Predicting Multiple Sclerosis Progression, Lesion Activity With Machine and Deep Learning

March 5, 2019

The director of the probabilistic vision group and medical imaging lab at McGill University spoke about machine learning’s potential to help physicians predict MS disease progression, treatment effectiveness, and more.

Tal Arbel, PhD

Across a number of diseases and in health care as a whole, machine learning and deep learning algorithms have begun to make an impact on the process of collecting patient outcome measurements. Now, they have begun to make its way into the field of multiple sclerosis (MS).

In a presentation at the Americas Committee for Treatment and Research in MS (ACTRIMS) Forum in Dallas, Texas, Tal Arbel, PhD, the director of the probabilistic vision group and medical imaging lab at McGill University, helped to show how these learning methods can be utilized for the prediction of future lesion activity and disability progression. Notably, Arbel explained since new T2 and gadolinium-enhancing lesions are indicators of disease activity, predicting their appearance in future images could help predict disease worsening, as well as treatment responders.

To find out more about this method, its potential impact in MS, and what work is being done right now by Arbel and her group, NeurologyLive spoke with the professor in the Department of Electrical and Computer Engineering at ACTRIMS.

NeurologyLive: Could you discuss your presentation and its relevance to the MS field?

Tal Arbel, PhD: I spoke about an area of machine learning that I've been working on in the context of multiple sclerosis. Machine learning is an area within the field of artificial intelligence and I spoke about how methods and machine learning can be used for difficult problems in neurology, specifically, in multiple sclerosis.

As we all know, machine learning, expressly a sub-area called deep learning, has revolutionized a lot of different fields. My area of research is computer vision, which is an area where we develop computer algorithms in order to get computers to understand what they're seeing. In the field of computer vision, there's been a huge sort of revolution in terms of methods that have been successful, and deep learning methods applied to problems in computer vision have led to an explosion of success in terms of identifying objects. Many applications such as Facebook, autonomous vehicles, and all this—all this growth has been generated due to the success of deep learning where it's outperformed other methods by huge margins, leading to huge success in terms of startups and corporate success. Some of the success is, in no small part, due to the fact that in computer vision there's a lot of data. Deep learning needs a lot of data and a lot of annotated data and, of course, advances in hardware.

There's a lot of room for machine learning to have similar success within the context of medicine, from diagnosis to prediction and precision medicine and so on. However, a lot of these techniques are not regularly used in the clinic. There are a couple of reasons for that. The first is that, typically, the scientists that develop these techniques are in computer science or electrical engineering where they don't have access to large annotated datasets and they don’t have access to clinicians and what their needs are, so they develop methods which may not be robust to different sites, and scanners, and stages of disease. They also may be focusing on metrics which may not be important for the specific clinical task that that is of interest.

In my group, I have a very long-standing collaboration with a company that does software analysis for the pharmaceutical industry in order to do analyses for clinical trials for MS. They have given us huge amounts of data from different clinical trials, from different scanners, from different sites, and each of those time points has multimodal MRI and temporal information, but also much more importantly, expertly annotated lesions and other annotations within the images. In addition to that, I'm part of the international progressive MS Alliance group that received funding to federate the first big data set of progressive MS patients with MRI. We're hoping to get up to 40,000 patients over time all their imaging and some clinical information and my group has already received up to let's say 20,000 so far. With all that data, I'm able to do some deep learning and some machine learning. My group, over the last 15 years, has been developing techniques in terms of detecting and segmenting lesions in patient MRI, mostly from large clinical trial datasets. We're also now, with the advent of this IPMSA dataset, looking at predicting future lesion activity in patients, and future disease progression from baseline imaging.

This is really important for a number of reasons. In order to predict, for example, how the patient is doing and then do the appropriate treatment selection; In order to better understand the disease. We're also looking at trying to understand whether we can get biomarkers—imaging biomarkers—that would be predictive of future progression. This is going to be important not only to understand progression a little bit better, but also to understand, for example, in the context of early phase clinical trials, to be able to enrich trials or determine if treatment is working.

What are your current research plans that you’re working on?

The work on predicting progression is very, very new. Actually, it was accepted for publication just a few days ago, and we're the first group to work on that. I'm very excited about that work.

We're also looking at embedding longitudinal information. We're trying to understand if we can use deep learning to develop tools to understand disease trajectories, like the temporal evolution of the disease, to try to see if we can get subtypes. For example, why do certain groups of patients progress in a certain way and others don't? Also, to bring in the treatment effects to try to understand why certain people respond to treatment and others don't, and what are the different courses of the disease to help us with our predictions. That's some current work in my group.

In order for it to be properly integrated into the clinic, we also have techniques to predict uncertainty in our predictions. For example, we would say, “This person will have a lot of lesions in future imaging or will progress in in terms of their disability,” and then it provides a degree of confidence. The algorithm will tell you whether they're confident in that assertion or whether they're not confident. Similarly, we have that for the lesion segmentation that we perform—we also have some sort of assertion of uncertainty, so that radiologist can then review where the system is very confident or not so sure about the lesions that they're looking at. They can go back and review it.

What is the potential impact that machine learning and methods like this could have in MS?

I think that machine learning and deep learning, with the state that were in now, has the opportunity—we have the opportunity—to have a huge revolution in terms of healthcare. There's a lot of promise in ways that we don't even really fully understand right now, so I'm very excited to be in this area. I've been very fortunate to work with clinicians at the Montreal Neurological Institute and I think that there's a lot more opportunity for success.

REFERENCE

Arbel T. Machine learning for MRI-based lesion segmentation and prediction of disease progression in MS. Presented at: ACTRIMS Forum; February 28 to March 2, 2019; Dallas, TX. actrims.confex.com/actrims/2019/meetingapp.cgi/Paper/3233. Accessed March 4, 2019.