LinkedIn – Connect with Heather.
Computer Vision Insights Newsletter – A biweekly newsletter to help bring the latest machine learning and computer vision research to applications in people and planetary health.
Computer Vision Strategy Session – Not sure how to advance your computer vision project? Get unstuck with a clear set of next steps. Schedule a 1 hour strategy session now to advance your project.
Foundation Model Assessment – Foundation models are popping up everywhere – do you need one for your proprietary image dataset? Get a clear perspective on whether you can benefit from a domain-specific foundation model.

[INTRODUCTION]

[0:00:02] HC: Welcome to Impact AI, brought to you by Pixel Scientia Labs. I’m your host, Heather Couture. On this podcast, I interview innovators and entrepreneurs about building a mission-driven machine learning-powered company. If you like what you hear, please subscribe to my newsletter to be notified about new episodes. Plus, follow the latest research and computer vision for people in planetary health. You can sign up at pixelscientia.com/newsletter.

[INTERVIEW]

[0:00:33] HC: Today, I’m joined by guest, Dirk Smeets, CTO of icometrix, to talk about imaging for neurological conditions. Dirk, welcome to the show.

[0:00:41] DS: Thank you very much, Heather. Please to be here.

[0:00:44] HC: Dirk, could you share a bit about your background and how that led you to Icometrix?

[0:00:48] DS: Yeah, sure. As you mentioned, I’m the Chief Technology Officer at icometrix. Actually, that means that I’m responsible for all our product developments, also including all of our AI, but also our strategic partnerships and trying to make sure that the technology that we develop finds a sustainable fit in the healthcare system, so that is called market access. That is – meaning, it’s being in touch with, for example, insurance companies and so on. I myself, I’m an engineer as a background. I also did a PhD in medical imaging and I joined Icometrix almost from the beginning of icometrix and tried to build out all the science and technology parts of the organization.

[0:01:29] HC: What does icometrix do? Why is this important in treating neurological conditions?

[0:01:34] DS: Well, if you look at it from a helicopter view, we are probably in a decade that neurological conditions will be really at the center, a bit similar as what how cancer has been perceived in the past as today. One in three people will suffer in their life from a neurological condition. The societal burden for neurological conditions is a sum of kidney disease, heart disease and diabetes together. It’s getting a real issue in society. You probably if you look around in your family or in your friend’s environment, there might be a person with a neurological condition, whether it’s a devastating one, like Alzheimer’s disease or a disease that can still be recovered quite well, like for example, a mild stroke.

There are so many neurological conditions that impact our lives today. We, as icometrix, we really try to be helping the way out there as many neurological conditions might be treated in the coming years. There is a disease called multiple sclerosis, where there are already many disease-modifying treatments available. These are medications. Yesterday, there was for the very first time in history, a treatment approved for Alzheimer’s disease. You see that the field of neurological conditions is moving.

There are treatments available, but the downside unfortunately, is that those medications are not working for everyone. It is still a lot of trial-and-error today. What we as Icometrix try to do is to, on the one hand accelerate the trial error process by better profiling patients. On the other hand, to just give the trial-and-error process and just build a digital twin of every patient to really think what is the best treatment for that particular patient and provide that best possible treatment at first. That’s why we believe it’s so important to match also these therapies with technology. AI is a super central there.

[0:03:42] HC: What role does AI and machine learning play? How does it help you in this technology?

[0:03:46] DS: Yeah. It’s playing a very central role, I would say. Not necessarily, because we see it as a purpose or purpose is to help people with neurological conditions, but because of the advancements in computer power, what is possible with machine learning and more particularly with deep learning was just not possible to imagine many years ago. You can also see it in with the recent outome of ChatGTP, for example. These were models that were impossible to build in the past as they need large computation power.

Also, in the field of neurological conditions and specifically in imaging for neurological conditions, we see a little bit of the same. It is quite intensive in terms of memory requirements for computation, but as that is coming more and more available, we can build machine learning models that can do tasks at the level of experts. For example, expert radiologists. That will change a lot the way we do current practice, as that means that the machine learning algorithms are able to provide this expertise to every single hospital; whether you’re in an expert academic medical center or in a rural hospital, the AI will bring that top expertise there. That is, I think, something that will change the way how we see care being done in the coming years. Yeah, that’s how I see the role of machine learning in our technology.

[0:05:19] HC: What kinds of tasks do you use machine learning for specifically with neurological conditions?

[0:05:24] DS: We use machine learning quite a lot. Where I would see its most central role at icometrix is in the analysis of imaging data, more particularly magnetic resonance and computer tomography images, MRI or CT in short. They’re acquired by specialized machines in the hospitals, and they provide actually quite big data sets, typically three-dimensional data sets. You can see it as a breadth that is sliced, and that’s how an MRI image looks like. It can be the MRI image, for example, of a brain where it is actually containing different slices that contain different parts of the brain, and altogether, you have a three-dimensional view of those images.

That is, of course, particularly interesting as those treaty images, they also have technical challenges attached to it. It’s different from, for example, a 2D photography in the sense that it is also bigger in terms of memory requirements. It needs fundamental adaptations of the machine learning technology to also work on those types of images. I think that makes it super challenging for more perspective as a company active in the field of AI.

[0:06:37] HC: Then how do you adapt to the standard approaches in deep learning and computer vision for these three-dimensional images and for the challenges you mentioned?

[0:06:47] DS: Yeah. That’s a good question. Actually, typically, we would start with traditional 2D technology, just treating those 3D images as different – a stack of 2D images. Of course, that doesn’t work optimally, because – you lose information. The second step is what we would call 2.5D. It means that we would take one slice of the three-dimensional image together with some neighboring slices, and we apply the deep learning algorithms on those slices. It’s still a little bit the ID that comes from 2D processing, but it takes already some context of the neighborhood into account. The big last step is, of course, to work in truly 3D models. That is luckily possible now with the big computational power that is available.

Another way that we can also use to make it possible is to divide that large three-dimensional image into blocks of small cubes that we would call patches. We would apply the deep learning algorithms on those patches. These are two ways that we use to also apply techniques, originally coming from 2D images into the world of 3D scans and make sure that it really can-do meaningful things on those images.

[0:08:09] HC: Applying deep learning, what specific tasks are you using it for? Are you trying to predict the presence or absence of a particular condition or is it more fine-grained, like measuring different structures in the brain and quantifying their size?

[0:08:22] DS: Yeah, indeed. There are actually different levels of information that can be extracted from such images. Typically, it starts from just extracting simple measurements that also rate audiologists or neurologists would theoretically be able to do, but then in an automated manner. For example, in people with multiple sclerosis, there are little spots of inflammation. They are called plaques inside the brain of those patients. They appear as white spots on the MRI scan. What we will do is we will try to identify them and measure their volumes. It’s a task that theoretically could also be done by a radiologist that would take a lot of time, but machine learning algorithms would automate that task.

The big advantage of this approach is that this feels very familiar to the radiologist and neurologist and makes it much easier to adopt in their clinical workflow, because it’s something that they might be willing to do anyway, but now it’s done in a more robust, more accurate and faster way than they would ever be able to do so. The second level is that we try to get more information out of it, instead of just measuring it. We can try to classify a lesion, whether it’s a multiple sclerosis lesion or another type of lesion. That intelligence can also be trained for by a machine learning algorithm.

That is something that we can, definitely, also implement with current technology. It’s already one level more difficult to implement in a clinical practice, because then the radiologists or neurologists need to have more trust in the algorithm. Then the last level of information that the ML algorithms can provide are a true diagnosis. The scan can be given and the full diagnosis can be done. That is something that is currently still at the very early stages in development. In the medical field, there is not much of those algorithms really deployed in a clinical practice.

On the one hand, that has to do with the challenges to build such tools in a reliable manner. It’s not easy, every patient looks different on the one hand, but also, because it requires a fundamental trust by neurologists, radiologists, other clinicians in the technology. That is not so easy. It will take time. It will take time also, to build the evidence for those machine learning algorithms to really prove that they can make the right diagnosis based on an image, for example.

[0:10:57] HC: The easiest place to start is, like you said, with the first level set of tasks where you’re making a radiologist more efficient and removing some of the mundane measurements they might otherwise do, but they can understand the outputs, they can verify that they’re correct. These higher-level tasks, as you said, it’s the future, but it’s still in progress. Is that accurate

[0:11:20] DS: Absolutely. If you look also in the field, for example, for radiology, most of the AI applications that are currently being used are definitely in that first category. It’s only in the last years that, for example, in the United States, the FDA is already accepting the first computer aided diagnostic and detection tools, which is like a one level higher. I would say there are about 10% of all the algorithms that have been cleared by the FDA. You see, that second stage is starting to be reality, but the third level is something that is still very, very early.

[0:11:58] HC: In working with applications across these three different levels and in deciding what specific tasks to apply machine learning to, how do you ensure that the decisions you make and the areas that you target are the right ones to ensure that the technology your team develops will fit in with the clinical workflow and provide the right assistance to doctors and patients?

[0:12:19] DS: Yeah. Clinical workflow is key in what you’re asking here. I think it is easy to, relatively easy to develop AI algorithms, take a data set. You probably can even take open-source tools to design an ML algorithm and to develop it, but to make sure that it is seamlessly integrated into the workflow of a clinician, that will make a difference of the tool being used or not. For example, from that perspective, we have chosen to be a background service. We actually didn’t want our users being the neurologist and the radiologist to interact with the software.

It is built that everything is done automatically in the background after the acquisition of an MRI or CT scan and before the radiologist starts the reading. As such, the radiologist has just more information and his availability. He is empowered to make more difficult decisions as he has more knowledge. That is something that we have chosen to do. I think that seems to be the right approach.

[0:13:26] HC: Taking another step back and thinking back towards the beginning of your product development or the start of a new feature that you’re developing, how does your team plan and develop for a new machine learning product or feature?

[0:13:40] DS: Yeah. It’s very difficult to weigh different feature requests. They can come from customers or other stakeholders. They can come from, for example, or medical advisory board, or strategic advisory board, or they can come from our icometrix vision. As a company, we have a certain purpose, a purpose to help people with neurological conditions to find the right treatment and that also translates into features.

All of these features and perspectives of the product, they come together and they will be implemented in a way that we try to prioritize the ones that we feel are most important first. We do that in an agile approach. We try to implement them relatively quickly and test them with our current customers, so that they can give feedback on the feature. If they like it, we keep it in. In such a way, we built the tools to meet our vision, to meet our strategy, but also to meet our customer needs.

This is also something that we struggle with sometimes, as we’re in a very regulated environment or tools that are regulated by the medical device regulations from the FDA or the European medical device regulations. It’s not just every change that we can simply put in the market and ask customers to test. For that, we also try to build mechanisms around it to make still sure that customers have early access to new features and can evaluate it and really say whether that fits their needs, because that is ultimately what counts, of course.

[0:15:18] HC: You mentioned the regulatory process. How does this affect the way you develop machine learning models? Are you thinking about it from the very beginning as you’re deciding whether to develop a new feature or does the planning come in a bit later in the process?

[0:15:31] DS: Yeah. In our business, this is really an impactful aspect. Regulations, if you speak about regulations, I would say the most important regulations are related to medical device regulations and the privacy regulations. That can also entails, for example, ethical considerations, balancing of data and so on. These are aspects that we all need to consider typically before we start the implementation. That’s sometimes fundamentally is different than the approach to quickly test features in the market and to test whether it’s really helping a clinician in the daily practice.

Now, for example, if we zoom in into the medical device regulations, these actually are there to make sure that the technology is safe and effective. That is, of course, super important and to make sure that the technology is safe and effective, there is a lot of processes in place to control the process of building the tools. Those also include the necessary validations to make sure that the tool is really doing what it’s meant to do before it gets into production. That process is also on a regular basis audited, a team of auditors coming into the offices of icometrix and diving really into our processes, asking a lot of questions to different people involved to make sure that the tools are developed also according to those regulations.

[0:17:02] HC: Your team has published a number of research articles. What benefits have you seen from publishing your work? l [0:17:07] DS: Yes. As icometrix we find science important. It’s actually almost in our DNA. The reason why is that we believe that the technology we build should be scientifically sound. It should do something that matters from the scientific point of view. There is a medical advantage of the tools we develop. To show that also to our stakeholders, of course, at first our customers would, but definitely also patients or other stakeholders, we have chosen to be quite active in publishing the work in scientific journals, preferably, but also at conferences or white papers to make sure that it’s clear what we’re doing.

This is also particularly important as machine learning of artificial intelligence in general. It’s often perceived as a black box, especially for people that are not so familiar with it. They feel that, yeah, it’s foggy. It’s unclear what it’s actually doing. With those publications, we also try to create transparency, to create openness to all of our stakeholders, to really show that those AI tools are designed in the right way and that they perform in the right way. That’s something that publications are a great tool for, actually.

[0:18:29] HC: It sounds like it really builds trust among patients, among clinicians, among investors, among everybody in your circle.

[0:18:37] DS: Absolutely. Absolutely. That’s what we really think it should do, indeed.

[0:18:41] HC: Is there any advice you could offer to other leaders of AI powered startups?

[0:18:45] DS: I can think of a lot of things, but what I would pick out particularly is on the generalizability of AI models. As I mentioned earlier, building an AI model based on a data set that you can find either on the web or through a collaboration, that is actually not that difficult. There are multiple toolboxes available. What makes it really challenging is to make the AI sufficient generalizable, to also allow it to be used in situations that the model has not seen in a straining data. That is what happens in the real world.

In our case, for example, we can collect MRI, CT images from one hospital, but that is not necessarily representative for the MRI and CT data from other hospitals. My recommendation from the technology perspective would be to think about the generalizability early onwards. There are different ways to cope with it. I think that is something that always should be kept in mind.

The second aspect that I would like to add more from the business perspective is that in my view, artificial intelligence is rather a mean and not the end goal. The end goal is the business question. In our case, it is the clinical value that we want to provide to the clinicians. That is the value that makes them willing to pay for the AI. AI is just a means to achieve it and not the purpose. I think that is something that we always try to keep in mind. Of course, cool AI could be a differentiator, but still, in my opinion, it is always a mean to a higher purpose.

[0:20:28] HC: Finally, where do you see the impact of icometrix in three to five years?

[0:20:33] DS: Well, allow me to be a little bit optimistic here and maybe a bit ambitious, but as I mentioned, in the very beginning, I think we’re in the decade that a lot will change in the field of neurological conditions. I think icometrix will play such an important role in that. I think the way we treat patients with neurological conditions will fundamentally change with more treatments available.

I think that icometrix will be an integral part of the Kappa that people with neurological conditions, they are monitored with digital tools like the icometrix digital tools, but I think also, other digital tools. This information will be used to drive the patients to the right therapy. With that, I think we can save so many lives. We can save so much quality of life for people with neurologic condition. That is something that is today untapped. I’m sure that we will be able to untap in maybe not the next three to five years, but definitely within the next 10 years.

[0:21:37] HC: This has been great. Dirk, your team at icometrix is doing some really interesting work for neurological conditions. I expect that the insights you’ve shared will be valuable to other AI companies. Where can people find out more about you online?

[0:21:49] DS: Well, we try to be quite active on social media, definitely LinkedIn. Of course, our website is also a way to find out more. Of course, feel free to reach out on LinkedIn or Twitter. We’ll be very happy to get in touch with all other companies being active in the field of artificial intelligence and whatever domain.

[0:22:06] HC: Perfect. Thanks for joining me today.

[0:22:08] DS: Thank you very much.

[0:22:09] HC: All right, everyone. Thanks for listening. I’m Heather Couture, and I hope you join me again next time for Impact AI.

[OUTRO]

[0:22:19] HC: Thank you for listening to Impact AI. If you enjoyed this episode, please subscribe and share with a friend. If you’d like to learn more about computer vision applications for people and planetary health, you can sign up for my newsletter at pixelscientia.com/newsletter.

[END]