Deep learning transforming neuroscience research

Using thousands of images from brain scans such as MRI, computers can learn to detect signs of neurological disease, opening up new possibilities in research and diagnosis.
Using thousands of images from brain scans such as MRI, computers can learn to detect signs of neurological disease, opening up new possibilities in research and diagnosis.

By Shawn Hayward, The Neuro

In an article published in Nature on Feb. 15, 2017, researchers, including principal investigators from the Montreal Neurological Institute’s McConnell Brain Imaging Centre (BIC), used magnetic resonance imaging (MRI) to predict the development of autism in babies.

It was not a neurologist or medical doctor doing the predicting, however, but a computer trained to distinguish the brains of children at risk of autism. This was an application of “deep learning”, a form of artificial intelligence that will increasingly put computers in the driver’s seat of medical diagnosis and neuroscience research.

Developed from the concept of artificial neural networks in the 1960s, in the footsteps of the pioneering work of Donald Hebb, a former McGill psychology professor, deep learning has experienced a kind of renaissance in recent years, thanks to the increasing availability of powerful computational resources and access to vast amounts of digital data.

Deep learning involves training computers to make complex calculations after analyzing enough data to “learn” or detect certain patterns of interest. They do this via relatively simple algorithms that mimic the brain’s basic mechanisms for processing information.

If you are on Facebook, you probably have already experienced AI in action. Facebook can detect where faces are in images and will ask you if you want to tag that person. The program that makes this possible is called DeepFace, a deep learning application Facebook developed by training computers to recognize faces using four million photos manually tagged and uploaded by users.

Deep learning techniques are being used in many aspects of biomedical research. One objective is to develop computer-assisted techniques to improve diagnosis and prevention, by analyzing data of various kinds to see problems before they occur. Deep learning is particularly important to neuroscience, where data types are extremely diverse. Artificial intelligence is a promising tool to help neuroscientists discover new basic principles within the vast amount of data available.

The Nature article is just an example of how deep learning and other AI techniques are rapidly becoming important to medicine and medical research, among other fields affecting our daily lives.

Several labs at the Montreal Neurological Institute are already using deep learning and related AI techniques to conduct research, and the BIC is training the next generation of neuroscientists and brain imagers to use these new methods. In January of 2017, the BIC sponsored two hands-on educational sessions focusing on deep-learning for neuroimaging. The event was attended by 80 of the centre’s students and staff scientists.

“AI techniques are changing the game of how we do science. We want our research staff and trainees to be aware of and well prepared for this revolution,” says Sylvain Baillet, a McGill professor and director of the BIC. “We are fortunate that Montreal is emerging as an international hub for AI research and industry. To remain leaders in our field, we must embrace AI methods like deep learning together with building and using large neurodata repositories, and invest both human and technical resources to exploit the unique features of these powerful tools.”

March 23, 2017