Edit | Violet
A cubic millimeter of brain tissue may not sound like much. But with 57,000 cells, 230 millimeters of blood vessels, and 150 million synapses for a total of 1,400 terabytes of data, researchers at Harvard University and Google have just accomplished a huge achievement.
Together, the researchers have created the largest synaptic resolution, 3D reconstructed segment of the human brain to date, showing in vivid detail every cell in the human temporal cortex, about half a grain of rice, and its network of neural connections.
It's the latest in a nearly 10-year collaboration between Harvard University and Google scientists, who combined Lichtman's electron microscopy imaging with AI algorithms to color-code and reconstruct the mammalian brain's extremely complex circuitry.
今天，这一壮举以《A petavoxel fragment of human cerebral cortex reconstructed at nanoscale resolution》为题，发表在《Science》杂志上。
Link to paper: https://www.science.org/doi/10.1126/science.adk4858
The ultimate goal of this collaboration is to create a high-resolution map of the neural circuitry of the entire mouse brain, which would require about 1,000 times the amount of data they had just generated from a 1 cubic millimeter fragment of the human cortex.
"The word fragmentation is ironic," said Jeff Lichtman, professor of molecular and cellular biology in the Department of Molecular and Cellular Biology at Harvard Brain Science Center and the new dean of the Academy. "For most people, 1 terabyte is huge, but for humans, it's only a tiny fraction of the human brain."
The latest map contains never-before-seen details of the brain's structure, including a rare but powerful set of axons connected by up to 50 synapses.
The team also noticed strange phenomena in the tissue, such as a small number of axons forming a wide range of spirals. Since their samples were taken from people with epilepsy, they were not sure if the unusual structure was pathological or just rare.
Brain "fragments"
Brain fragments were taken from a 45-year-old woman who was undergoing epilepsy surgery at the time. It comes from the cerebral cortex, which is responsible for learning, problem-solving, and processing sensory signals. Samples are immersed in preservatives and stained with heavy metals to make the cells easier to see. Lichtman and his colleagues then cut the sample into about 5,000 slices, each only 34 nanometers thick, that could be imaged using an electron microscope.
The team of Google neuroscientist Viren Jain then built an AI model capable of stitching together microscope images to reconstruct the entire sample in 3D. "I remember this moment, going into the map, looking at a single synapse in this lady's brain, and then shrinking down to millions of other pixels."
Illustration: A single neuron (white) with 5,600 axons connected to it (blue). The synapses that make these connections are shown in green. (Source: Google Research & Harvard Lichtman Lab)
In examining the model in detail, the researchers found unconventional neurons, some of which had up to 50 connections with each other. "In general, you'll find a maximum of a few connections between two neurons," Jain says. Elsewhere, the model shows that the tendrils of neurons form knots around themselves. "No one has ever seen anything like this before," Jain added.
The team also found pairs of neurons that are almost perfect mirror images of each other. "We found that the two groups would send dendrites in two different directions, sometimes with some kind of mirror symmetry." Jain En said. It's not clear what role these traits play in the brain.
Electron microscopy imaging combined with AI algorithms
Image data was segmented using a multi-resolution flood-filling network (FFN) to generate basic fragments, and then clustered using FFN re-segmentation to generate more complete reconstructed cells and processes.
To identify synaptic sites, the researchers trained a classifier based on the U-Net architecture to label three categories: background, presynapse, and postsynaptic. A two-class ResNet-50 classifier was trained to classify each recognized synapse as excitatory or inhibitory based on its electron microscopic appearance, postsynaptic structure type, and presynaptic neuron type (if known).
Illustration: Segmentation, division correction, and merge error correction by neuronal subregion classification and synaptic prediction. (Source: Paper)
Petabillion-level dataset, online open source
Lichtman's field is "connectomics," which is similar to genomics in that it aims to create a comprehensive catalog of brain structures, down to individual cells and circuits. These complete maps will point the way to new insights into brain function and disease, which scientists still know very little about.
Google's state-of-the-art AI algorithms can reconstruct and map brain tissue in three dimensions. The team has also developed a set of publicly available tools that researchers can use to inspect and annotate connection groups.
Researchers used high-throughput serial sectioning electron microscopy to image samples from the human temporal cortex, generating a quadrillion-scale dataset and analyzing it using new tools and computationally intensive methods. Thousands of neurons, more than 100 million synaptic connections, and all the other tissue elements that make up the human brain, including glial cells, the vascular system, and myelin, were reconstructed.
Because the dataset is large and not fully reviewed, researchers share all data in an online resource (https://h01-release.storage.googleapis.com/landing.html) and provide tools for analysis and proofing.
Figure: Shared H01 dataset. (Source: Paper)
"Given the huge amount of money invested in the project, it was important to present the results in a way that everyone could benefit from," Jain said.
A proofreader is needed
The map is so large that much of it hasn't been manually checked, and it can still contain errors made in the process of stitching so many images together. "Hundreds of units have been proofread, but that's obviously only a few percent of 50,000 of them," Jain says. He asked people to help proofread the parts of the map that they were interested in. He said the team plans to make similar maps of brain samples from others, but is unlikely to map the entire brain in the coming decades.
"This paper is truly a masterpiece of the Human Cortex Dataset," said Hongkui Zeng, director of the Allen Institute for Brain Science in Seattle. She added that the vast amount of data that is freely available will "allow the community to gain a deeper understanding of the microcircuits of the human cortex".
A deeper understanding of how the cerebral cortex works can provide clues on how to treat certain psychiatric and neurodegenerative diseases. "This diagram provides unprecedented detail that can reveal new rules of neural connectivity and help decipher the inner workings of the human brain." said Yongsoo Kim, a neuroscientist at Penn State.
Next, the team will study the formation of the hippocampus in mice, which is important for neuroscience because it plays an important role in memory and neurological disorders.
Dataset open source address: https://h01-release.storage.googleapis.com/landing.html
References: https://www.nature.com/articles/d41586-024-01387-9
https://medicalxpress.com/news/2024-05-human-brain.html