GCIB-SEM: A path to 10 nm isotropic imaging of cubic millimeter volumes

Focused Ion Beam Scanning Electron Microscopy (FIB-SEM) generates 3D datasets optimally suited for segmentation of cell ultrastructure and automated connectome tracing but is limited to small fields of view and is therefore incompatible with the new generation of ultrafast multibeam SEMs. In contrast, section-based techniques are multibeam-compatible but are limited in z-resolution making automatic segmentation of cellular ultrastructure difficult. Here we demonstrate a novel 3D electron microscopy technique, Gas Cluster Ion Beam SEM (GCIB-SEM), in which top-down, wide-area ion milling is performed on a series of thick sections, acquiring < 10 nm isotropic datasets of each which are then stitched together to span the full sectioned volume. Based on our results, incorporating GCIB-SEM into existing single beam and multibeam SEM workflows should be straightforward and should dramatically increase reliability while simultaneously improving z-resolution by a factor of 3 or more.

sectioning reliability while simultaneously improving z-resolution and maintaining compatibility with MultiSEM imaging.
In FIB-SEM a tightly focused (~1 µm) beam of high-energy (30 kV) gallium ions is directed at an almost parallel angle (< 1 o ) to the surface of a tissue block, ablating material while avoiding deeper damage. Repeated FIB/SEM cycles produce a 3D stack, but milling artifacts accumulate in the direction of the FIB beam due to the shallow milling angle used, and these limit the area which can be imaged 13 . We reasoned that replacing FIB with a more 'top-down' approach might overcome this, but the energy of the ions would have to be lowered considerably. We therefore decided to try GCIB milling, a lowenergy surface polishing method used in semiconductor fabrication and mass spectroscopy 15 . In GCIB, clusters of argon atoms are ionized and accelerated while a magnetic field selects clusters containing a particular number of atoms. In this way the average energy per atom can be tuned to just a few electron volts. Molecular dynamics simulations 16 show that such clusters 'splat' against the surface, scattering atoms at glancing angles thereby removing asperities.
Our prototype system consists of a GCIB gun (GCIB-10s from Ionoptika) mounted to a single beam SEM (Ultra SEM from Zeiss) (Supplementary Fig. 1). We first tested milling a variety of ultrathin tissue sections, exploring a range of angles, cluster sizes and energies, and embedding resins. We SEM imaged milled surfaces with 1.2 kV landing energy (chosen to provide adequate z-resolution) using energy selective backscatter (ESB) and in-lens secondary electron (InLens-SE) detectors. ESB (500 V filtering, see Methods) produced high quality images under virtually all milling conditions, but InLens-SE showed rough surfaces in some. InLens-SE detection is much more sensitive to surface topography and charging than ESB detection but approximates MultiSEM imaging 11 . Epon samples showed rough InLens images under all but the shallowest milling angles. Durcupan and Sprurr's samples gave high-quality InLens-SE images at glancing angles up to 30 o (Supplementary Fig. 2).
Sections thicker than 100 nm charged during imaging, and GCIB milling did not correct this (unlike FIB-SEM). Charging lessened with extended imaging and we reasoned that this was due to the slow conversion of the polymer embedding into a more conductive amorphous carbon 17 . We subsequently demonstrated that such electron irradiation could be used to make sections at least 10 µm thick sufficiently conductive to allow quality InLens-SE imaging (Supplementary Fig. 3,4). We found that a dose of ~0.5x10 27 eV/cm 3 was needed to completely eliminate charging. GCIB milling of irradiated sections remained sufficiently smooth to produce 3D volume images but the sputter removal rate of irradiated regions dropped by up to a factor of six (Supplementary Fig. 5). Our current 'optimized' protocol recommends irradiating to a dose of 0.5x10 27 eV/cm 3 , then milling with 10 kV Ar2000 clusters at a glancing angle of 30 o while the sample is rotating. Fig. 1a outlines this process.
To test, we sectioned a Durcupan-embedded Drosophila brain and collected three sequential 1 µm sections onto gold-coated silicon. We performed ~250 GCIB/SEM cycles (using InLens-SE detection) completely milling through all three sections (6x6x4 nm voxels). Fig. 1b shows images of one section at 200 nm intervals. Fig. 1c shows a Z-reslice through the image stacks of all three sections showing an unevenness at the bottoms of each section, the result of milling rate variance (~10%). We choose to collect on gold-coated silicon because the InLens-SE and ESB signals for gold are beyond the maximum signals produced by heavy-metal stain. Remarkably this 10% milling variance is sufficiently small and sufficiently spatially uniform that these variances could be corrected by software. We wrote algorithms to find and 'flatten' this unevenness (see Methods). Once flattened (Fig. 1d), we stitched the sections into a single volume (Fig. 1e). Membranes and synapses are clear and resemble those we typically see using FIB-SEM. A close inspection (Supplementary Video 1) shows some slight artifacts relative to FIB-SEM, particularly a light texture of milling streaks is visible in the Z-reslice.
This sample was embedded in Durcupan resin which appears to be insufficiently resilient to be sectioned this thick (diamond cut surface showed some damage), so we have now switched to Spurr's embedding which can be room-temperature sectioned with high quality surfaces at up to 1 µm, and can be hot knife sectioned at up to 25 µm. Fig. 2a shows a Z-reslice of a GCIB-SEM volume (using InLens-SE detection) covering three sequential 500 nm thick sections of Spurr's-embedded mouse cortex (8x8x6 nm voxels). Ultrastructural features like post synaptic densities and vesicles are clearly defined. Fig. 2b shows the simultaneously-acquired ESB volume which looks almost identical under these milling conditions. Cut surfaces were smooth and were easily flattened and stitched with an estimated loss between sections of ~30 nm (see Methods), and visual tracing of all neuronal processes appeared straightforward (Fig. 2c, Supplementary Video 2). These datasets demonstrate that wide-area GCIB milling and InLens-SE detection can be performed over multiple sequential thick sections and can produce datasets approaching the quality of FIB-SEM.
To address the applicability of GCIB-SEM to mammalian connectomics we sectioned Spurr'sembedded mouse cortex at 1 µm and collected ten sequential sections, then performed ~170 GCIB/SEM cycles (8x8x6 nm voxels) completely milling through all ten sections. We computationally flattened each section and stitched them together into the single volume shown in Fig. 2d, Supplementary Fig. 6-11, and Supplementary Video 3. The cut surfaces of these Spurr's-embedded sections were of high quality and we estimated the loss between sections at ~30 nm. A test segmentation of the ESB stack was performed using a flood-filling neural network 14 that had previously been trained on a quite different SBEM dataset and then 'bootstrap-trained' on this dataset (see Methods). Fig. 2e shows an example spiny dendrite spanning the ten section GCIB-SEM volume along with synapsing axons. Fig. 2f shows views through a single spiny synapse.
Cutting at 1 µm appears to be approaching the limit for room temperature sectioning of Spurr's since we saw significant surface damage when we attempted sectioning at 2 µm. Cutting much thinner (e.g. 150 nm) resulted in less loss and higher quality surfaces (see Methods), which may be preferred despite the fact that it results in more stitch boundaries. An alternative is hot knife ultrathick sectioning which uses a heated diamond knife and oil lubrication to minimize stresses on the tissue 18 .
To test this, we 'hot knife' sectioned Spurr's-embedded mouse cortex at 10 µm thickness and flat embedded two sequential sections against gold-coated silicon. The resulting GCIB-SEM volume (10x10x12 nm voxels) spanning the two hot knife sections is shown in Fig. 3a and a test segmentation based on the InLens-SE stack using a flood-filling network is shown in Fig. 3b. Supplementary Figs. 12-16 show Z-reslice views of the GCIB-SEM stack spanning the two 10 µm thick sections before and after flattening and stitching as well as additional segmentation examples. The cut surfaces were of high quality but the flat embedding procedure resulted in some contamination. We estimated the loss between sections at ~30 nm (see Methods). Supplementary Video 4 shows a side-by-side comparison of the ESB vs. InLens-SE volumes (also Fig. 2c,d).
We noticed that GCIB milling produces a slight drop in the InLens-SE membrane contrast which progresses over the first ~500 nm of milling. We speculate this is the result of a few nm thick layer of disrupted surface material persisting in equilibrium with milling which changes SE yield. Initial electron irradiation also produces a slight drop in InLens-SE contrast. These effects are quantified in Supplementary Fig. 17. We also verified that GCIB-SEM imaging was generally compatible with ATUM tape collection 9 by imaging two 1 µm thick sections collected on copper coated tape ( Supplementary  Fig. 18).
Finally, we verified that the InLens-SE detection used in our prototype was an adequate proxy for MultiSEM imaging. We GCIB-SEM milled and imaged the first 250 nm of a 500 nm section of mouse cortex then shipped the half-milled section to Zeiss for imaging on their MultiSEM system. MultiSEM images of the GCIB-milled surface (acquired with similar electron dose) were essentially identical to our InLens-SE images of the same surface (Supplementary Figs. 19,20).
We see Serial Thick-Section GCIB-SEM as a potential route to 'industrial-scale' 3D EM and connectomics as it addresses key limitations of existing techniques: GCIB-SEM resolution is not limited by section thickness, and imaging dose is not limited by interactions with sectioning 19 , so resolution and dose can be adjusted as needed to achieve fully automated segmentation. GCIB-SEM utilizes thick sectioning which should be reliable even over arbitrarily large volumes. GCIB is dramatically less sensitive to beam position and focus than FIB, allowing lossless restarts. The gas cluster source, unlike FIB, does not require periodic reheating and is comparatively simple and reliable. GCIB-milling is widearea and fast (up to 450 µm 3 /s, 1 mm 3 /month, see Methods), and, if necessary, can be performed across multiple GCIBs in parallel with imaging when sections are spread across multiple wafers. This should allow close to 100% utilization of MultiSEM imaging time and should allow projected 20 MultiSEM improvements to be taken full advantage of. We envision a day when acquiring automatically segmented EM volumes of large tissue samples (e.g. whole invertebrate and small vertebrate brains) is routine; where such volumes are reliably thick sectioned and spread across multiple wafers which are then robotically shuttled between multiple GCIB-milling and MultiSEM-imaging stations in a setting reminiscent of today's semiconductor fabrication plants.

Online Methods: Sample preparation:
All experimental protocols were conducted according to US National Institutes of Health guidelines for animal research and were approved by the Institutional Animal Care and Use Committee at Janelia Research Campus. A three-month-old adult C57/bl6 mouse was euthanized with an overdose of isoflurane and immediately perfused via the heart with a buffered solution of 1.25 % glutaraldehyde and 4 % paraformaldehyde (0.1M PB pH 7.4). Two hours after perfusion the brain was removed and left in the same fix overnight at 4°C. Coronal sections of the brain were then cut with a vibratome (Leica VT1200). These were stained with 1.5% potassium ferrocyanide, on ice, followed by 2% osmium, each diluted in phosphate buffer. Sections were stained in 1% thiocarbohydrazide for 20 minutes before transferring them into 2% osmium tetroxide (Agar Scientific) for a further 30 minutes. They were then placed in 1% uranyl acetate at 4 o C overnight then in lead aspartate solution at 50 o C for 2 hours. The sections were finally dehydrated in increasing concentrations of alcohol, for 5 minutes each change and then transferred to increasing concentrations of epoxy resin (Spurs, EMS) until 100%. These were placed between two glass slides, coated with mold separating agent (Glorex, Switzerland), and hardened at 60 ° C for 24 h.
A six-day-old male adult Canton S G1 x w1118 fly of the Drosophila was GCIB-SEM imaged producing the micrographs shown in Fig. 1b-e. The preparation method used was a heavy metal contrast enhancement on C-PLT, an optimized preparation protocol for Drosophila brains with optimal contrast and morphological preservation for FIB-SEM. Chemical/ Progressive Lowering of Temperature (C-PLT) fixation/dehydration with en bloc staining is a modified conventional chemical fixation method 13,18 . Isolated whole brains were fixed in 2.5% formaldehyde and 2.5% glutaraldehyde in 0.1 M phosphate buffer at pH 7.4 for 2 hours at 22 o C. After washing, the tissues were post-fixed in 0.5% osmium tetroxide in 0.05M sodium cacodylate buffer for 40 min and then treated with 0.8% potassium ferricyanide in buffer for 2 hours at 4 °C. After washing, tissue was incubated with 0.5% aqueous uranyl acetate for 30 min at 4 o C then followed by lead aspartate en bloc staining at 4 o C for overnight. A PLT procedure started from 1 o C when the tissues were transferred into 10% ethanol after secondary osmication with 0.8% osmium tetroxide for 20 min at 4 o C. The temperature was progressively decreased to −25 o C while the ethanol concentration was gradually increased to 97%. The tissue was incubated in 1% osmium tetroxide and 0.2% uranyl acetate in ethanol for 32 hours at −25 °C. After PLT and low temperature incubation, the temperature was increased to 22 o C, and tissues were rinsed in pure ethanol following by acetone, then infiltrated and embedded in Durcupan (ACM Fluka).

Sectioning, Irradiation, Milling and Imaging:
Drosophila brain (Fig. 1b-e): We sectioned a Durcupan-embedded Drosophila brain at 1 µm thickness on an ultramicrotome (Leica UC7) using a room temperature diamond knife (Diatome, 35 o clearance angle) and collected three sequential sections from the water boat onto a gold-coated silicon wafer. Silicon wafers used for this and other runs were pre-diced 5 x 7 mm p-type silicon support chips (Ted Pella). We deposited 5 nm of chromium (to promote gold adhesion) followed by 50 nm of gold onto the silicon surface using a Precision Etching and Coating System (Gatan). We then glow-discharged the gold surface just prior to collection to promote wetting. Electron irradiation parameters used: 10,000 μm 2 area, 2 hrs at 6 kV, 4 hrs at 10 kV, 1x10 27 eV/cm 3 total dose. Electron irradiation for this and other samples was performed using the SEM's beam set to its highest aperture, defocused and scanned across the area to be irradiated. GCIB parameters used: 22nA of Ar2000 at 10 kV spread over a 10 mm 2 area, 21 o glancing angle, 360 o rotation (3 steps), 900 s per mill cycle giving 4 nm of removal per image (44 µm 3 /s milling rate). GCIB milling for this and other samples was performed by slightly defocusing the GCIB beam (~0.5 mm) and scanning the area to be milled with a 50 x 50 position grid. SEM parameters used: 1.2 kV, 2 nA electron beam, InLens-SE detection, 6 nm pixels, 2 MHz acquisition rate.
Mouse cortex (Fig 2d-f) : We sectioned Spurr's embedded mouse cortex at 1 µm thickness on an ultramicrotome (Leica UC7) using a room temperature diamond knife (Diatome, 35 o clearance angle) and collected ten sequential sections onto a gold-coated silicon wafer. Electron irradiation parameters used: 10,000 μm 2 area, 1.6 hrs at 8kV, 2.4x10 26 eV/cm 3 total dose. GCIB parameters used: 22nA of Ar2000 at 10kV spread over a 18 mm 2 area, 30 o glancing angle, 360 o rotation (3 steps), 240s per mill cycle giving 6nm of removal per image (450 µm 3 /s milling rate). SEM parameters used: 1.2kV, 2nA, simultaneous ESB (500 V filtering) and InLens-SE detection, 8nm pixels, 1.25MHz acquisition rate. Fig 3): We 'hot knife' sectioned Spurr's embedded mouse cortex at 10 µm thickness using the method described in (Hayworth et al. 2015) 18 . Hot knife parameters used: Diamond knife (Diatome cryo, 25 o clearance angle) lubricated with 0.22 µm filtered tapping oil (Master Plumber), 65 o C knife temperature, no oscillation, 0.1 mm/s cutting speed. Two sequential sections were collected, washed of oil by dipping repeatedly in uncured Durcupan resin, and flat embedded against a gold coated silicon wafer (while still covered in uncured Durcupan) by sandwiching sections between the wafer and a glass optical flat that had a Kapton release film taped over it. This sandwich was put in a 65 o C oven to cure under a weight. This flat embedding procedure resulted in the two thick sections having a thin covering of Durcupan that had to be GCIB-milled to reveal their tissue surfaces (see text). A scalpel and laser ablation were used to remove large regions of the surrounding Durcupan 'flash' to prevent peripheral charging, then matching regions in each thick section were electron irradiated to make them conductive. Electron irradiation parameters used: 110,000 μm 2 area, 117 hrs at 30 kV, 0.9x10 27 eV/cm 3 total dose. GCIB parameters used: 22nA of Ar2000 at 10kV spread over a 13 mm 2 area, 30 o glancing angle, 360 o rotation (6 steps), 360 s per mill cycle giving 12 nm of removal per image (400 µm 3 /s milling rate). SEM parameters used: 1.2 kV, 2 nA electron beam, simultaneous ESB (500 V filtering) and InLens-SE detection, 10 nm pixels, 1.25 MHz acquisition rate. (MultiSEM test sample, Supplementary Figs. 19, 20): We sectioned Spurr's embedded mouse cortex at 500 nm thickness on an ultramicrotome (Leica UC7) using a room temperature diamond knife (Diatome, 35 o clearance angle) and collected a single section from the water boat onto a gold-coated silicon wafer. A region was electron irradiated to make it conductive, then GCIB-SEM imaged to a depth of ~250 nm before shipping to Zeiss for MultiSEM imaging tests. Electron irradiation parameters used: 188,000 μm 2 area, 89 hrs at 6 kV, 0.9x10 27 eV/cm 3 total dose. GCIB parameters used: 32 nA of Ar2000 at 10kV spread over a 8 mm 2 area, 30 o glancing angle, 360 o rotation (3 steps), 360 s per mill cycle giving ~12 nm of removal per image. SEM parameters used for GCIB-SEM imaging on Janelia Ultra SEM: 1.2 kV, 2 nA electron beam, InLens-SE detection, 8 nm pixels, 1.25 MHz acquisition rate.

Computational flattening and stitching of sections:
Processing of all GCIB-SEM runs proceeded in three steps (overviewed in Fig. 1): 1.) Align the SEM images of each thick section individually to generate a set of 3D volumes, one for each thick section. 2.) Computationally flatten the bottoms of each section individually. 3.) Align and stitch these 3D volumes together to form a final spanning 3D volume. We used the SIFT and BUnwarpJ alignment tools in the Fiji software 21 to perform steps #1 and #3. For step #2 we wrote a custom Matlab script to perform flattening.
To perform flattening we first identified the grayscale threshold that best separated tissue pixels from substrate pixels, which could be reliably distinguished by grayscale as long as the thick sections were collected on a gold substrate. Our flattening script used this grayscale threshold to determine, for each x,y pixel position, the image index (z-position) where the tissue broke through to the substrate, creating a "bottom index map". This "bottom index map" was lightly smoothed spatially and then used to stretch (with linear interpolation) each column of pixels in the z-direction so that all substrate breakthroughs would occur in a single z-plane.
We have provided this custom Matlab flattening script in the Supplementary Software along with an example dataset and a brief manual describing its operation.

Section-to-section loss estimation:
We estimated section-to-section tissue losses for each of the GCIB-SEM runs described in the text as well as for a test run performed with a series of 150 nm thick sections. Supplementary Figs. 21-25 summarize these results.
Our method for estimating section-to-section loss was as follows: First the z-thickness of each GCIB-SEM mill step was estimated by dividing the microtome-set cutting thickness by the number of mill steps spanning a thick section. Then each image in the final stitched GCIB-SEM stack covering multiple serial thick sections was first spatially filtered (using FIJI) with a 2D Gaussian (r = 4 pixels), and a matrix of mean pixel-wise image differences (absolute valued) spanning the full GCIB-SEM volume was computed using a custom Matlab script. We used this to plot the expected image difference vs. mill steps (shown in each supplementary figure). The images right next to the cut surfaces display differences in contrast and features relative to images in the interior of a stack. For example, the first few images of each thick section often have somewhat higher contrast which diminishes after milling (see Supplementary Fig. 17), and the last few images of a flattened stack show contrast and texture artifacts from the flattening algorithm where pixel grayscale values start to blend with the gold substrate. Looking at these 'boundary' images makes clear that they contain information crucial for tracing but directly computing the image difference between such 'top' and 'bottom' images would lead to an erroneously large estimated loss dominated by such contrast and flattening-artifact effects. Instead we picked images slightly removed from the boundary to compare and computed the loss based on these.
For example, the GCIB-SEM dataset shown in Fig. 2a-c spanned three 500 nm microtome cut sections. The flattened stack for the first thick section from this run contained 81 images giving an estimated mill step size of 500 nm / 81 = 6.2 nm. The computed image difference matrix is shown in SOM Fig. 22 along with a plot of the expected image difference vs. mill steps derived from looking at only 'interior' images. The boundary between the second and third thick section occurred between image #162 and #163 in this stack, but image #162 was not sufficiently clean for direct comparison due to the contrast and flattening artifacts described above. Instead we backed up to image #159 (which was clean) and compared it to image #163. The difference between these images (according to the computed mean difference matrix) was 14.5 units which corresponded to 8.7 mill steps according to the plot. This implies that there is an equivalent of 8.7 -3 = 5.7 mill steps between #162 and #163. But if there was no loss of tissue during sectioning then there should only be 1 mill step between #162 and #163. So the loss is equivalent to 5.7 -1 = 4.7 mill steps or 4.7 * 6.2 nm = 29 nm.
As described, the flattening procedure results in boundary images containing artifacts and, depending circumstances (for example, how sensitive one's segmentation algorithm is to noise), one may decide to discard one or more of the boundary images. To account for this Supplementary Figs. 21-25 display both the boundary image we based our loss estimate on and the next one that would be used if that one was discarded along with loss estimates for both.

Flood-filling network segmentation:
As described in the text, test segmentations were performed on an ESB-detected GCIB-SEM volume spanning ten 1 µm sections of mouse cortex which was acquired with 8x8x6 nm voxels ( Fig. 2d-f,  Supplementary Figs. 6-11) and on an InLens-SE-detected GCIB-SEM volume spanning two 10 µm 'hot knife' sections of mouse cortex which was acquired with 10x10x12 nm voxels (Fig. 3, Supplementary  Figs. 12-16). The two GCIB-SEM stacks were segmented with flood-filling networks (FFN) 14 . Instead of producing de novo dedicated volumetric training data for these GCIB-SEM volumes, an existing model trained on 10x10x25 nm SBEM data of zebra finch Area X tissue (provided by Jorgen Kornfeld) was used to bootstrap the segmentation, and object-based proofreading was utilized to correct errors as described next.
Using the ESB-detected GCIB-SEM volume spanning ten 1 µm sections, a prototype variant of the CycleGAN-based transfer procedure described in (Januszewski & Jain 2019) 22 was first applied between (a) 20x20x25 nm SBEM and 16x16x24 nm GCIB-SEM data (downsampled from its native 8x8x6 nm acquisition resolution), as well as 10x10x25 nm SBEM and (b) 16x16x12 nm, and (c) 8x8x18 nm GCIB-SEM data. Reduced-resolution downsampled variants of the GCIB-SEM volume were obtained via areaaveraged resampling of the full resolution data.
Segmentations obtained for multiple CycleGAN checkpoints from the (a) and (b) variants were screened for merge errors, and combined with oversegmentation consensus after upsampling to a common resolution. 12 neurite fragments were manually assembled from correct fragments in the base segmentation. These fragments were then used to train an FFN model for 16x16x12 nm GCIB-SEM data, with loss masking in 2 images before and after every 1 μm thick section seam. The segmentation from this new model correctly resolved the larger structures in the volume, but the model could not track finer processes due to insufficient downsample resolution of input data and a lack of representative training examples. To compensate for that, oversegmentation consensus between the FFN-generated reduced resolution segmentation, and the full resolution segmentation from SBEM-transfer (see (c) above) was computed, an additional 13 neurites were reconstructed by manual fragment assembly to form the training set for a full-resolution GCIB FFN model. Once trained, a segmentation was generated, 450 objects were proofread and selected for another round of training. All objects larger than 100k voxels were inspected in the final segmentation volume and split errors were manually corrected where necessary.
A similar procedure was applied to segment the second GCIB-SEM volume (InLens-SE-detected GCIB-SEM volume spanning two 10 µm 'hot knife' sections, acquired at 10x10x12 nm voxel resolution). First, the network trained on the first GCIB-SEM volume at 8x8x6 nm resolution was transferred to the InLens-SE 10x10x12 nm stack with the help of a CycleGAN. No attempt was made to correct the mismatched voxel resolution.
Candidate segmentations were generated for multiple checkpoints of the CycleGAN using forward-backward oversegmentation consensus 14 . The following iterative procedure was then applied to identify the best checkpoint. Starting with a random checkpoint, 3D meshes of objects in the segmentation were screened for mergers in descending order of object voxel count until 10 merge locations were found and added to a list of known mergers. The segmentations were then evaluated using all recorded merge locations, and the segmentation with the fewest mergers was screened again. Screening was terminated once a single segmentation without known errors remained, at which point 61 distinct merge errors were recorded. FFN agglomeration with standard settings 14 was then applied to the selected segmentation, with all merge pairs from agglomeration subjected to manual review. All objects larger than 100k voxels after agglomeration and touching 0 or 1 faces of the volume were then manually inspected, and if necessary connected to other objects within the volume. For the final segmentation, an FFN model was retrained on this proofread segmentation with the voxelwise loss computed only over voxels belonging to objects larger than 100k voxels.
The FFN network architecture and training procedures described in (Januszewski et al. 2018) 14 were used everywhere, but the field of view was extended to (33, 33, 33) voxels to account for the more isotropic resolution of the GCIB-SEM data, and the depth of the network for the second GCIB-SEM volume was extended to 12 residual modules. The manual inspections performed did not reveal any objects that could not be unambiguously traced. No manual voxel-level corrections were performed at any step of the reconstruction.     Ar2000, 10kV). This is a time-of-flight plot made by pulsing the GCIB source while the beam is hitting a sample mounted at a known distance. The GCIB was tuned to give clusters containing an average of 2000 argon atoms each, but the plot shows that this beam actually contains a range of cluster sizes from mainly within the range of Ar1000 to Ar3000.