Pixelating the brain

Pixelated apple over abstract background of binary code.

MS/MRI research group applies machine learning techniques to neuroimaging data.

“Think of an apple. An apple is round, sometimes it’s red, sometimes it has a stem,” says Marco Law. “We know one when we see one. But what if there’s more to an apple than we’re aware of? What if there are things about an apple that we haven’t noticed, or that are invisible to the eye?”

For Law, a graduate student on a biomedical engineering track, and supervisor Dr. Roger Tam, it’s what we can’t see that has captured their attention. 

“When it comes to images of the brain, a computer can ‘learn’ to interpret details that can be hard to detect,” explains Law. “It can be difficult to quantify things like lesion patterns on appearance alone. But if we can use machine learning to ascribe value to biomarkers in diseases like multiple sclerosis (MS), we believe we can identify changes in the brain via magnetic resonance imaging (MRI) scans before they become symptomatic.”

Machine learning is an area of study within the broader computer sciences field that uses statistical methods to “teach” computer systems outside of conventional programming methods. Deep learning is a subfield of machine learning, and attempts to mimic the human brain in its ability to recognize patterns and attribute meaning to the data that it interprets.

Dr. Tam has been working on the background technology for deep-learning related to MS and neuromyelitis optica (NMO) diagnosis for the past seven years, and sees machine learning as the future of prognostic testing.

“So far, our program has been able to distinguish MS and NMO with more than 80 per cent accuracy, entirely based on deep-learned lesion and tissue integrity features,” says Dr. Tam. “This is one of the most potentially clinically impactful projects we’ve worked on to date.”

To get to this point, Law has been inputting brain images from a closed clinical trial into the system to “train” the computer in what to look for. MS and NMO symptoms are highly variable, but Dr. Tam and his team are uniquely positioned for this work, with a repository of images gleaned from their work as a core analysis centre for clinical trials from around the world.

Working with clinical imaging data, Law has been inputting scans of brains in various stages of disease progression. As more and more data are uploaded to the computer, the program becomes better able to find a correlation between the image itself and learned markers of degeneration.

“The faster we can identify scans in which the disease appears to be progressing, the sooner we can intervene and either begin or change treatment for real patients,” says Dr. Tam.

The team is still working out the details, but they are optimistic.

“We’re hoping to begin inputting larger data sets,” says Law. “Right now we’re limited to the scans we have from clinical trials, which tend to only reflect a certain segment of the population – those who fit standard clinical trial criteria. To ensure the programming is robust, we need to ensure we’re capturing people with more progressive forms of disease, and people from a more diverse age range and socioeconomic spectrum.”

Through the MS & NMO Research Program at the Djavad Mowafaghian Centre for Brain Health, they are hoping to increase their data set via a pan-Canadian study that aims to recruit more than 1500 patients and controls (300 locally), which will enable them to establish better benchmarks, especially for the more progressive, harder-to-treat form of MS. 

“This is a huge field, and there’s a lot of excitement building around the possibilities for machine learning in healthcare,” says Law. “It’s my hope that we’ll be able to use technology like this to identify changes in a person’s disease course more quickly and adapt individual care programs to ensure better outcomes for people living with MS and NMO.”