ConnectomX

Excited to announce the installation of the first katana microtome @america @WashU , housed in Professor Eduardo Rosa-Molinar’s incredible superlab! We’re absolutely blown away by the cutting-edge instruments available here—it’s a playground for innovation. The introdution of katana microtome opens up a world of possibilities of integrate it with various imaging techniques to create a plethora of corrective imaging workflows that could redefine #volumeEM and #SBFSEM.

A huge thank you to the brilliant engineers from @Hitachi_EM for their support in integrating the katana with the SU7000. That microscope is an absolute powerhouse with its large chamber size and abundance of available ports. We’re thrilled to see how this setup will push the boundaries of what we can achieve.

A fruit fly has no blood vessels in its brain. Oxygen arrives through air tubes. A new preprint traces every one of them in a single adult brain, and they don't join up.

Insects breathe through a tracheal system, gas-filled tubes that run in from spiracles on the body. In the adult fly brain the finest branches, the tracheoles, are under 1 µm across. That is too thin for ordinary fluorescence imaging, and the usual genetic reporters fade in adults. So nobody had a full map.

Cho, Ng, Schoenberg and colleagues at Mount Holyoke College used FAFB, the open serial-section EM volume of a 7-day-old female fly brain that also underlies the FlyWire connectome. Tracers placed skeleton nodes at the centre of every tracheole lumen in CATMAID, and expert reviewers checked each one. To test completeness, two people went through every 250th section of the whole brain and confirmed each lumen had a node. The skeletons were then matched to FlyWire's automatic segmentation. That segmentation had often glued glia and neuron fragments onto the tubes, so those were split off by hand.

The result is 117 tracheoles with 12.98 cm of total length. The same brain holds 149.2 m of neuronal path. They end in 1,376 blind tips inside the tissue and never form a connected network like our capillaries do. Most enter through the optic lobes (63 of 116 entry points), and none enter the central complex. One tracheole sits entirely inside the brain with no entry point at all.

Coverage is tight. 59.3% of points on the network are within 10 µm of a different tracheole, and 91.8% within 20 µm. Volume fraction runs from 0.028% in the lateral horn to 0.14% in the central complex. Left and right hemispheres get the same amount, but the branches are not mirror images. Even after registering each neuropil pair, only about a third of points had a partner within 5 µm on the other side.

Then the synapses. Across 78.3 million presynaptic sites, the median distance to a tracheole was 7.87 µm, and 97.3% were within 20 µm. But compared with non-synaptic stretches of the same neuron, synapses sat slightly farther away, a median 0.36 µm, and only within about 15 µm of a tube. The authors suggest mitochondria along the axon may set where oxygen is used, and glia hugging the tubes may keep synapses back.

Limits: one brain, one sex. The air sacs were removed in dissection, so origins outside the brain are inferred. The 386 odd membrane pinches they call tracheole link connectors, and the isolated tracheole, could be fixation artifacts. Glia aren't segmented in FAFB, so that idea can't be tested here yet.

What I like is the reuse. A volume imaged for neurons held a second organ system that the segmentation treated as clutter. The every-250th-section audit is a cheap completeness check anyone tracing a sparse structure in their own volume could copy.

Cho H, et al. A complete adult Drosophila melanogaster brain tracheolar network. bioRxiv 2026. doi.org/10.64898/2026.…

#vEM #connectomics

How many axons are hiding in an OCT scan? A new monkey glaucoma study used serial block-face SEM to count them.

Clinics track glaucoma by measuring retinal nerve fibre layer (RNFL) thickness around the optic disc with OCT. But a micron of RNFL does not hold the same number of axons everywhere. Axon size changes around the disc, and glia and blood vessels take up space too. Kharal, Sullivan and Patel at the University of Houston set out to measure both, then turn OCT thickness into an axon count.

They induced glaucoma in one eye of 22 rhesus monkeys with laser scarring of the drainage angle and followed them with OCT. In five animals they cut roughly 0.75 x 0.75 mm pieces of retina from six spots around the optic disc, in both eyes, and gave them a heavy-metal stain (osmium, thiocarbohydrazide, uranyl, lead). The blocks went into a Tescan MIRA3 SEM with a Gatan 3View 2. They sectioned to about halfway through the tissue, roughly where the OCT scan path sits, and took a single 8192 x 8192 backscatter image of the block face at 9.2 to 30.5 nm per pixel and 7 to 12 kV.

They traced 431,768 axons by hand. Mean axon area was 0.46 µm² in control eyes and 0.43 µm² in glaucoma eyes, with no significant difference. Location mattered more. Temporal axons were the smallest (0.28 µm² in controls) and superior nasal axons the largest (0.66 µm²).

The non-axonal part was the surprise. As a share of the layer it rose from 28% in controls to 45% in glaucoma eyes. In absolute area it barely moved (1592 vs 1441 µm² in a 60 µm strip). The disease takes away axons and leaves the glia and vessels behind.

The OCT data agree. Plotting RNFL thickness against optic nerve axon counts, the line hits zero axons at 31.4 µm of thickness, about 28% of a normal layer. The slope works out to an average axon of 0.43 µm², matching the histology. A six-sector model built on the EM numbers predicted total optic nerve axon counts with a slope of 0.99 and a bias of about 9,500 axons, against counts of around 1.44 million in healthy eyes.

So a thin RNFL is not all lost axons. There is a floor of other tissue, and it differs by region. The authors expect a location-aware model to tie OCT more closely to visual field loss.

The authors list the limits. Only five animals went through EM, the laser model progresses faster than human glaucoma, the images can't split astrocytes from microglia and capillaries, and the continuous model has no independent validation set yet.

What strikes us is that they ran an SBF-SEM mostly for single faces. One check went deeper. They imaged three depths about 10 µm apart, 94,155 axons in all, and mean size stayed put (0.39, 0.39, 0.38 µm²). A full serial stack of the same blocks could separate the glia from the vessels, follow axons through the layer and remove the oblique-cut images they had to drop (6 of 60).

Kharal A, Sullivan M, Patel NB. Axon Content of the Circumpapillary Nerve Fiber Layer in Experimental Glaucoma. IOVS 2026;67(13):8. doi.org/10.1167/iovs.6…

#VolumeEM

Every big connectome ends with the same chore. The segmentation has glued two neurons together, a proofreader clicks near the bad join, and the software has to decide where to cut. A new arXiv preprint from Notre Dame and Korea University looks hard at that step.

Byunghyun Ko, Won-Ki Jeong and Chaoli Wang call their method ContactCut. Their point is that a click doesn't define the split on its own. The segmentation is stored as a graph of supervoxels, and every place where two supervoxels share voxel faces is a contact. Within 500 nm of the click there are usually several contacts that could act as the two seeds for a min-cut, and each gives a different split of the whole object, even with the same solver.

So instead of seeding at the nearest contact, ContactCut tries every nearby contact, runs a seeded minimum cut on the full graph for each one, and scores the resulting partitions. The scoring looks only at structure, never at the EM images. It prefers splits where both sides stay in one connected piece, where the smaller side isn't a tiny crumb, and where the cut runs through a narrow connection. A ranker with two learned weights balances the last two. A second mode tries pairs of contacts and offers a proofreader five alternatives.

They tested it on real edits that proofreaders made in MICrONS minnie65 (mouse visual cortex) and FlyWire (adult fly brain). On 733 MICrONS test edits the main score was 0.657, against 0.598 for the best graph-based method and 0.532 for the best learned one. On 2,929 FlyWire edits it scored 0.641, against 0.608 and 0.541. A ranker trained only on MICrONS scored 0.6405 on FlyWire, almost the same as 0.6414 when trained on FlyWire itself. It runs on a CPU, with a median of 0.30 s per edit on MICrONS and 0.11 s on FlyWire.

It matters because false merges are the expensive error. One wrong fusion corrupts the shape and wiring of several neurons.

There are limits. The reference is the edit people actually made, so this measures agreement with past proofreading, not verified neuron identity. It only handles two-way splits. A score around 0.65 still leaves plenty for a human to check, and the best single-contact cut available scores 0.843 on MICrONS, so the ranker doesn't yet find the best option every time. It's also a preprint, not yet peer reviewed.

The part I like most is that it ignores the images completely. It works on the shape of the supervoxel graph alone, which is probably why a model trained on mouse cortex carries over to the fly brain with almost no loss. The flip side is that the fix can only be as good as the supervoxels. If a section is damaged or misaligned there may be no clean contact to cut through. Clean, evenly cut volumes at acquisition still save time later.

Ref: Ko, Jeong and Wang (2026), "ContactCut: A Partition-Aware Contact Selection Framework for False-Merge Correction in Connectomics", arXiv:2611.01457. arxiv.org/abs/2611.01457 Code: github.com/bentake/Contac…

#connectomics

Vimentin filaments move through the cell along microtubules, and they need kinesin-1 and dynein to do it. Yet a 3D FIB-SEM study from the Gelfand and Lippincott-Schwartz labs found the filaments sit on average 233 nm from the nearest microtubule, and over 85% of microtubule length is more than 50 nm from any vimentin filament. So how does a filament use a track it barely touches?

A new bioRxiv preprint from the same groups at Northwestern and Janelia tests a simple idea. The filaments hitchhike on mitochondria, which already carry the motors.

The volume EM work reuses an existing dataset. They went back to a FIB-SEM volume of the perinuclear region of a COS-7 cell, acquired at 2 nm isotropic sampling by Janelia's CellMap team, in which 572 vimentin filaments and 39 microtubules had already been traced by hand. They added the surfaces of 15 mitochondrial profiles from an 8 nm segmentation, modelled the filaments as tubes of 6 nm and 12 nm radius, and measured the gaps in both directions.

In that volume, 4.61% of the mitochondrial surface lay within 50 nm of a vimentin filament, against 1.15% within 50 nm of a microtubule. Single filaments curve along a mitochondrion at sites that only partly overlap with where microtubules touch it. The two are not simply stuck to the same track.

The causal tests are light microscopy. Knocking down TRAK1, the adaptor that links mitochondria to kinesin-1, moved mitochondria toward the nucleus in RPE cells, with mean radial position falling from 0.52 to 0.40. Vimentin moved with them, from 0.56 to 0.40, and its peripheral fraction fell from 36.0% to 13.5%. Microtubules stayed spread out. Then, in cells with no kinesin-1, they tethered a tail-less kinesin motor to the mitochondrial outer membrane. That motor can't bind vimentin. Mitochondria moved outward, 0.34 to 0.60, and vimentin followed, 0.34 to 0.51. Pulling mitochondria inward with a plant minus-end motor did the reverse.

Hitchhiking is already known for peroxisomes, lipid droplets, ER and RNA granules. This adds a cytoskeletal filament to that list.

The authors are open about the limits. The EM is one cell, one volume and 15 mitochondrial profiles, and the contact fractions are small. Not all vimentin follows. When mitochondria were dragged inward, part of the peripheral network stayed put, so they propose a hybrid model with other routes as well. The physical linker, plectin 1b or the vimentin head domain, is still open.

What I like here is that the puzzle itself came from volume EM. Light microscopy can't measure a 50 nm gap. A well-traced 3D volume showed the paradox, and the same volume, reopened with a new question, pointed at the vehicle.

Ref: Renganathan, Moore, Adam, Lippincott-Schwartz, Gelfand. Mitochondria-coupled hitchhiking drives vimentin intermediate filament transport. bioRxiv, posted 4 Nov 2026. doi.org/10.64898/2026.…

#VolumeEM