,

Current Opinion in Neurobiology
Large 3D EM volumes require overlapping fields of view to form 2D mosaics of each section.The time to acquire each FOV and the overhead for each mosaic and section varies for the different techniques.Typical values for each technique are intended to serve only as estimates.Given these estimates, the duration of an experiment to acquire a volume the size of a mouse cortical column (TEM or SEM) are provided in the table.Note that the required voxel resolution will be dependent on characteristics of the tissue under examination, for example, minimum neurite diameter and staining contrast.
the duration of these steps can sum to substantial overhead (see actual time estimates, Figure 3).

Post-acquisition alignment
Once a series of images has been acquired, an alignment step is necessary to stitch the images into a 3D volume.
An advantage of block-face SEM imaging is the inherent registration that comes with acquiring images before sectioning; typically only a simple translational shift of images is required [24 ].Alignment is more complicated for ssTEM and ATUM-SEM sections, often requiring local warping algorithms to compensate for section stretching, folds in sections, and distortions that occur during imaging.While such alignment procedures were, historically, daunting, the use of high-performance computing makes automating complex alignment algorithms possible [11 ,28,29].

Tissue preparation and correlative techniques
The electron dense staining of tissue primarily still relies on chemical compounds first described decades ago during the early years of biological EM.Stains based on high-Z number elements such as osmium, uranium, and lead are most common for each of the techniques described [30].The use of en bloc staining methods [30] is, in particular, essential for the block-face methods in which post-staining of sections is not possible.But the art of staining is by no means dead.The development of indicators to highlight structures of interest, such as membrane-targeted HRP [31,32] or genetically encoded fluorescent reactive oxygen generators for photoconversion [33] is geared toward selectively increasing contrast in genetically controllable ways.
While electron micrographs provide structural information at resolutions unreachable by light microscopy, correlative light microscopy can provide the labeling necessary to bridge the gap to both circuit function and patterns of molecular expression and localization.Array tomography was developed to take advantage of the z resolution of ultrathin sections combined with the ability to multiplex the labeling of antigens [34].By sequentially exposing a section to fluorescently labeled antibodies, a number of different synaptic proteins (for example) can be localized to the same synapse [35].The tissue processing steps needed to label endogenous antigens are, however, often incompatible with good ultrastructural preservation; nevertheless, the labeling of experimentally introduced antigens has been demonstrated with good preservation of ultrastructure [36,37].Additionally, near-infrared branding can be used to place a high-resolution fiducial mark near a fluorescent object of interest, for visualization using both light and EM [38].
Another approach to connect circuit function to structure is to directly measure the activity of populations of neurons from the same piece of tissue to be reconstructed.Given these section dimensions and the resolutions described above, the techniques are currently best suited for different applications.Because FIB-SEM currently obtains the highest 3D resolution but is limited to small volumes, it is best suited for questions of local synaptic circuitry.SBEM is currently best suited for moderately sized (several hundred microns on a side) volumes containing circuits such as those within brain nuclei or perhaps cortical columns.ssTEM and ATUM-SEM, due to the current ability to cut and image the largest area sections, are best suited to multi-millimeter scale connectivity questions, perhaps even of whole brains of, for example, fruit flies or zebrafish larvae.
We anticipate that, as more labs begin to adopt these new techniques, a number of technical advances will follow.We anticipate further increases in z-resolution for each of the described techniques.For ssTEM, a more reliable sectioning and mounting procedure, possibly combined with ATUM technology, will likely aid data reliability.We also anticipate that the TEMCA approach offers substantial headroom for increased imaging throughput.The time estimates for each of the techniques provided in Figure 3 should therefore be taken only as current best estimates and are expected to significantly improve over the next few years.
Parallelizing acquisition across multiple microscopes is an obvious way to further increase throughput.When sections are cut before imaging as in ssTEM or ATUM-SEM, parallelization is straightforward: sections can be imaged in separate microscopes.Parallelizing block-face imaging between multiple microscopes is more difficult because the block would need to be cleaved in such a way to not lose any material along the cut.However, this may be possible using a recently described 'hot-knife' method [40].

Is it worth it?
As neuroscientists interested in the structure of circuits, a common question we are asked is something along the lines of: 'Will wiring diagrams tell us how circuits work?' Our answer is, 'Not on their own.'We argue that wiring diagrams, in the absence of any other information about cell type, synapse type, firing dynamics, etc. are likely insufficient to define circuit function.However, we think wiring diagrams may be necessary to understand circuit function.At a minimum, wiring diagrams in conjunction with complementary data will continue to generate, constrain, and falsify hypotheses about the functional organization of neural circuits.
We have mainly focused on methods for acquiring image data prerequisite to the determination of wiring diagrams, but wish to point out that data acquisition times, while daunting, are still miniscule compared to data analysis times for complete dense reconstruction [27].
In the short term, sparse reconstruction of a subset of the connections contained in vast image datasets will likely continue to yield new insights [11 ,24 ].In the medium-to-long term, we anticipate that the recent progress in automating and speeding the acquisition of volume EM data will carry over to automation of the analysis of these large datasets [41,42], eventually permitting complete reconstruction of mammalian circuits across millimeter scales.

Figure 3 field
Figure 3