One of the most popular techniques to procure synapse-resolution datasets relies on combining serial-section electron microscopy (EM) imaging technology with cutting edge artificial intelligence (AI). Here’s how the process works.

We start with a piece of brain that’s been removed from an animal. No microscope exists that can image a live brain at anywhere close to the resolution of EM, so if we want to see and chart the sizes of every synapse, EM is necessary. Most modern EM research projects pair activity with connectivity measured from the cells that are later imaged. This yields rich functional maps of neurons and their synapses, allowing teams to decipher computational circuits. This field is called Connectomics.

Since neurons are naturally clear, we first stain them with heavy metals that attached to cell membranes. This makes the boundaries between cells visible. Stained neural tissue is then embedded in a hard block of resin so it can be sections into tiny slices just 1/500th the width of an average hair thin. Each slices is imaged at nanoscale resolution using an electron microscope and aligned into a 3D volume, allowing the block of brain to be digitally stored and processed.

We can now use AI to reconstruct individual 3D neurons that have grown within this volume.

Even with the best software, it’s incredibly laborious to reconstruct neurons. It can take years to do even a few hundred neurons. The neural circuits that scientists want to investigate are composed of hundreds of thousands of neurons, so we need to find a faster way.

We use a convolutional neural network (CNN) to mark whether each pixel in each image is the boundary of a neuron or not.

A CNN is a state-of-the-art artificial intelligence method that’s based on how simple models of neurons learn to recognize objects. That’s right -- we’re using simple models of the brain to help map the brain. Hopefully we’ll use those maps to make more accurate models of the brain to complete the virtuous cycle.

Automated neuron reconstruction using AI is good but far from perfect. That's where humans come in. Working alongside the partial reconstructions, humans are able to identify and solve tricky cases that stump even the best AI.

//While we reconstruct the neurons, we also need to identify their connections where two cells form a synapse. If the tissue is stained and imaged properly, synapses can be visible as a really dark smudge between two neurons neighbored by a set of cellular vesicles transporting neurotransmitter.

We use another convolutional neural network to look through all of the images and label those dark smudges as synapses. Then we combine this map with our reconstructed neurons to say that one neuron sends a signal to another neuron at a particular location. We can even estimate the size of the synapse, which helps us infer its strength.

And there you have it! With the reconstructed neurons and the synapses, we have a connectome. Well, at least a part of one. Connectomics researchers have started small and are working their way up to larger datasets. There have already been exciting discoveries revealing uncharted circuits and inviting many new questions.

Why the Nanoscale?

We need the nanoscale to track every branch of a neuron as it wires its way through the brain. Unfortunately, trying to track something at the nanoscale for centimeters to meters is a Herculean task because of the vast difference in scale. Fortunately, many fields of neuroscience are working together to tackle this challenge and there’s a lot of additional information we can see at the nanoscale. There are mitochondria, nuclei, microtubules & microfilaments, endoplasmic reticulum, golgi bodies, vesicles, postsynaptic densities. Also fortunately this task isn’t left to labs alone thanks to the participation of citizen scientists.