The dense reconstruction of neuronal circuits has become an increasingly realistic goal in the neurosciences thanks to developments in large-scale 3D EM (Denk and Horstmann, 2004; Bock et al., 2011; Hayworth et al., 2015; Kasthuri et al., 2015; Xu et al., 2017) and circuit reconstruction techniques (Saalfeld et al., 2009; Boergens et al., 2017). First locally dense neuronal circuits have been mapped using these techniques (Helmstaedter et al., 2013; Kasthuri et al., 2015; Berck et al., 2016; Wanner et al., 2016). However, in mammalian nervous systems, any given local neuronal circuit receives numerous synaptic inputs from distant neuronal sources. In most layers of the mammalian cortex, for example, the fraction of distal input synapses in a dense local circuit is estimated to comprise 70–90% of all synapses (Stepanyants et al., 2009). Similarly, subcortical structures as the striatum or amygdala (Figure 1a) receive the vast majority of their inputs from distal projection sources. Uncovering the synaptic logic of such multi-source circuits together with the local dense neuronal connectivity requires the identification of multiple input sources in the same connectomic experiment. Since 3D EM is still limited to volumes that do not routinely encompass entire mammalian brains, the origin of a large fraction of input synapses remains thus unidentified in current high-resolution connectomic experiments in mammals.

Figure 1

Download asset Open asset

At the same time, the labeling of multiple neuronal populations and their axonal projections by fluorescent dyes visible in LM is routine (Livet et al., 2007; Mao et al., 2011; D'Souza et al., 2016). Combining light-microscopic fluorescent labels in axons from multiple sources with large-scale 3D EM would be ideal. Such direct correlative light- and electron microscopy (CLEM) has been successful for high-resolution subcellular label localization (Agronskaia et al., 2008; Murphy et al., 2011; Liv et al., 2013) and synaptic identification (Micheva and Smith, 2007; Rah et al., 2013), but can so far not be combined with 3D EM at both resolution and scale allowing for dense large-scale circuit mapping in mammalian neuropil.

Similarly, en-bloc immunolabeling in conjunction with EM is limited to thin tissue slices (Faulk and Taylor, 1971; Pallotto et al., 2015). Approaches to interlace EM and LM imaging on alternating tissue slices, though successfully applied in invertebrates (Shahidi et al., 2015), impede dense EM-based circuit reconstruction in mammalian nervous tissue. Rather, methods to obtain electron-dense labels in a subset of neurons have been developed over the decades based on induced axonal degeneration (Gray and Hamlyn, 1962; Colonnier, 1964; White, 1978), introduction of HRP (Hersch and White, 1981; Hamos et al., 1985; Horikawa and Armstrong, 1988; Anderson et al., 1994a; Anderson et al., 1994b; Markram et al., 1997; da Costa and Martin, 2011) or fluorophores (Maranto, 1982; Grabenbauer et al., 2005; Knott et al., 2009) into a subset of neurons which could be utilized to create electron contrast via enzymatic or photo-oxidative DAB polymerization. More recent improvements enabled the genetically targeted expression of electron-dense labels in diverse cellular compartments, potentially allowing to encode multiple label classes in a single EM experiment (Shu et al., 2011; Martell et al., 2012; Atasoy et al., 2014; Lam et al., 2015; Joesch et al., 2016). However, generating multiple electron-dense label classes in the very same experiment while ensuring ultrastructural preservation is challenging and to our knowledge has been accomplished in Drosophila (Lin et al., 2016) but so far not in mammals. Furthermore, the number of cellular compartments usable for such axonal identification may be rather limited.

Methods to directly co-register light microscopic labels in axons and dendrites with 3D EM data (Bishop et al., 2011; Maco et al., 2013; Gala et al., 2017) have been restricted to local volumes of up to 10 µm on a side, using LM-burnt fiducial marks visible in EM; these do not scale to the large volumes required for neuronal circuit reconstruction (sized hundreds of micrometers per dimension, that is at least 1000-fold larger). Direct co-alignment between fluorescently labeled cell bodies and EM data has been routinely applied (Bock et al., 2011; Briggman et al., 2011; Lee et al., 2016; Tsang et al., 2018), but provides LM-to-EM alignment only at a coarse scale that does not allow for the identification of single axons.

Apart from the experimental limitations currently hampering the use of multi-color labels in large 3D EM data, current computational approaches are challenged when attempting submicron registration precision of neurites in volumes required for connectomics (~10⁶ µm³ or larger). Previously proposed multi-modal graph-based neurite registration (Serradell et al., 2012; Fua and Knott, 2015; Serradell et al., 2015) can in principle be used for LM-to-EM morphological matching, but did not yet scale to graphs of thousands or tens of thousands of nodes as required for connectomics.

Thus, a method that could directly employ the full multi-color space of LM in large 3D tissue samples while enabling dense EM-based circuit reconstruction would be ideal. Here, we report FluoEM, a set of experimental and computational tools to achieve this goal. Instead of directly attempting alignment of LM-labeled axons to the EM dataset at the nanometer scale, we exploit the availability of locally dense neurite reconstructions in current volume EM. Here, the alignment problem can be re-formulated as identifying the most likely axon (out of all axons, reconstructed in EM) that best explains an LM-imaged axonal fluorescence signal. We show that this is a well-constrained problem based on the actual geometry of axonal fibers in nerve tissue from mouse cerebral cortex. We report the experimental methods allowing for subsequent correlative imaging of samples in LM and EM and the computational tools to register the labeled LM axons to their EM correspondences. We exemplify our method for long-range inputs to layer 1 of mouse cortex.

The FluoEM workflow (Figure 1b) comprised the following steps: acquisition of a 3D fluorescence-LM dataset from a given piece of tissue; 3D imaging of that very same piece of tissue in the electron microscope; reconstruction of the fluorescently labeled axons in the LM data; reconstruction of locally all axons in small subvolumes of the EM data; computational determination of the most likely EM axons that explain the LM signal.

In a first step, we used blood vessels as intrinsic fiducials in the neuropil to obtain a coarse co-registration between LM and EM (Figure 1b). Coarse LM-EM registration based on cell bodies and blood vessels has been successfully performed before (Bock et al., 2011; Briggman et al., 2011) and is expected to yield an alignment precision ${\displaystyle {\lambda }_{align}}$ of about 5–10 µm (Figure 1c). We next had to determine the typical length of faithful 3D axon reconstruction ${\displaystyle d_{recon}}$ in fluorescence data (Figure 1d). Then, the problem of matching EM axons to LM data can be phrased as follows (Figure 1e): Given all axons that traverse a cube of edge length ${\displaystyle {\lambda }_{align}}$ in the EM data, will the trajectory of these axons become unique on a length scale ${\displaystyle d_{unique}}$ that is smaller than or comparable to the LM-reconstruction length ${\displaystyle d_{recon}}$: ${\displaystyle d_{unique}}$ (${\displaystyle {\lambda }_{align}}$) ≤ ${\displaystyle d_{recon}}$? In other words: What is the travel length ${\displaystyle d_{unique}}$ after which all axons from a center cube of size ${\displaystyle {\lambda }_{align}}$ will have no other of these axons closer than ${\displaystyle {\lambda }_{align}}$?

In order to understand whether this approach could work, we had to measure ${\displaystyle d_{recon}}$ in LM data with realistic labeling density and determine ${\displaystyle d_{unique}}$ for a realistic ${\displaystyle {\lambda }_{align}}$ in EM data. These measurements are reported next, followed by a proof-of-principle co-registration experiment and results from the virtual labeling of axons in EM data.

Reconstruction length and axon uniqueness

We first performed a fluorescence imaging experiment (Figure 2a–c) in which we injected adeno-associated viruses expressing fluorescent markers into the primary motor cortex (M1; virus expressing the green marker eGFP) and the secondary somatosensory cortex (S2; virus expressing the red marker tdTomato) of a 28-day-old mouse, followed by transcardial perfusion and tissue extraction at 46 days of age. Using a confocal laser scanning microscope, we then acquired a dataset sized (828 × 828 × 75) µm³ at a voxel size of (104 × 104 × 444) nm³ from layer 1 of a part of cortex in which both of these projections converged (located to lateral parietal association cortex (LPtA), Figure 2b). We then 3D-reconstructed all axons traversing a (40 × 40 × 48) µm³ bounding box in this LM image dataset using our annotation tool webKnossos (Boergens et al., 2017) and measured the length over which the axons were faithfully reconstructable as judged by expert annotators (for this we asked the annotators to stop reconstructing whenever they encountered a location of any ambiguity, the loss of axon continuity or an unclear branching). The reconstruction length ${\displaystyle d_{recon}}$ per axon was 70 ± 44 µm (mean ± s.d., n = 46 axons) for the green and 134 ± 62 µm (n = 42 axons) for the red fluorescence channel. Ninety-five percent of axons were reconstructable above 20 µm length, some for up to 200 µm (Figure 2c, independent reconstruction by two experts).

Figure 2 with 1 supplement see all

Download asset Open asset

Figure 2—source data 1

Axonal uniqueness analysis for cortical layer 1 (Figure 2g).

https://doi.org/10.7554/eLife.38976.005

Download elife-38976-fig2-data1-v2.xlsx

Figure 2—source data 2

Axonal uniqueness analysis for cortical layer 2/3 (Figure 2h).

https://doi.org/10.7554/eLife.38976.006

Download elife-38976-fig2-data2-v2.xlsx

Figure 2—source data 3

Axonal uniqueness analysis for cortical layer 4 (Figure 2i).

https://doi.org/10.7554/eLife.38976.007

Download elife-38976-fig2-data3-v2.xlsx

Figure 2—source data 4

Axonal uniqueness summary boxplot (Figure 2j).

https://doi.org/10.7554/eLife.38976.008

Download elife-38976-fig2-data4-v2.csv

Figure 2—source data 5

Axonal uniqueness dependence on alignment error (Figure 2k).

https://doi.org/10.7554/eLife.38976.009

Download elife-38976-fig2-data5-v2.xlsx

We then measured the length ${\displaystyle d_{unique}}$ over which the trajectory of axons in cortex becomes unique, assuming a macroscopic alignment precision of ${\displaystyle {\lambda }_{align}}$ = 5 µm (Figure 2d–k). For this, we acquired a 3D EM dataset sized (130 x 110 x 85) µm³ at a voxel size of (11.24 x 11.24 x 30) nm³ from layer 1 of LPtA cortex (Figure 2d,e) using SBEM (Denk and Horstmann, 2004) and reconstructed all n = 220 axons that traversed a cubic bounding box of size (5 x 5 x 5) µm³ in the center of the dataset (Figure 2e, blue box). We then analyzed for each axon how many of the other 219 axons were still within ${\displaystyle {\lambda }_{align}}$ proximity at a Euclidean distance d from the center cube (Figure 2f). For a given axon, the uniqueness length was then computed as the distance after which no other axon was in the ${\displaystyle {\lambda }_{align}}$-surround. Figure 2g shows the number of remaining axons after distance d for a dense reconstruction of axons in L1 of mouse cortex. The uniqueness length per axon was 22 ± 10 µm (mean ± s.d., range 7 – 78 µm, n = 220 axons); after 31, 35, 41 and 69 µm, 85%, 90%, 95% and 98% of axons had no other axon in their proximity, respectively. We repeated this analysis in two additional 3D EM datasets obtained from layers 2/3 and 4 of mouse somatosensory cortex (datasets 2012-09-28_ex145_07x2 and 2012-11-23_ex144_st08x2, KM Boergens & MH, unpublished, Figures 2h,i). Here, uniqueness lengths per axon were 21 ± 9 µm (mean ± s.d., range 7 – 62 µm, n = 214, L2/3) and 17 ± 8 µm (mean ± s.d., range 6 – 55 µm, n = 128, L4), with 90% of axons unique after ${\displaystyle d_{unique}^{90}}$= 35 µmin L2/3 and ${\displaystyle d_{unique}^{90}}$= 32 µm in L4 (Fig. 2j, $d_{unique}^{90}$was indistinguishable between layers 1, 2/3 and 4, Figure 2—figure supplement 1).

Importantly, when comparing $d_{unique}^{90}$ in layer 1 as determined in 3D EM data to the reconstruction lengths ${\displaystyle d_{recon}}$ obtained in the fluorescence datasets (Figure 2c), we find that about 45% (green channel) and about 80% (red channel) of the faithfully LM-reconstructed axonal stretches were at least of length ${\displaystyle d_{unique}^{90}}$. Since for a first LM-EM co-registration a set of the longest LM-reconstructable axons can be used, these data strongly suggested that the FluoEM approach for matching-based co-alignment between LM and EM reconstructed axons could in principle be successful. Furthermore, we expected the uniqueness length $d_{unique}^{90}$ to depend on the macroscopic alignment precision ${\displaystyle {\lambda }_{align}}$ achieved by the blood-vessel-based coarse pre-registration of the LM and EM datasets (Figure 2k). When re-analyzing our dense axonal reconstructions in the 3D EM dataset from L1 for smaller ${\displaystyle {\lambda }_{align}}$, we find that the uniqueness length strongly decreased for smaller ${\displaystyle {\lambda }_{align}}$ (to $d_{unique}^{90}$ = 5 µm at ${\displaystyle {\lambda }_{align}}$ = 1 µm; ${\displaystyle d_{unique}^{90}=d_{recon}^{90}}$(eGFP) = 11 µm at ${\displaystyle {\lambda }_{align}}$ = 2 µm; ${\displaystyle d_{unique}^{90}=d_{recon}^{90}}$(tdTomato) = 20 µm at ${\displaystyle {\lambda }_{align}}$ = 3 µm).

Correlated LM and 3D EM dataset acquisition

We then set out to perform correlated 3D LM-EM imaging for the application of FluoEM (Figure 3). For this, we continued the experiment described above (Figure 2a) and acquired an additional fluorescence channel to image blood vessels (during perfusion, the rinsing solution was supplemented with DiD, a lipophilic dye for blood vessel staining with fluorescence in the far red spectrum, around 650 nm). Then we applied a slightly modified version of our standard volume EM staining protocol (Hua et al., 2015) to the tissue block (see Materials and methods), embedded and mounted the sample and imaged it using SBEM (Denk and Horstmann, 2004) (Figure 3a). We first acquired EM images with a large field of view (dataset dimensions (1500 × 1000 × 10) µm³) at lower resolution of (270 × 270 × 200) nm³ to identify the location of the LM-imaged dataset relative to the EM imaged sample (Figure 3b,c). For this, we used the pattern of superficial blood vessels as coarse alignment fiducials (Figure 3d,e). We then acquired a high-resolution 3D EM dataset sized (130 × 110 × 85) µm³ at a voxel size of (11.24 × 11.24 × 30) nm³ in a region previously imaged using LM with suitable fluorescent labeling density (Figure 3c,g). In both the high-resolution EM dataset and the corresponding LM volume we then reconstructed the blood vessels (Figure 3h,i) and annotated their branch points as control points for fitting a coarse affine transformation AT_BV from the LM to the high-resolution EM dataset (Figure 3h–n, n = 7 control points). To estimate the precision of this coarse alignment, we measured its residuals (Figure 3o; 2.7 ± 1.1 µm, mean ± s.d., n = 7, median = 2.7 µm). Since this coarse alignment precision was well in the range previously assumed for the analysis of axonal uniqueness (${\displaystyle {\lambda }_{align}}$≈ 5 µm, Figure 3g–k), we then attempted a first LM-to-EM matching.

Figure 3

Download asset Open asset

Figure 3—source data 1

Blood vessel based affine registration error (Figure 3o).

https://doi.org/10.7554/eLife.38976.011

Download elife-38976-fig3-data1-v2.xlsx

LM-to-EM axon matching

We chose an axon next to a blood vessel in the LM data (Figure 4a). The axon was skeleton-reconstructed at the LM level (Figure 4b) using webKnossos (Boergens et al., 2017). We then transformed this LM reconstruction (path length of 105 µm) to the EM dataset using the coarse transformation AT_BV based on blood vessel correspondence as determined before (Figure 4c). Evidently, based solely on this coarse transformation, an identification of the corresponding axon at the EM level was impossible (Figure 4d). Rather, we then defined a cubic region sized (5 µm)³ in the EM data next to the blood vessel at which the LM reconstruction had been seeded (Figure 4e,f) and reconstructed all axons that traversed that region in the EM dataset (Figure 4g, n = 193 axons). We then used a spherical search kernel with 5 µm radius (Figure 2f) to determine which of the 193 EM-reconstructed axons were within that surround of the transformed LM-reconstructed axon along all of its trajectory (Figure 4h,i). Beyond a distance of 42 µm along the transformed LM axon, only one candidate EM-reconstructed axon remained persistently within the search kernel (Figure 4i). So, in fact, for this example fluorescent axon, we had found a corresponding axon at the EM level that matched its trajectory.

Figure 4 with 1 supplement see all

Download asset Open asset

Figure 4—source data 1

Axonal uniqueness analysis for initial axon match (Figure 4i).

https://doi.org/10.7554/eLife.38976.014

Download elife-38976-fig4-data1-v2.xlsx

But how could we be sure that this EM axon was the one that had given rise to the fluorescence signal at the LM level? We first investigated the detailed morphology of this axon and identified axonal varicosities in the EM data. We then reconstructed the varicosities visible in the LM data and overlaid both (Figure 4j). The similarity was visually striking (distance between LM and EM varicosities, 0.7 µm ± 0.6 (mean ± s.d., n = 11); distance between LM and random EM varicosity distributions: 4.9 µm ± 2.1, (mean ± s.d., n = 11 varicosities, n = 10⁵ draws, p < 10⁻⁵, Randomization test, Figure 4—figure supplement 1)). With this, we had found the EM axon that most likely explained the fluorescence signal reconstructed in the LM dataset (Figure 4k).

We then continued to match other fluorescent axons given this first matched axon (Figure 5). For this, we chose an axon in the LM data in proximity to the previously matched axon and selected an LM varicosity close to the axonal apposition as a seed point for the correspondence search in the EM data (Figure 5a). Since each matched axon provided additional registration constraints beyond the initial blood-vessel-based constraints, we could now use each matched axon to further refine an elastic-free-form LM-to-EM transformation using the matched varicosities, branchpoints and other prominent structural features of the axons seen in both LM and EM (Figure 4b,k). The iteratively refined registration together with the usage of varicosities as search criteria allowed us to substantially reduce the work load for matching subsequent fluorescent axons to their EM counterparts (Figure 5b–e; in this example 34 axons were within a (2 × 2 × 3) µm³ search volume around the LM-varicosity, but only six of these had a varicosity). While the first match involved the reconstruction of 193 axons, the subsequent matches were successful after only several or even one reconstruction attempt and a few minutes of reconstruction time (Figure 5e). The fact that all subsequent matches were successful (see varicosity correspondence as shown in Figure 5f) implies that the previous matches were correct (since the probability of finding a correctly matched axon by chance after several chance matches is virtually zero). Using this strategy, we matched a total of 38 axons (27 in the red fluorescence channel, 11 in the green channel), using a total of 284 control points (7.5 ± 3.8 cps per axon (mean ± s.d), Figure 5g).

Figure 5 with 1 supplement see all

Download asset Open asset

Figure 5—source data 1

Reconstruction effort (Figure 5e).

https://doi.org/10.7554/eLife.38976.017

Download elife-38976-fig5-data1-v2.xlsx

Figure 5—source data 2

Axon based affine and free-form registration error (Figure 5h).

https://doi.org/10.7554/eLife.38976.018

Download elife-38976-fig5-data2-v2.xlsx

Figure 5—source data 3

Averaged axon based affine and free-form registration error depending on control point numbers (Figure 5i).

https://doi.org/10.7554/eLife.38976.019

Download elife-38976-fig5-data3-v2.xlsx

Figure 5—source data 4

Matched LM signal fraction (Figure 5o).

https://doi.org/10.7554/eLife.38976.020

Download elife-38976-fig5-data4-v2.xlsx

To quantify the improved registration precision obtained from the iteratively constrained LM-to-EM axon matches, we measured the residual registration error of both affine and free-form registration (constrained by a random subset of 250 control points) on 30 other control points. Compared to the affine registration, the residual registration error was reduced by one third from 725 ± 408 nm (mean ± s.d) to 487 ± 290 nm (mean ± s.d, Figure 5h; Figure 5—figure supplement 1), thus about six-fold smaller than in the blood vessel based coarse pre-alignment obtained before (cf. Figure 3o; see Figure 5i for dependence on control point number).

To test whether we could perform a close-to-complete matching of LM-to-EM reconstructions for a certain depth range of the fluorescence dataset we matched all clearly visible fluorescently labelled axons at a depth of 17–20 µm in the red (tdTomato) channel (n = 27 axons, total path length of 3 mm). We then computed a segmentation of the LM data and measured the fraction of the segmentation voxels overlapping with the EM-matched reconstructions (Figure 5j–n). We found that in fact about 90% (86 ± 3%, mean ± s.d, n = 7 LM image planes) of the fluorescent voxels were explained by the matched EM axons in this depth range (Figure 5o).

In summary, investing a total of 195 work hours (178.8 hr for the first axon; 16.6 hr for the remaining axons) we obtained the EM equivalents of 38 fluorescently labeled axons in two fluorescence channels (total axonal path length of 4.6 mm). The total reported time included LM-to-EM matching (181.1 hr), full axon reconstruction in EM (4.8 hr) and control point placement in LM and EM (9.5 hr). Since the matching of the first axon consumed most invested time, we checked whether by using the distribution of varicosities along EM and LM axons also for the initial match the time investment could be further reduced. In fact, when asking a second independent expert to perform the first LM-to-EM axonal match by also using varicosities as a search criterion, the invested total time was only 30.3 min for the initial match (7.5 min for EM-to-LM matching, 6.8 min full axon reconstruction in EM, 16.0 min control point placement in LM and EM). While this reduction in matching effort is well applicable to settings in which axonal varicosities can be detected, the more general but more laborious FluoEM approach of locally dense axonal reconstruction is expected to be applicable to other settings (for example transit axons that rarely form synapses). Together, we conclude that FluoEM is effective and efficient for LM-to-EM registration.

Comparison of axon and synapse detection in LM and EM

The axons matched between the LM and EM volumes provided a dataset to calibrate in hindsight, how well axonal trajectories and branches can be reconstructed and how faithfully synapses can be detected in fluorescence LM data (Figure 6).

Figure 6

Download asset Open asset

Figure 6—source data 1

Relation between sphere equivalent diameter in LM and EM for synaptic and non-synaptic axonal varicosities (Figure 6n).

https://doi.org/10.7554/eLife.38976.022

Download elife-38976-fig6-data1-v2.xlsx

Figure 6—source data 2

Sphere diameter dependent synaptic and non-synaptic varicosity likelihood (Figure 6o).

https://doi.org/10.7554/eLife.38976.023

Download elife-38976-fig6-data2-v2.xlsx

Figure 6—source data 3

Reconstruction effort reproduction experiment (Figure 6r).

https://doi.org/10.7554/eLife.38976.024

Download elife-38976-fig6-data3-v2.xlsx

We first asked a different expert annotator to reconstruct the previously matched axons in the LM dataset, this time without an explicit bias to stop reconstruction at uncertain locations (cf. Figure 2c), and without knowledge of the matched EM axon’s shape or trajectory. We then compared these axonal reconstructions performed in the LM dataset with the corresponding EM trajectories of the axons, which we considered the ground truth for the detection of branchpoints, the determination of the axonal morphology and the presence of presynaptic boutons (Figure 6a–f). We found that while a large fraction of axons were reconstructed faithfully at the LM level (27 of 38 axons, 71%), in 29% of cases the LM reconstruction contained an error (Figure 6f). The largest fraction of errors was related to incorrectly terminated reconstructions (‘stop’, 38%, Figure 6b,f), followed by an incorrect continuation of the LM-reconstructed axon at locations where multiple fluorescent axons overlapped (31% of errors, Figure 6c,f). Furthermore, missed (23%) and added (8%) branches occurred (Figure 6d,e,f).

Together, this data illustrates that at the chosen labeling density, fluorescence-based axon reconstruction in cortex is error-prone. Labeling density was 1.2 ± 0.2% (mean ± s.d., n = 115 planes) for the tdTomato channel and 1.4 ± 0.4% (mean ± s.d., n = 115 planes) for the eGFP channel; error rates evaluated separately for axons in each channel were 30% of axons (tdTomato) and 27% of axons (eGFP), respectively.

We then identified 255 axonal varicosities in the LM data and investigated their ultrastructure at the EM level based on the successfully matched axonal counterparts (Figure 6a,g–o). As the examples indicate (Figure 6g–m), many varicosities in the LM in fact correspond to presynaptic boutons (63%, n = 161 of 255, Figure 6h,k,l). A substantial fraction of LM-varicosities, however, are without evidence for a presynaptic specialization and instead contain only mitochondria (n = 54, Figure 6g,j), or are hollow thickenings (n = 39, Figure 6i) or myelinated (n = 1, Figure 6m). Notably, while the inference of synaptic boutons from LM varicosities was dependent on the size of the LM varicosity (Figure 6n) with larger varicosities more likely synaptic, there was no LM varicosity size range that allowed unequivocal inference of a synapse from the LM morphology (Figure 6n,o).

We report FluoEM, a set of experimental and computational tools for the virtual labeling of multiple axonal projection sources in connectomic 3D EM data of mammalian nervous tissue. At its core is the notion that dense EM circuit reconstruction allows to reformulate the alignment between data obtained at micrometer resolution (LM) to data obtained at nanometer resolution (EM) to the best-matching between a sparse subset of labeled axons (LM) and all possible axons (EM). We find that in mammalian neocortex, densely reconstructed axons have unique trajectories on a spatial scale that is compatible with currently achievable 3D EM dataset sizes, allowing the FluoEM concept to be successfully applied. With this, FluoEM allows the enriching of EM-based connectomes in mammalian neuropil with the large range of multi-color encoded information already successfully expressed in axons at the light-microscopic level.

One key prerequisite for FluoEM is the uniqueness of axonal trajectories in dense neuropil. We have so far determined the uniqueness length for several layers of the mouse cerebral cortex (Figure 2), where this length is consistently below 40 µm (90th percentile, see Figure 2), thus in a range that can be faithfully reconstructed at the LM level (Figure 2a–c). For a broad range of cortical and subcortical nervous tissue (basal ganglia, thalamus, hypothalamus, nuclei of the brain stem etc.) in which locally dense circuits coexist with incoming projection axons from distant sources, FluoEM can be applied on the spatial scales described here. Cases in which dense neuropil contains highly anisotropic axons (for example axon bundles in the white matter of mammalian brains) will likely imply a longer scale on which axonal trajectories become unique, and therefore require longer reconstruction lengths in the fluorescence data.

We have developed FluoEM using a particular 3D EM imaging technique, SBEM (Denk and Horstmann, 2004). Since FluoEM only relies on the underlying geometry of axonal trajectories, which we have shown to be unique on scales achievable with all current 3D EM techniques (Figure 2g–k), FluoEM is not restricted to the combination with SBEM. Rather, other 3D EM imaging approaches, especially FIB-SEM (Heymann et al., 2006; Knott et al., 2008; Hayworth et al., 2015; Xu et al., 2017) and ATUM-SEM (Schalek et al., 2011; Kasthuri et al., 2015) that provide faithful axonal reconstructions over at least about 40–50 µm extent in three dimensions will be useable for the presented approach.

While FluoEM in principle allows the identification of as many axonal projection sources in a single connectomic experiment as can be encoded at the light-microscopic level, it is practically limited by the number of fluorescence channels that can be acquired in sequence without exerting damage to the tissue. In our experiments, tissue exposed to 2 hr of continuous LM imaging with acquisition of three parallel fluorescence channels was still well usable for subsequent 3D EM circuit reconstruction. To acquire more fluorescence channels without increasing the local photon dose in the tissue, one can exploit the fact that in FluoEM, it is not necessary to image the axons of interest in their entirety at the LM level. Rather, it is sufficient to LM-image volumes of 40–50 µm in extent, as long as all axons of interest pass through that volume, and to reconstruct the complete axons in EM. Therefore, a setting in which the volume of interest is split into several subvolumes, in each of which a limited number of fluorescence channels is acquired, could multiplex the color range of FluoEM (Figure 7a–b) without increasing the exposure time in the LM step.

Figure 7

Download asset Open asset

In LM imaging of mammalian nervous tissue, labeling density of axons poses a notorious limitation as long as axonal trajectories were to be reconstructed from the LM data (labeling more than about 1 in 1000 to 10,000 axons impedes reconstruction (Helmstaedter, 2013); and LM reconstruction errors may increase substantially with increased labeling density, see Figure 6b–f). With FluoEM, labeling density is only limited by the need to follow axons over the uniqueness length (Figures 2 and 3). Since this depends on the overall alignment precision (Figure 2k), a setting in which one fluorescence channel contains sparsely labeled axons to provide an accurate pre-alignment (see Figure 5g,h), and the other fluorescence channels are densely labeled is a possible solution, in which LM-reconstructable stretches of axons as short as 5 µm would be sufficient for co-registration (Figures 2k and 5h). For the experiment described here, we titrated the labeling density by adjusting the injection volume while keeping the virus titer constant (see Materials and methods). Depending on the distance and connectivity of the projection sources to the FluoEM target area this parameter might need to be adjusted. An extreme example of sparse labeling could be the parallel intracellular injection of two or more (neighboring) neurons in one part of cortex (Figure 7c) and the investigation of their convergent projections in other parts of the brain (Figure 7c); in this case, FluoEM could be used to study the synaptic output logic of multiple identified neurons at single-cell and single-synapse level.

Finally, the reconstruction work load in FluoEM requires a quantitative discussion. In settings where axonal varicosities can be used as additional constraints on the matching (Figure 5a–d), the LM-to-EM match consumed about 20–30 min per axon (including the full EM reconstruction of the axon and the placement of control points). With this, a full matching of a sparsely stained volume sized (1 × 1 × 0.1) mm³, for example a large area of layer one in the mouse neocortex, would thus likely consume about 5300 work hours in FluoEM (using webKnossos (Boergens et al., 2017), see Materials and methods for estimates). While this is a considerable time investment, it favorably compares to other resource investments in biomedical imaging. Furthermore, with the development of increasingly reliable automated approaches for the reconstruction of axons and synapses at the EM level (Berning et al., 2015; Dorkenwald et al., 2017; Staffler et al., 2017; Januszewski et al., 2018), a further speedup of FluoEM is conceivable in which partially automated matching contributes. In mammalian circuits where axons may not show local diameter variations that can be used for co-registration, the FluoEM approach using dense axonal reconstruction (Figure 2) is applicable.

Together, the presented methods overcome the major hurdle of limited color space for identifying multiple axonal input sources in parallel in large-scale and dense electron microscopic reconstruction of mammalian neuronal circuits.

All imaging data is available for online browsing and annotation at demo.webknossos.org as detailed in the data availability section of the Methods.

1. 1. Drawitsch F
   2. Helmstaedter M

   (2018) FluoEM low-res EM dataset

   Openly accessible via webknossos.org (FluoEM_2016-05-23_FD0144-2_st001_v1).

   https://demo.webknossos.org/datasets/FluoEM_2016-05-23_FD0144-2_st001_v1/view
2. 1. Drawitsch F
   2. Helmstaedter M

   (2018) FluoEM high-res EM dataset

   Openly accessible via webknossos.org (FluoEM_2016-05-26_FD0144-2_v2s2s).

   https://demo.webknossos.org/datasets/FluoEM_2016-05-26_FD0144-2_v2s2s/view
3. 1. Drawitsch F
   2. Helmstaedter M

   (2018) FluoEM LM dataset

   Openly accessible via webknossos.org (FluoEM_2016-06-02-FD0144_2_Confocal).

   https://demo.webknossos.org/datasets/FluoEM_2016-06-02-FD0144_2_Confocal/view

1. 1. Agronskaia AV
   2. Valentijn JA
   3. van Driel LF
   4. Schneijdenberg CT
   5. Humbel BM
   6. van Bergen en Henegouwen PM
   7. Verkleij AJ
   8. Koster AJ
   9. Gerritsen HC

   (2008) Integrated fluorescence and transmission electron microscopy

   Journal of Structural Biology 164:183–189.

   https://doi.org/10.1016/j.jsb.2008.07.003

   • PubMed
   • Google Scholar
2. 1. Anderson JC
   2. Douglas RJ
   3. Martin KA
   4. Nelson JC

   (1994a) Map of the synapses formed with the dendrites of spiny stellate neurons of cat visual cortex

   The Journal of Comparative Neurology 341:25–38.

   https://doi.org/10.1002/cne.903410104

   • PubMed
   • Google Scholar
3. 1. Anderson JC
   2. Douglas RJ
   3. Martin KA
   4. Nelson JC

   (1994b) Synaptic output of physiologically identified spiny stellate neurons in cat visual cortex

   The Journal of Comparative Neurology 341:16–24.

   https://doi.org/10.1002/cne.903410103

   • PubMed
   • Google Scholar
4. 1. Atasoy D
   2. Betley JN
   3. Li WP
   4. Su HH
   5. Sertel SM
   6. Scheffer LK
   7. Simpson JH
   8. Fetter RD
   9. Sternson SM

   (2014) A genetically specified connectomics approach applied to long-range feeding regulatory circuits

   Nature Neuroscience 17:1830–1839.

   https://doi.org/10.1038/nn.3854

   • PubMed
   • Google Scholar
5. 1. Berck ME
   2. Khandelwal A
   3. Claus L
   4. Hernandez-Nunez L
   5. Si G
   6. Tabone CJ
   7. Li F
   8. Truman JW
   9. Fetter RD
   10. Louis M
   11. Samuel AD
   12. Cardona A

   (2016) The wiring diagram of a glomerular olfactory system

   eLife 5:e14859.

   https://doi.org/10.7554/eLife.14859

   • PubMed
   • Google Scholar
6. 1. Berning M
   2. Boergens KM
   3. Helmstaedter M

   (2015) SegEM: efficient image analysis for High-Resolution connectomics

   Neuron 87:1193–1206.

   https://doi.org/10.1016/j.neuron.2015.09.003

   • PubMed
   • Google Scholar
7. 1. Bishop D
   2. Nikić I
   3. Brinkoetter M
   4. Knecht S
   5. Potz S
   6. Kerschensteiner M
   7. Misgeld T

   (2011) Near-infrared branding efficiently correlates light and electron microscopy

   Nature Methods 8:568–570.

   https://doi.org/10.1038/nmeth.1622

   • PubMed
   • Google Scholar
8. 1. Bock DD
   2. Lee WC
   3. Kerlin AM
   4. Andermann ML
   5. Hood G
   6. Wetzel AW
   7. Yurgenson S
   8. Soucy ER
   9. Kim HS
   10. Reid RC

   (2011) Network anatomy and in vivo physiology of visual cortical neurons

   Nature 471:177–182.

   https://doi.org/10.1038/nature09802

   • PubMed
   • Google Scholar
9. 1. Boergens KM
   2. Berning M
   3. Bocklisch T
   4. Bräunlein D
   5. Drawitsch F
   6. Frohnhofen J
   7. Herold T
   8. Otto P
   9. Rzepka N
   10. Werkmeister T
   11. Werner D
   12. Wiese G
   13. Wissler H
   14. Helmstaedter M

   (2017) webKnossos: efficient online 3D data annotation for connectomics

   Nature Methods 14:691–694.

   https://doi.org/10.1038/nmeth.4331

   • PubMed
   • Google Scholar
10. 1. Briggman KL
    2. Helmstaedter M
    3. Denk W

    (2011) Wiring specificity in the direction-selectivity circuit of the retina

    Nature 471:183–188.

    https://doi.org/10.1038/nature09818

    • PubMed
    • Google Scholar
11. 1. Colonnier M

    (1964)

    Experimental degeneration in the cerebral cortex

    Journal of Anatomy 98:47–53.

    • PubMed
    • Google Scholar
12. 1. D'Souza RD
    2. Meier AM
    3. Bista P
    4. Wang Q
    5. Burkhalter A

    (2016) Recruitment of inhibition and excitation across mouse visual cortex depends on the hierarchy of interconnecting areas

    eLife 5:e19332.

    https://doi.org/10.7554/eLife.19332

    • PubMed
    • Google Scholar
13. 1. da Costa NM
    2. Martin KA

    (2011) How thalamus connects to spiny stellate cells in the cat's visual cortex

    Journal of Neuroscience 31:2925–2937.

    https://doi.org/10.1523/JNEUROSCI.5961-10.2011

    • PubMed
    • Google Scholar
14. 1. Denk W
    2. Horstmann H

    (2004) Serial block-face scanning electron microscopy to reconstruct three-dimensional tissue nanostructure

    PLoS Biology 2:e329.

    https://doi.org/10.1371/journal.pbio.0020329

    • PubMed
    • Google Scholar
15. 1. Dorkenwald S
    2. Schubert PJ
    3. Killinger MF
    4. Urban G
    5. Mikula S
    6. Svara F
    7. Kornfeld J

    (2017) Automated synaptic connectivity inference for volume electron microscopy

    Nature Methods 14:435–442.

    https://doi.org/10.1038/nmeth.4206

    • PubMed
    • Google Scholar
16. 1. Faulk WP
    2. Taylor GM

    (1971)

    An immunocolloid method for the electron microscope

    Immunochemistry 8:1081–1083.

    • PubMed
    • Google Scholar
17. 1. Fua P
    2. Knott GW

    (2015) Modeling brain circuitry over a wide range of scales

    Frontiers in Neuroanatomy 9:42.

    https://doi.org/10.3389/fnana.2015.00042

    • PubMed
    • Google Scholar
18. 1. Gala R
    2. Lebrecht D
    3. Sahlender DA
    4. Jorstad A
    5. Knott G
    6. Holtmaat A
    7. Stepanyants A

    (2017) Computer assisted detection of axonal bouton structural plasticity in in vivo time-lapse images

    eLife 6:e29315.

    https://doi.org/10.7554/eLife.29315

    • PubMed
    • Google Scholar
19. 1. Grabenbauer M
    2. Geerts WJ
    3. Fernadez-Rodriguez J
    4. Hoenger A
    5. Koster AJ
    6. Nilsson T

    (2005) Correlative microscopy and electron tomography of GFP through photooxidation

    Nature Methods 2:857–862.

    https://doi.org/10.1038/nmeth806

    • PubMed
    • Google Scholar
20. 1. Gray EG
    2. Hamlyn LH

    (1962)

    Electron microscopy of experimental degeneration in the avian optic tectum

    Journal of Anatomy 96:309–316.

    • PubMed
    • Google Scholar
21. 1. Hamos JE
    2. Van Horn SC
    3. Raczkowski D
    4. Uhlrich DJ
    5. Sherman SM

    (1985) Synaptic connectivity of a local circuit neurone in lateral geniculate nucleus of the cat

    Nature 317:618–621.

    https://doi.org/10.1038/317618a0

    • PubMed
    • Google Scholar
22. 1. Hayworth KJ
    2. Xu CS
    3. Lu Z
    4. Knott GW
    5. Fetter RD
    6. Tapia JC
    7. Lichtman JW
    8. Hess HF

    (2015) Ultrastructurally smooth thick partitioning and volume stitching for large-scale connectomics

    Nature Methods 12:319–322.

    https://doi.org/10.1038/nmeth.3292

    • PubMed
    • Google Scholar
23. 1. Helmstaedter M
    2. Briggman KL
    3. Denk W

    (2011) High-accuracy neurite reconstruction for high-throughput neuroanatomy

    Nature Neuroscience 14:1081–1088.

    https://doi.org/10.1038/nn.2868

    • PubMed
    • Google Scholar
24. 1. Helmstaedter M
    2. Briggman KL
    3. Turaga SC
    4. Jain V
    5. Seung HS
    6. Denk W

    (2013) Connectomic reconstruction of the inner plexiform layer in the mouse retina

    Nature 500:168–174.

    https://doi.org/10.1038/nature12346

    • PubMed
    • Google Scholar
25. 1. Helmstaedter M

    (2013) Cellular-resolution connectomics: challenges of dense neural circuit reconstruction

    Nature Methods 10:501–507.

    https://doi.org/10.1038/nmeth.2476

    • PubMed
    • Google Scholar
26. 1. Hersch SM
    2. White EL

    (1981) Thalamocortical synapses with corticothalamic projection neurons in mouse SmI cortex: electron microscopic demonstration of a monosynaptic feedback loop

    Neuroscience Letters 24:207–210.

    https://doi.org/10.1016/0304-3940(81)90157-9

    • PubMed
    • Google Scholar
27. 1. Heymann JA
    2. Hayles M
    3. Gestmann I
    4. Giannuzzi LA
    5. Lich B
    6. Subramaniam S

    (2006) Site-specific 3D imaging of cells and tissues with a dual beam microscope

    Journal of Structural Biology 155:63–73.

    https://doi.org/10.1016/j.jsb.2006.03.006

    • PubMed
    • Google Scholar
28. 1. Horikawa K
    2. Armstrong WE

    (1988) A versatile means of intracellular labeling: injection of biocytin and its detection with avidin conjugates

    Journal of Neuroscience Methods 25:1–11.

    https://doi.org/10.1016/0165-0270(88)90114-8

    • PubMed
    • Google Scholar
29. 1. Horn BKP

    (1987) Closed-form solution of absolute orientation using unit quaternions

    Journal of the Optical Society of America A 4:629–642.

    https://doi.org/10.1364/JOSAA.4.000629

    • Google Scholar
30. 1. Hua Y
    2. Laserstein P
    3. Helmstaedter M

    (2015) Large-volume en-bloc staining for electron microscopy-based connectomics

    Nature Communications 6:7923.

    https://doi.org/10.1038/ncomms8923

    • PubMed
    • Google Scholar
31. 1. Januszewski M
    2. Kornfeld J
    3. Li PH
    4. Pope A
    5. Blakely T
    6. Lindsey L
    7. Maitin-Shepard J
    8. Tyka M
    9. Denk W
    10. Jain V

    (2018) High-precision automated reconstruction of neurons with flood-filling networks

    Nature Methods 15:605–610.

    https://doi.org/10.1038/s41592-018-0049-4

    • PubMed
    • Google Scholar
32. 1. Joesch M
    2. Mankus D
    3. Yamagata M
    4. Shahbazi A
    5. Schalek R
    6. Suissa-Peleg A
    7. Meister M
    8. Lichtman JW
    9. Scheirer WJ
    10. Sanes JR

    (2016) Reconstruction of genetically identified neurons imaged by serial-section electron microscopy

    eLife 5:e15015.

    https://doi.org/10.7554/eLife.15015

    • PubMed
    • Google Scholar
33. 1. Kasthuri N
    2. Hayworth KJ
    3. Berger DR
    4. Schalek RL
    5. Conchello JA
    6. Knowles-Barley S
    7. Lee D
    8. Vázquez-Reina A
    9. Kaynig V
    10. Jones TR
    11. Roberts M
    12. Morgan JL
    13. Tapia JC
    14. Seung HS
    15. Roncal WG
    16. Vogelstein JT
    17. Burns R
    18. Sussman DL
    19. Priebe CE
    20. Pfister H
    21. Lichtman JW

    (2015) Saturated reconstruction of a volume of neocortex

    Cell 162:648–661.

    https://doi.org/10.1016/j.cell.2015.06.054

    • PubMed
    • Google Scholar
34. 1. Knott G
    2. Marchman H
    3. Wall D
    4. Lich B

    (2008) Serial section scanning electron microscopy of adult brain tissue using focused ion beam milling

    Journal of Neuroscience 28:2959–2964.

    https://doi.org/10.1523/JNEUROSCI.3189-07.2008

    • PubMed
    • Google Scholar
35. 1. Knott GW
    2. Holtmaat A
    3. Trachtenberg JT
    4. Svoboda K
    5. Welker E

    (2009) A protocol for preparing GFP-labeled neurons previously imaged in vivo and in slice preparations for light and electron microscopic analysis

    Nature Protocols 4:1145–1156.

    https://doi.org/10.1038/nprot.2009.114

    • PubMed
    • Google Scholar
36. 1. Lam SS
    2. Martell JD
    3. Kamer KJ
    4. Deerinck TJ
    5. Ellisman MH
    6. Mootha VK
    7. Ting AY

    (2015) Directed evolution of APEX2 for electron microscopy and proximity labeling

    Nature Methods 12:51–54.

    https://doi.org/10.1038/nmeth.3179

    • PubMed
    • Google Scholar
37. 1. Lee S
    2. Wolberg G
    3. Shin SY

    (1997) Scattered data interpolation with multilevel B-splines

    IEEE Transactions on Visualization and Computer Graphics 3:228–244.

    https://doi.org/10.1109/2945.620490

    • Google Scholar
38. 1. Lee WC
    2. Bonin V
    3. Reed M
    4. Graham BJ
    5. Hood G
    6. Glattfelder K
    7. Reid RC

    (2016) Anatomy and function of an excitatory network in the visual cortex

    Nature 532:370–374.

    https://doi.org/10.1038/nature17192

    • PubMed
    • Google Scholar
39. 1. Lin TY
    2. Luo J
    3. Shinomiya K
    4. Ting CY
    5. Lu Z
    6. Meinertzhagen IA
    7. Lee CH

    (2016) Mapping chromatic pathways in the Drosophila visual system

    Journal of Comparative Neurology 524:213–227.

    https://doi.org/10.1002/cne.23857

    • PubMed
    • Google Scholar
40. 1. Liv N
    2. Zonnevylle AC
    3. Narvaez AC
    4. Effting AP
    5. Voorneveld PW
    6. Lucas MS
    7. Hardwick JC
    8. Wepf RA
    9. Kruit P
    10. Hoogenboom JP

    (2013) Simultaneous correlative scanning electron and high-NA fluorescence microscopy

    PLoS One 8:e55707.

    https://doi.org/10.1371/journal.pone.0055707

    • PubMed
    • Google Scholar
41. 1. Livet J
    2. Weissman TA
    3. Kang H
    4. Draft RW
    5. Lu J
    6. Bennis RA
    7. Sanes JR
    8. Lichtman JW

    (2007) Transgenic strategies for combinatorial expression of fluorescent proteins in the nervous system

    Nature 450:56–62.

    https://doi.org/10.1038/nature06293

    • PubMed
    • Google Scholar
42. 1. Maco B
    2. Holtmaat A
    3. Cantoni M
    4. Kreshuk A
    5. Straehle CN
    6. Hamprecht FA
    7. Knott GW

    (2013) Correlative in vivo 2 photon and focused ion beam scanning electron microscopy of cortical neurons

    PLoS One 8:e57405.

    https://doi.org/10.1371/journal.pone.0057405

    • PubMed
    • Google Scholar
43. 1. Mao T
    2. Kusefoglu D
    3. Hooks BM
    4. Huber D
    5. Petreanu L
    6. Svoboda K

    (2011) Long-range neuronal circuits underlying the interaction between sensory and motor cortex

    Neuron 72:111–123.

    https://doi.org/10.1016/j.neuron.2011.07.029

    • PubMed
    • Google Scholar
44. 1. Maranto AR

    (1982) Neuronal mapping: a photooxidation reaction makes Lucifer yellow useful for electron microscopy

    Science 217:953–955.

    https://doi.org/10.1126/science.7112109

    • PubMed
    • Google Scholar
45. 1. Markram H
    2. Lübke J
    3. Frotscher M
    4. Roth A
    5. Sakmann B

    (1997) Physiology and anatomy of synaptic connections between thick tufted pyramidal neurones in the developing rat neocortex

    The Journal of Physiology 500:409–440.

    https://doi.org/10.1113/jphysiol.1997.sp022031

    • PubMed
    • Google Scholar
46. 1. Martell JD
    2. Deerinck TJ
    3. Sancak Y
    4. Poulos TL
    5. Mootha VK
    6. Sosinsky GE
    7. Ellisman MH
    8. Ting AY

    (2012) Engineered ascorbate peroxidase as a genetically encoded reporter for electron microscopy

    Nature Biotechnology 30:1143–1148.

    https://doi.org/10.1038/nbt.2375

    • PubMed
    • Google Scholar
47. 1. Micheva KD
    2. Smith SJ

    (2007) Array tomography: a new tool for imaging the molecular architecture and ultrastructure of neural circuits

    Neuron 55:25–36.

    https://doi.org/10.1016/j.neuron.2007.06.014

    • PubMed
    • Google Scholar
48. 1. Murphy GE
    2. Narayan K
    3. Lowekamp BC
    4. Hartnell LM
    5. Heymann JA
    6. Fu J
    7. Subramaniam S

    (2011) Correlative 3D imaging of whole mammalian cells with light and electron microscopy

    Journal of Structural Biology 176:268–278.

    https://doi.org/10.1016/j.jsb.2011.08.013

    • PubMed
    • Google Scholar
49. 1. Pallotto M
    2. Watkins PV
    3. Fubara B
    4. Singer JH
    5. Briggman KL

    (2015) Extracellular space preservation aids the connectomic analysis of neural circuits

    eLife 4:e08206.

    https://doi.org/10.7554/eLife.08206

    • PubMed
    • Google Scholar
50. 1. Rah JC
    2. Bas E
    3. Colonell J
    4. Mishchenko Y
    5. Karsh B
    6. Fetter RD
    7. Myers EW
    8. Chklovskii DB
    9. Svoboda K
    10. Harris TD
    11. Isaac JT

    (2013) Thalamocortical input onto layer 5 pyramidal neurons measured using quantitative large-scale array tomography

    Frontiers in Neural Circuits 7:177.

    https://doi.org/10.3389/fncir.2013.00177

    • PubMed
    • Google Scholar
51. 1. Rueckert D
    2. Sonoda LI
    3. Hayes C
    4. Hill DL
    5. Leach MO
    6. Hawkes DJ

    (1999) Nonrigid registration using free-form deformations: application to breast MR images

    IEEE Transactions on Medical Imaging 18:712–721.

    https://doi.org/10.1109/42.796284

    • PubMed
    • Google Scholar
52. 1. Saalfeld S
    2. Cardona A
    3. Hartenstein V
    4. Tomancak P

    (2009) CATMAID: collaborative annotation toolkit for massive amounts of image data

    Bioinformatics 25:1984–1986.

    https://doi.org/10.1093/bioinformatics/btp266

    • PubMed
    • Google Scholar
53. 1. Schalek R
    2. Kasthuri N
    3. Hayworth K
    4. Berger D
    5. Tapia J
    6. Morgan J
    7. Turaga S
    8. Fagerholm E
    9. Seung H
    10. Lichtman J

    (2011) Development of High-Throughput, High-Resolution 3D reconstruction of Large-Volume biological tissue using automated tape collection ultramicrotomy and scanning electron microscopy

    Microscopy and Microanalysis 17:966–967.

    https://doi.org/10.1017/S1431927611005708

    • Google Scholar
54. Conference

    1. Serradell E
    2. Glowacki P
    3. Kybic J
    4. Moreno-Noguer F
    5. Fua P

    (2012) Robust non-rigid registration of 2D and 3D graphs

    Paper Presented At: 2012 IEEE Conference on Computer Vision and Pattern Recognition.

    https://doi.org/10.1109/CVPR.2012.6247776

    • Google Scholar
55. 1. Serradell E
    2. Pinheiro MA
    3. Sznitman R
    4. Kybic J
    5. Moreno-Noguer F
    6. Fua P

    (2015) Non-Rigid graph registration using active testing search

    IEEE Transactions on Pattern Analysis and Machine Intelligence 37:625–638.

    https://doi.org/10.1109/TPAMI.2014.2343235

    • PubMed
    • Google Scholar
56. 1. Shahidi R
    2. Williams EA
    3. Conzelmann M
    4. Asadulina A
    5. Verasztó C
    6. Jasek S
    7. Bezares-Calderón LA
    8. Jékely G

    (2015) A serial multiplex immunogold labeling method for identifying peptidergic neurons in connectomes

    eLife 4:e11147.

    https://doi.org/10.7554/eLife.11147

    • PubMed
    • Google Scholar
57. 1. Shu X
    2. Lev-Ram V
    3. Deerinck TJ
    4. Qi Y
    5. Ramko EB
    6. Davidson MW
    7. Jin Y
    8. Ellisman MH
    9. Tsien RY

    (2011) A genetically encoded tag for correlated light and electron microscopy of intact cells, tissues, and organisms

    PLoS Biology 9:e1001041.

    https://doi.org/10.1371/journal.pbio.1001041

    • PubMed
    • Google Scholar
58. 1. Staffler B
    2. Berning M
    3. Boergens KM
    4. Gour A
    5. Smagt PV
    6. Helmstaedter M

    (2017) SynEM, automated synapse detection for connectomics

    eLife 6:e26414.

    https://doi.org/10.7554/eLife.26414

    • PubMed
    • Google Scholar
59. 1. Stepanyants A
    2. Martinez LM
    3. Ferecskó AS
    4. Kisvárday ZF

    (2009) The fractions of short- and long-range connections in the visual cortex

    PNAS 106:3555–3560.

    https://doi.org/10.1073/pnas.0810390106

    • PubMed
    • Google Scholar
60. 1. Tsang TK
    2. Bushong EA
    3. Boassa D
    4. Hu J
    5. Romoli B
    6. Phan S
    7. Dulcis D
    8. Su CY
    9. Ellisman MH

    (2018) High-quality ultrastructural preservation using cryofixation for 3D electron microscopy of genetically labeled tissues

    eLife 7:e35524.

    https://doi.org/10.7554/eLife.35524

    • PubMed
    • Google Scholar
61. 1. Wanner AA
    2. Genoud C
    3. Masudi T
    4. Siksou L
    5. Friedrich RW

    (2016) Dense EM-based reconstruction of the interglomerular projectome in the zebrafish olfactory bulb

    Nature Neuroscience 19:816–825.

    https://doi.org/10.1038/nn.4290

    • PubMed
    • Google Scholar
62. 1. White EL

    (1978) Identified neurons in mouse sml cortex which are postsynaptic to thalamocortical axon terminals: a combined Golgi-electron microscopic and degeneration study

    The Journal of Comparative Neurology 181:627–661.

    https://doi.org/10.1002/cne.901810310

    • PubMed
    • Google Scholar
63. 1. Xu CS
    2. Hayworth KJ
    3. Lu Z
    4. Grob P
    5. Hassan AM
    6. García-Cerdán JG
    7. Niyogi KK
    8. Nogales E
    9. Weinberg RJ
    10. Hess HF

    (2017) Enhanced FIB-SEM systems for large-volume 3D imaging

    eLife 6:e25916.

    https://doi.org/10.7554/eLife.25916

    • PubMed
    • Google Scholar

Author details

1. Florian Drawitsch

   1. Department of Connectomics, Max Planck Institute for Brain Research, Frankfurt, Germany
   2. Donders Institute, Faculty of Science, Radboud University, Nijmegen, Netherlands

   Contribution

   Resources, Data curation, Software, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing—original draft

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0001-9543-1417

2. Ali Karimi

   Department of Connectomics, Max Planck Institute for Brain Research, Frankfurt, Germany

   Contribution

   Data curation, Validation, Writing—review and editing

   Competing interests

   No competing interests declared

3. Kevin M Boergens

   Department of Connectomics, Max Planck Institute for Brain Research, Frankfurt, Germany

   Contribution

   Data curation, Writing—review and editing

   Competing interests

   No competing interests declared

4. Moritz Helmstaedter

   1. Department of Connectomics, Max Planck Institute for Brain Research, Frankfurt, Germany
   2. Donders Institute, Faculty of Science, Radboud University, Nijmegen, Netherlands

   Contribution

   Conceptualization, Investigation, Visualization, Methodology, Writing—original draft, Project administration

   For correspondence

   mh@brain.mpg.de

   Competing interests

   Reviewing editor, eLife

   "This ORCID iD identifies the author of this article:" 0000-0001-7973-0767

• 3,256

  views

• 532

  downloads

• 25

  citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.