Multiple sclerosis is an autoimmune inflammatory disorder that affects the brain and the spinal cord, causing motor, sensory and cognitive deficits. The diagnosis of multiple sclerosis is largely based on the visual detection of lesions on brain scans, which can be challenging and cumbersome for radiologists and neuroscientists. New developments in artificial intelligence could allow part of this process to be done automatically, reducing potential mistakes and providing healthcare professionals in Canadian hospitals with valuable insights on the severity and evolution of the disease.

Multiple sclerosis (MS) is a condition where the immune system « attacks » the myelin sheath that covers and protect nerve fibers, creating lesions in the brain and spinal cord that can cause mobility problems, pains, fatigue and more.  In 2020, the Multiple Sclerosis Society of Canada reported that almost 12 Canadians are diagnosed each day. While there is currently no cure for MS, being able to measure the presence and evolution of such lesions allows clinicians to know if a patient will improve or worsen and help them decide which treatment will provide patients the best quality of life possible. However, diagnosing multiple sclerosis on patients’ scans takes a lot of time and is susceptible to human error and variability when done manually. The goal of my research project at Polytechnique Montreal is to develop AI algorithms to automatically analyze MS lesions, by training them on brain scans from multiple Canadian hospitals. The hope is to help people suffering from MS receive optimal treatment while allowing radiologists do more with less.

MS is characterised by the presence of nerve lesions visible on magnetic resonance imaging (MRI) scans, which capture images of the brain or spinal cord in 3 dimensions. A careful examination of these scans is required since lesions in different parts of the brain and spinal cord may have different consequences. For example, lesions in the spine may lead to balance issues or paralysis, while lesions in specific regions of the brain may lead to memory loss and trouble concentrating. Recently, a team of researchers at Polytechnique Montreal found that the size and locations of lesions were associated with the severity and evolution of the patient’s physical disability.

Quantifying the size and location of MS lesions requires “segmenting” individual lesions on each brain slice generated by an MRI scan, which can go into the hundreds for a single patient. This means having an expert manually draw over each possible lesion on each slice, which can take hours for one patient. Because of this heavy workload, this method is not currently part of the clinical routine.

AI can help automate this process by training an algorithm on scans that have already been segmented by a radiologist, so it can learn to replicate the expert’s decisions on other similar scans. When well trained, such an automated algorithm could help detect lesions that a human eye could miss. Quantifying the size of the lesions is also challenging for a human: it is relatively “easy” to count them but much harder to tell how much they have grown since the last scan. Something an AI can do in a split second.

Algorithms need a lot of MRI scan examples with manual segmentation of MS lesions from radiologists. This kind of data is sparse due to privacy concerns and the high costs of making and segmenting those scans. Algorithms are thus often trained on small sets of brain images. This is a problem since MRI scans can look drastically different from one another due to the machine manufacturer or acquisition parameters. The difference in these scans mainly lies in the contrast or colors that are used to see specific brain structures, such as putting emphasis on white matter, grey matter or blood vessels. Current algorithms are mostly trained with scans using the same type of contrasts, and thus fail to detect lesions on images recorded with other settings. Unlike humans, algorithms are bad at learning a concept and generalizing it to new situations. For example, if an algorithm has been trained to detect bears and has only been given visual examples of black bears, it might not be able to detect a polar bear when it sees one. The same applies for lesions in scans coming from different MRI settings.

Our current research at Montreal Polytechnique tries to overcome this challenge. My goal is to make a more general algorithm that could be used in numerous hospitals regardless of the machine’s manufacturer or acquisition parameters. Building such a general algorithm is still an unresolved problem, but we believe the task could be accomplished by training it with a larger amount of diverse data coming from different hospitals. This will be made possible with the new CODA platform, a Canadian initiative allowing us to securely use data from various hospitals to train our algorithm. CODA’s training mechanism is based on a technique called federated learning, where we first train custom algorithms for each hospital separately, which are then merged into one more general algorithm. This prevents sharing confidential data between hospitals, giving researchers safe access to broadly different MRI images, allowing the algorithm to be exposed to “all species of bears” so it will not fail on a different fur color.

The CODA platform is a promising initiative for developing more efficient algorithms to help diagnose multiple sclerosis. These improvements in the state-of-the-art automatic detection and quantification of lesions could one day be generalizable to other pathologies through the private and safe use of clinical data, potentially providing vital help to healthcare professionals and their patients.