A cubic millimetre of human brain

The sample

A woman of forty five had epilepsy that medicine could not control. The focus was in her hippocampus, and to reach it the surgeons had to remove the piece of cortex sitting on top of it. That piece would otherwise have been discarded. Instead it was fixed within minutes, stained with heavy metals, set in resin, and cut into 5,019 slices averaging 33.9 nanometres thick. Every slice was photographed by a multibeam electron microscope at four nanometres per pixel.

The block came from the front of the middle temporal gyrus and runs the full depth of the cortex, from the surface down into the white matter. A neuropathologist examined it and found it normal. It is just over one cubic millimetre, about the size of a grain of coarse sand, and photographing it produced 1.4 petabytes.

That is the sample. What is on this page is what was reconstructed from it: which is to say, roughly one part in a million of one human brain.

Most of it is not neurons

The release ships a table of every cell body in the volume. It has 49,379 rows. Sorted by what each cell is, the shape of a piece of human cortex turns out to be mostly not what people picture.

Why these groups and not others. The table names eleven classes and does not say which of them are neurons. The grouping used here is the one that reproduces the paper's own figures from the table exactly, to the individual cell: 16,087 neurons, 32,315 glia, and 10,531 of the neurons spiny. The class that decides it is the C shaped cell. Counted as a neuron the totals match nothing. Counted with the glia, all three of the paper's numbers come out right. So the paper groups them with glia, and so does this page.

Oligodendrocytes alone outnumber every neuron in the block. They make myelin, and they are densest where the axons leave the cortex, which is why filtering to them in the volume above draws the white matter for you.

One of each

Eleven classes, eleven cells. Each was picked by measuring eleven candidates of its own class and taking the one closest to the middle, so none of them is an outlier. They are drawn at the same height on screen whatever their real size, because a 44 micrometre oligodendrocyte beside a 509 micrometre pyramidal cell would otherwise be a single pixel.

How deep is deep

Depth in cortex is measured from the pial surface, and in this volume that is harder than it sounds. The block is a tilted slab and the cortical surface curves inside it, so depth cannot be read off any axis. Subtracting the X coordinate spreads layer 1 across 2.7 millimetres. Projecting onto the best straight line through the layers still spreads it across 2.6.

Depth here is instead the true distance from each cell to the outward surface of the release's own layer 1 region, a patch of about a third of a square millimetre. The check that this is right is that nobody fitted it: measured that way, each layer's lower edge lands on the next layer's upper edge to within a few tens of micrometres, all the way down.

The one band not to trust is layer 1. Testing all 49,379 labels against all seven of the release's layer meshes, every label's best match is its own mesh, so the column is sound. But layer 1 is the weakest: 78 per cent of the cells carrying it fall inside the layer 1 mesh, against 85 to 99 per cent for the others, and 7 per cent fall inside no layer mesh at all, out in the corners of the block. The label is doing some duty as a catch-all. Its band is left off the ladders on this page rather than drawn as though it were as good as the rest.

The hundred and four

Almost every cell in the volume was reconstructed by machine and never checked. A hundred and four were checked by hand: every branch followed, every wrong merge cut, every missing piece added back. Those are the ones below, with the numbers the release itself publishes for them.

Each cell is drawn at the same height on screen whatever its real size, so that a small one is still worth looking at. The true extent is printed beside it. The mesh is level of detail 3 of the release's multiresolution mesh, roughly 22 thousand faces where the full one is 2.7 million; the levels simplify the surface rather than prune the arbor, so every branch is still there and the skin is smoother than the microscope saw it.

The other end of the scale

Everything above is one cubic millimetre. The companion page steps back about eight orders of magnitude: the white matter of the whole brain, 87 named cable bundles averaged over 1,065 Human Connectome Project subjects, inside that project's own group-average cortex, with a signal whose travel time along each tract is computed from that tract's real length.

A third page steps back along the other axis entirely: the brain from microseconds to decades, twenty things a brain does placed by how big they are and how long they take, with a real human action potential, a real alpha rhythm and real myelin measured across sixty years of adult life.

And a fourth: the map of the world on the back of your head, where visual space is laid out on one real person's cortex. And a fifth, from a whole brain to a single synapse, five orders of magnitude in one move. Click the brain and it tells you where that piece of tissue is looking.

The plumbing

Turn the blood vessels on in the volume at the top. The paper puts about 230 millimetres of vessel in this one block: a quarter of a metre of pipe in a grain of sand, threaded between the cells so that nothing is ever far from a supply. This is 1,432 segmented vessel objects, and they needed no alignment at all, because the vasculature and the cell bodies were measured in the same voxel frame.

It is also the physical reason functional MRI exists. That method does not watch neurons, it watches blood arriving where it is needed, several seconds late. This is the network it is watching.

What is in the release and not on this page

Still a great deal. The 1.4 petabytes of electron microscopy itself, browsable in Neuroglancer. All 149,871,669 automatically detected synapses, of which 111,272,315 were called excitatory and 38,599,354 inhibitory. Every myelin sheath. The full c3 segmentation, which is millions of fragments rather than a hundred cells.

Everything on this page was pulled from that release with the two scripts in scripts/, which say exactly where each number came from. Nothing was smoothed, sampled or invented, and where the data did not support a claim the claim was dropped rather than approximated.