I had the chance to interview Benjamin De Leener about a paper on the Spinal Cord Toolbox. SpinalCordBox is an open-source software that allows the automatic segmentation of the spinal cord from MRI images. In particular, we talked about the first version that dates from 2012.

 

• Can you tell me about your background? 

 

I’m Belgian, I did my bachelor’s degree at the Université Libre de Bruxelles, there it’s a common core. I oriented myself towards computer science and biomedical in my master’s degree. Then I started a double degree in Biomedical Engineering at Polytechnique where I did a research master. I was in Samuel Kadoury’s laboratory where I developed tools for spinal cord segmentation. At the end of my master’s degree, I went to Julien Cohen-Adad’s laboratory for a PhD to develop more diverse tools for spinal cord imaging. Then at the end of my PhD I went to McGill University for a post-doctorate with Julien Doyon to do a little more application development, but for patients who are affected by Parkinson’s disease. Then afterwards I was hired by Polytechnique as a professor in the department of Computer Engineering and Software Engineering. 

 

• You are part of a laboratory that is in a hospital, is it a will or was it because of an opportunity? 

 

It’s always an opportunity, usually we have relatively few choices, but this was a great opportunity. To tell the truth, I had a choice: to put him either at Polytechnique at the TransMedTech institute or directly at the Sainte Justine UHC. This second option is still a great added value for a laboratory because when you do medical imaging, the ideal is to have access to the machines fairly easily. It’s interesting, because you can see how the images are acquired, meet patients and not just work behind your computer. You can understand why you’re doing this job. The other advantage of being in a hospital is that you have the application side. You meet the doctors, the clinicians, they’re right on the other side of the door, it’s very easy to meet them. So it’s easy to see how the two fit together between the clinic and the research. That’s also an added value because it’s also a question that we ask ourselves, is what we do for two years really used by doctors. So it allows us to take stock a little more easily. 

 

• Does this allow you to work in pairs? 

 

Yes, that’s it. There are a lot of doctors who are hired at Sainte Justine UHC to do research. So they have their own research program but often it’s very clinical research. They are going to use imaging tools, the tools they have access to at the hospital to answer more clinical questions. Whereas we are trying to develop new methods, new tools that will be useful to them, but sometimes there is not that transmission between methodological development and clinical use. 

 

•  I think it’s related to your background, but why are you interested in the spinal cord? What makes it different from other parts of the body like the brain for example? 

 

Yes, so did I really choose it at some point, that’s always the question. The idea of looking at the spinal cord is that it’s often said that it’s just a tube that connects the brain with the outer limbs, in reality it’s a little bit more than that. It’s true that it’s just a tube, which makes some things easier, but there are more and more things that we’re realizing are taking place in the spinal cord. For example many diseases that affect the brain, which are neurodegenerative diseases like multiple sclerosis. That’s the idea of the spinal cord: it’s part of the central nervous system so it’s affected by the disease in much the same way as the brain, and it’s a simpler structure to analyze. So the biomarkers that you’re going to have for diagnosis, the progression of the disease, will be potentially more robust than the ones you can extract from the brain. It also has disadvantages, which is why we’re looking at it a little later than the brain, because the spinal cord is very deep in the body, protected by the spine. So it’s only recently that we’ve had high-performance acquisition tools for analysis. Then there is still no one who has done anything on the spinal cord, so that leaves a field of opportunities that is quite important.

 

• So it is all the advances in MRI image acquisition that have paved the way for the study of the spinal cord.

 

Yes, absolutely, that’s exactly what happened in the brain. 

 

• I guess it’s not the first software for MRI image processing. What are for you the most innovative features? 

 

There were some tools that existed before. Since then, a lot of tools have been developed.  The big difference between these tools is mainly automation.  

The big difficulty is that you can take any MRI image that has the spinal cord in it and make sure that it locates and segments it. That means that you can do what you want in MRI in terms of acquisition: you can have T1, T2, which are going to have tissue separation mechanisms that are different from a physical point of view. You can also have the edges that are offset: Between a T1 and T2 image the marrow is not the same size, yet it is the same person and there is only one minute between acquisitions. In addition to the images you can take the spinal cord from the neck, or head and neck or only the body. There is so much variability in the acquisition procedures that it is difficult to develop a tool that is 100% automatic. 

There is the second part:whether or not it’s a good segmentation. Because there are many tools that will make segmentations that are relatively interesting but that we will have to correct. Depending on how you correct it, it can be disturbing to have something semi automatic. It’s a tool that can adapt to the field of view. That’s where I developed a deformable model propagation tool so that no matter the size of the image we can have a method that works. 

The image above illustrates the segmentation of the grey matter of the spinal cord. This figure illustrates what Benjamin said: if we compare the automatic and manual segmentation the results are very good.

• SCT is compatible with mpIRMs. What are their particularities compared to “classical” MRIs?

 

It’s a question of name, it’s a general name for MRI. The idea is that MRI has the ability to provide very different images by manipulating the parameters whereas structural MRI is called T1 and T2 MRI. 

 

• To perform the segmentation of the spinal cord, the software uses a PropSeg algorithm that you developed. It consists of two steps: the detection of the center of the spinal cord and then the propagation of a tubular surface along the spinal cord. Can you explain the second step in a little more detail? 

 

In the first version of the software, the center of the spinal cord was detected in the middle of the image. We only had one point which was our starting point. Around that we built a tubular surface that we will put in the right direction. From that moment on, we will try to deform the cylinder so that it adapts to the spinal cord.

We have a surface that is going to be deformed based on certain parameters: external and internal energy. The external energy tries to get closer to the edges of the spinal cord. We add internal constraints to this because without if the image has certain defects, with the gradient, the surface will be attracted by the edges. This internal constraint will try to confine in something that is still circular. We have an energy equation that will try to balance these two parameters. 

What we’re going to do next is we’re going to take the top part and we’re going to copy and paste the surface in the direction we’re going. We move it by a certain amount, we’re going to detect the rotation on the new surface, and then we’re going to deform the surface again. We do this upwards and then downwards. And if all goes well, we have our model that works well. And since we have discontinuities in our model every time we copy and paste, we will make a global deformation of the whole surface afterwards.

• In the era of Deep Learning why not use this method? 

 

In 2012 it was not fashionable. It was starting. There weren’t as many tools. 

I’m still happy with the choice I made because it’s an algorithm that is totally deterministic. It means that you know what characteristics you’re looking for in the image, you have control over that. The whole method you can parameterize it. They’re not necessarily features that are easy to adjust. But it still adapts well to the problem which is the segmentation of the adult spinal cord because it’s always about the same size. Whereas with DeepLearning you can’t do anything. You can never have a method that won’t have an image where it doesn’t work. With a completely deterministic method you can adapt it to make it work. A mix between the two can be a good approach. 

 

• Why did you choose to make open-source software? 

 

Why not ?

 

• I see the advantage of the open source, the reproducibility, but otherwise there can be financial advantages. 

 

It depends on what you’re going to do with it. When you develop software you have a lot of possibilities. It depends on your interests and the potential of the software. Let’s say you develop software that can be directly applied to an industry problem that you solve. Then you’re not necessarily going to put it open source. 

You can protect your software, patent it, license it, and try to market it to the industry so that they can implement it. The question is: how do you promote it? What we often see is software developed by researchers, who try to encapsulate it, protect it, and then in the end they don’t do anything because they can’t find anybody interested. 

On the other hand, there is another dimension, which is that if you develop software at the research level, your software is not perfect. Because you’re always looking for ways to do it. By making the tool open source and providing good documentation you have a tool for researchers, which makes some things easier. Your tool is not perfect: you don’t sell it so it’s normal. If it crashes you can say “calm down, we are all researchers”. Plus people use it for research, they do something with it and that’s also rewarding. As we use it, as we improve it, as we add components around it, it’s going to be more and more useful from a clinical research point of view. So sometimes it’s the research community that’s going to help you move in the right direction to achieve the clinical goal. 

Afterwards, you should not fall into the trap of only doing open source with the “guys I can’t do any more, you can’t do any more, you’re on your own” mentality. From the moment you say that, you take away potential value from your software because you don’t accept to take it to the next level where it will be useful.  

 

• The software seems to give very good results, do you plan a new version of the software in the future? 

 

There are a lot of different versions that have been made. The model I explained to you earlier with the point has a very big flaw is that if at some point you get it wrong, the spread goes outside the spinal cord and continues. To this we have added tools that allow us to detect the central line of the spinal cord before segmenting it. We then follow a thread rather than calculating the progressive orientation. 

And there are probably more to come. Because the tool was developed for adults and that’s what you see in a lot of medical imaging software. It’s developed for people between 20 and 25 years old because when we do research we use volunteers to be able to take images and the volunteers are often students. Here, it will be the adaptation of the tool for pediatrics.