How the Human Brain Project Maps the Brain Faster with Deep Learning

by Jens Neuschäfer

The Human Brain Project has ambitions to advance brain research, cognitive neuroscience and other brain-inspired sciences like few other projects before it. Created in 2013 by the European Commission, the project’s aims include gathering, organizing and disseminating data describing the brain and its diseases, and simulating the brain itself.¹

To do so, they’re going to need a good map. That’s because the human brain, with about 100 billion neurons and 100 trillion connections, is one of the most complex systems known to man.

To map something so complicated, scientists at the Jülich Research Center (Forschungszentrum Jülich), in Germany, are developing a 3D multi-modal model of the human brain. They do this by analyzing thousands of ultrathin histological brain slices using microscopes and advanced image analysis methods — and then reconstructing these slices into a 3D computer model.²

Analyzing and registering high-resolution 2D image data into a 3D reconstruction is very data- and compute-intensive. This makes it an ideal match for Jülich’s JURON supercomputer, which is stocked with NVIDIA Tesla P100 GPU accelerators that offer the power to quickly solve simultaneous registrations.

Julich lateral dorsal brain images
Images courtesy of Forschungszentrum Juelich.

Deep Learning Speeds 3D Map-Making

To build a digital 3D atlas of the brain, scientists traditionally have analyzed the patterns of cell distributions in histological slices. By staining the cell bodies with dye, they can identify borders between different areas of brain. Doing this across many different brains yields a common 3D reference model.³ It also yields maps that indicate the probability that a given area of the brain is found in a specific location in the brain. Two hundred such areas have now been documented, corresponding to roughly 70 percent of the human brain.

But that’s a time-consuming process and doesn’t scale with high throughput imaging, so a science team at Jülich under the lead of Professor Katrin Amunts and Dr. Timo Dickscheid have now applied deep learning to speed things along.4 They trained a convolutional neural network that combined topological information from probabilistic atlases with the texture features learned from high-resolution cell-body stained images.

The team applied the model to visual areas and trained it on a sparse set of partial annotations. The model worked, making predictions that are spatially consistent and reproducible on sections of previously unseen brains.

Such 3D brain models mark an important advancement in science’s ability to better understand the structure and function of healthy brains. Using it, physicians ought to be able to compare the information from the brain map with patient data gained from advanced imaging techniques. The differences between the two will provide helpful information for clinical practice about a whole range of neurological or neurodegenerative diseases, such as Parkinson’s, Alzheimer’s and strokes.

The JURON cluster is one of two pilot systems delivered by IBM and NVIDIA to the Jülich Research Center. It is composed of 18 IBM Minsky servers, each with four Tesla P100 GPU accelerators with NVIDIA NVLink interconnect technology. Learn more about the JURON system.

  1. Katrin Amunts, Christoph Ebell, Jeff Muller, Martin Telefont, Alois Knoll, and Thomas Lippert (2016): The Human Brain Project: Creating a European Research Infrastructure to Decode the Human Brain. Neuron 92, November 2, 2016
  2. Amunts, K., Lepage, C., Borgeat, L., Mohlberg, H., Dickscheid, T., Rousseau, M., Bludau, S., Bazin, P., Lewis, L., Oros-Peusquens, A., Shah, N., Lippert, T., Zilles, K., Evans, A., 2013. BigBrain – an ultra-high resolution 3D human brain model. Science Vol. 340 no. 6139 pp. 1472-1475
  3. Amunts K, Zilles K. (2015) Architectonic Mapping of the Human Brain beyond Brodmann. Neuron. 2015 Dec 16;88(6):1086-107. doi: 10.1016/j.neuron.2015.12.001. Review
  4. H. Spitzer, K. Amunts, S. Harmeling, T. Dickscheid (2017): Parcellation of Visual Cortex on high-resolution histological Brain Sections using Convolutional Neural Networks. IEEE International Symposium on Biomedical Imaging (ISBI)