By: Krista Davidson
2 Feb, 2022
The brain’s complexity is as deep and enigmatic as the cosmos, which has led a team of scientists at CIFAR to apply artificial intelligence (AI) techniques typically used to detect stars to identify neurodegeneration in the brain.
The idea for the project grew from a CIFAR meeting that brought together interdisciplinary researchers to examine the use of powerful algorithms in space and health applications. The approaches astronomers take to map stars in the sky could help neuroscientists pinpoint the macroscopic manifestations of neurodegenerative diseases such as Parkinson’s and ALS well before patients experience symptoms. Even more, it holds great promise for unlocking the mysteries surrounding both human and machine intelligence.
“In preparation for a meeting at CIFAR, I found myself studying images of the Universe and I realized they looked very similar to my microscopy images of the brain. The challenges were very similar for both, in terms of trying to identify objects that are difficult to find,” says Flavie Lavoie-Cardinal, an assistant professor at Université Laval and Canada Research Chair in Intelligent Nanoscopy of Cellular Plasticity.
“I thought perhaps the key to understanding one lies in understanding the other — the infinitely small seems to fit neatly with the infinitely large,” she says.
It was a moment of illumination that led to the formation of a team including Lavoie-Cardinal, CIFAR Azrieli Global Scholar Renée Hložek and two Canada CIFAR AI Chairs from Mila, Audrey Durand and Christian Gagné. The team successfully applied for a CIFAR Catalyst Fund, which provides funding for high-risk, high-reward projects at the earliest stages of research.
“The infinitely small fits neatly together with the infinitely large.”
The goal of the project is to detect neurodegeneration so that health-care practitioners can apply preventative medicine versus curative medicine, but Lavoie-Cardinal, Hložek, Gagné and Durand all agree that the research has the potential to advance their fields of research.
“The Catalyst Funds from CIFAR were perfect for this type of collaboration because it gave us the rapid start required to prove to ourselves that we could do this,” says Hložek, an assistant professor in Astrophysics at the Dunlap Institute in the department of Astronomy and Astrophysics at the University of Toronto. Hložek is also a CIFAR Azrieli Global Scholar.
Lavoie-Cardinal’s team is collecting image samples of primary neuronal cultures from the hippocampus and the cerebral cortex of newborn mice. The team then prepares a transfection on the sample, a process of artificially introducing a DNA protein into the neurons from models of Parkinson’s or ALS. This process also allows the team to tag proteins of interest in neurons with molecules that emit light and can be detected using fluorescence microscopy. Super-resolution microscopic images capture synaptic structure with a resolution that is 10 times better than standard optical microscopy, allowing scientists to observe otherwise non-visible patterns and identify potential anomalies.
The images bear many similarities to the type of data found in radio observational imaging. In astronomy, scientists extract signals from large volumes of data. The data is captured from different telescopes, resulting in varying configurations of the sky.
“When you’re looking into a telescope you see only a projection of the brightest bursts of light in the sky. In this way, astronomy is similar to microscopy because you’re presented with a nanoscale of one tiny cell at a time,” says Hložek.
AI systems are highly successful in identifying patterns, but anomalies in the brain follow no discernible pattern, colour or shape. In addition, microscopes introduce noise to the data, which can make it challenging to distinguish a brain anomaly from an instrumental error — what is referred to as “noisy” data. Manual detection is impossible due to the number of images of the brain that one must capture to form the larger picture of the brain’s synaptic structure and function.
The challenges are similar in astronomy in that telescopes also introduce noisy data in their images, but astronomers have been able to use machine learning to train systems on the parameters needed to weed out irrelevant data found in the sky.
“Machine learning is useful in this project because you can automate processes to find variations in the data. It can identify patterns and new structures for knowledge discovery,” says Christian Gagné. Gagné is a full professor in the department of electrical and computer engineering at Université Laval and director of Institute Intelligence and Data.
The machine learning approaches that astronomers use to detect transient images in the Universe could be the breakthrough needed for neuroscience, but it can also advance our knowledge of machine intelligence. While machine learning has mastered tasks related to identifying patterns, it struggles, much like humans, with the unknown.
“This project could transform the way machines learn.”
Durand and Gagné are employing both traditional and nontraditional approaches, including convolutional neural networks, a deep learning network that is useful for identifying patterns and eliminating irrelevant data; and generative adversarial networks (GANs), a framework in which two algorithms play against each other with the goal of “winning” while gathering important insights about their task, improving their performance without explicit guidance.
“This project could transform the way machines learn,” says Durand, an assistant professor in computer science and software engineering at Université Laval. “The more we look at how objects behave across different images, rather than looking only at one fixed image, provides a lot of insight and information. It’s common in both neuroscience and astrophysics, but less common for machine learning researchers who are working in computer vision, where many tasks rely on a single image.
“If we can get algorithms to extract concepts without much knowledge of their environment, it will bring artificial intelligence much closer to human intelligence,” says Durand. “There is so much we can do just in fundamental neuroscience to understand how the brain works by looking at anomalies. This project has the potential to not only advance our understanding of neurodegenerative disease in general, but to address molecular mechanisms in the brain,” says Lavoie-Cardinal.
For Hložek, it represents an opportunity to perfect the algorithms used for detecting fast bursts in the sky and reduce the false detection rate.
“What we’ve learned so far is that while we’re helping advance biological methods, they’re actually informing our decisions around imaging in astronomy, which I didn’t expect,” says Hložek.
CIFAR is a registered charitable organization supported by the governments of Canada, Alberta and Quebec, as well as foundations, individuals, corporations and Canadian and international partner organizations.