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]

[00:00:03] 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 in computer vision for people in planetary health. You can sign up at pixelscientia.com/newsletter.

[INTERVIEW]

[0:00:33.6] HC: Today, I’m joined by guest Erez Naaman, co-founder and CTO of Scopio Labs, to talk about blood cell morphology. Erez, welcome to the show.

[0:00:42.2] EN: Hi Heather, it’s great to be here, thank you.

[0:00:44.8] HC: Erez, could you share a bit about your background and how that led you to create Scopio Labs?

[0:00:48.0] EN: So, my background is physics and math, my cofounder is also a physicist, his name is Itai and the way that it started was, we were both students for vast degrees of universities. The university and we drove back and forth every day for about an hour to school and tried to think of ideas of what to do and we both realized we wanted to do something that was for the good of the world, so something in healthcare or in that vicinity.

So, we started walking the halls of hospitals, looking for needs and talking to people there and trying to understand what is it that doesn’t work and one day, we passed by the counting room, which is where they sit next to microscopes and look at the cells of you know, blood cells under the microscope and count them manually, and we said, “There’s no way that this is how things are done in the 21st century.”

From there, it started off by wanting to do something with AI to help them and then found out that there is no platform for scanning those samples. So, it ended up being that we started off wanting to start with AI, then found out that the problem is actually in physics, in how to do the actual imaging so we solved that problem first and only then we created that unique data and then we start applying AI to the analysis, so we kind of closed the circle on it. So, it was more of a bit of luck, a bit of tenacity that led to where we are.

[0:02:11.5] HC: So, what does Scopio do today and why is this so important for healthcare?

[0:02:16.1] EN: So, Scopio manufactures platforms for analysis of morphology of blood and bone marrow. There is a common misconception that a complete blood count, which is the most common lab test in the world is an automated test but is actually not a diagnostic test, it’s a screening test. Every sample that has any kind of clinical indication of something being out of the ordinary, the complete blood count raises a flag and it actually goes to someone and looks at it under a microscope.

That’s done about 600 million times a year, which is a little-known fact but it’s actually extremely manual and so what Scopio does is provides the platform that digitizes this process, makes it a lot more consistent, faster, more automated, and in the future, will enable new diagnostic horizons directly from simple blood tests.

[0:03:08.1] HC: And what role does machine learning play in this technology?

[0:03:11.7] EN: So, it is actually quite essential to the analysis. We scan the blood sample, which contains vast numbers of cells and we analyze them using machine learning. We actually have to differentiate between many different kinds of cells in the blood and that is done using machine learning, which is kind of the basis for our analysis, right? So, there’s a first step of the imaging, which is done without machine learning.

It’s based on super-resolution under a technology that we’ve developed but after that comes the machine learning for the analysis of the sample.

[0:03:43.8] HC: So, the general premise is that we use machine learning on these images, you can detect each individual cell and then classify it into one of several different types that are important for the complete blood count.

[0:03:55.5] EN: Exact. So, it’s for the differential, which comes after the complete blood count but other than that, it’s exactly what you said. So, we scan the sample, we run the textures that find all the different cells and then we classify them into many different classes, and then report that to the user so they can review the results and decide if they accepted or if they reject or want to correct anything, and that goes directly to the patient’s records.

[0:04:22.2] HC: When you founded Scopio in 2014, deep learning was still quite new. How’s your approach to AI evolved over the past nine years?

[0:04:30.2] EN: In terms of the approach, I think we saw it the same way that we do today in terms of its importance but I think that the infrastructure for using ML has greatly evolved. Back when we started, a lot of the things had to be developed in-house in order to simply have the infrastructure to build, to train, to run the AI, to run it on a GPU, and today, so many of these things are things you can just buy tools for or use online tools for.

And they speed up the operation, make it a reliable process, a cost-effective process, and requires a lot less engineers to get to the same result. So, I think it’s more about how we do it than why we do it that has changed.

[0:05:18.2] HC: What kinds of challenges do you encounter in working with blood sphere images in a particular and training models based off of them?

[0:05:25.1] EN: I think the most severe problem is that when you’re dealing with blood cells, a lot of the cells are very prevalent but then you have classes that are abnormal that you’re looking for, which are extremely rare. They only happen in specific kinds of disease and these are the ones you want to catch in many cases, right? So, you don’t want to just – the things that everyone sees every day.

But for the things that people will miss because they rarely see it, because there are very few of them in a specific sample and they come and they’re rare between samples as well and so, getting the large enough data set to get a reliable classification on specific more rare cell types is the most difficult problem in my opinion.

[0:06:08.2] HC: In selling that, do you mainly focus on data, as you mentioned, collecting more examples of those rare classes or are there algorithmic ways you can tackle this as well?

[0:06:18.3] EN: So, it’s always a combination of the two. We have to collect a large enough data set and I think that’s always the case for machine learning. You do have to have enough samples for the algorithm to learn from that and we also wanted to be generalized well enough because the system then goes out to different areas in the world with different staining, with different procedures, and different people reading the results.

So, there’s a certain baseline that you need for something that’s on a clinical level but apart from that, I think that getting – just apart from getting enough data, there’s always the level of algorithm that you have to develop in order to get a good enough result, to be able to reject the corner cases, and prepare the data well enough for the machine learning algorithms. So, it’s a mix of the two.

[0:07:09.3] HC: How does a regulatory process affect the way you develop machine learning models? You know, for example, are the things you do differently or at a different point during development than you might not if you weren’t going for a regulatory approval?

[0:07:22.2] EN: So, there are a few differences I think between regulated and nonregulated machine learning but they’re not in the basic structure of what you’re doing but more in one healthcare full, you have to be, for example, to separate your training foundation test sets, that’s critical and to check how well your model generalizes because again, it goes out to the world and is used for clinical use.

It goes through a regulatory body that checks everything and that also means that you’re not able to update as quickly or as often as when you do it for something that’s called more consumer-type applications, which means that you have to go through more rigorous internal validation, internal verification, validation processes, and external ones. It’s very important to not lose the confidence of the medical community when putting a product out there in machine learning because they deal with human lives.

It’s as simple as that but I think other than being more careful in our processes, the machine learning in itself in the end is more or less the same.

[0:08:34.4] HC: So, it’s largely a case of thinking more carefully, more thoroughly, and perhaps earlier about validation.

[0:08:41.8] EN: Right, it’s kind of like doing the same things but with a higher bar. You know you’re going to be judged with that high bar at the end so you have to design your process from the get-go to meet that high bar.

[0:08:53.3] HC: How does your team plan and develop a new machine learning product or feature? What are some of the first steps you take in the process?

[0:09:00.2] EN: So, in a way, we look at it backward. Machine learning is a tool and not a goal. So, we always start with the patient in mind or the user. So, we always start with the doctor, the lab employee, the patient, and think about what the feature our product should serve, and then only then comes the question of where and how will machine learning serve that need. So, we never start with the technology, always with the product.

[0:09:31.3] HC: So, once you have that defined what they need as for the end user and at what point do you decide that machine learning is the solution and what are some of the first steps you’d take once you have identified machine learning as a solution?

[0:09:44.5] EN: At that point, we will define the criteria that the product has to meet. For example, what level of precision and then from that, we will determine the size of the dataset and the tools we would want to use for that, like does this require segmentation, does this require classification, detection? And from that, from understanding the size of the dataset, we will start a data collection process.

So, that is a first step, where we collect samples from partners in order to get enough data to train and that helps us build the timeline because they’re the first step of data acquisition and then there is a step of development, maturation, and then verification and of course, going through the regulatory process with it.

[0:10:33.0] HC: How do you decide what type of machine learning model to apply? How do you work through that process?

[0:10:38.2] EN: First of all, we have for many because we are already FDA approved for many of the tasks we already have tried and tested algorithms that we use and of course, we follow the state of the art but when we’re trying to do something new, we typically will try a few approaches at the same time and choose the best one. So, we’ll take two or three approaches that academically or by industry standards meet or are fit for the task that we want to perform, and then we will test those head-to-head and choose the best one.

[0:11:12.9] HC: So, it sounds like there might be a little bit of research to figure out what the state of the art is for solving that type of problem, and then certainly a fair amount of experimentation to evaluate each and compare them.

[0:11:24.5] EN: Yeah, and I do think that as time progresses that part becomes less and less central because one, we have a lot of tools already that we know how well they perform and are already out there and I think the industry also is maturing at such a rapid pace. Everyone is dealing with AI and so the front runners are clearly becoming the leaders with time. So, it is much easier to choose the right tools for every task as time progresses.

[0:11:53.7] HC: Yes, and machine learning is advancing quite rapidly right now. There are new advancements hitting the headlines more frequently than ever. Are there new developments in computer vision or AI more broadly that you’re particularly excited about and can see a potential use case for Scopio?

[0:12:08.1] EN: Yes. So, I think it relates to what I mentioned before about where our cell types, I think that there are two really important advancements that are constantly happening. It’s not just like one thing that’s happening, which is dealing with small datasets. So, how to use smaller datasets where you don’t have a lot of examples of something but are still able to generate a strong model for those, and the second one relates to that, which is when you have large differences between the classes.

So, for example, when you have certain classes, they all have to run on the same product and some you have very little information about and some are, you have tremendous amounts of information about it and still, you’re required to the level of maturity of the product on all of them at the same time and I think that that is also advancing quite quickly and I think we’ll be something that in five years, will be kind of just your run of the mill product available for AI, where today, these things are very much done per case still.

[0:13:16.5] HC: Is there any advice you could offer to other leaders of AI-powered startups?

[0:13:20.8] EN: Related to what I said before about product and features, it’s never about the AI unless what you’re doing is developing infrastructure for AI. It’s always about the product. AI is an incredible tool, I think probably the most amazing technology to hit the market in decades. Definitely one that advances the fastest but it is still a tool for achieving something. Your product needs to do something and so it’s best in my opinion to try to avoid starting from what the AI breakthrough is but what the product should do.

The technology, part of the reason for that is that the fact that the technology is moving so fast also means that whatever technology breakthrough you think you’re using right now in five years will be obsolete and that means that if you started off with a good product, you’ll use different tools to achieve the same goal in five years but if you started off with the technology you’ll be left behind.

[0:14:20.2] HC: That’s some good advice and finally, where do you see the impact of Scopio in three to five years?

[0:14:25.4] EN: Well, hopefully, since now we’re on the market that I think the main impact will be that I hope we’ll be in every hospital in every lab and we’ll make blood testing and bone marrow testing much better than it was until today and I think that is the first and of many impacts that I would want to see and the second is that I think that people will start seeing our new applications and future things that are coming, which are beyond what the industry has imagined until today.

And that will enable just a new wave of first-line diagnostics, first line being the test that you run at the moment you walk into a facility and they take your blood and they just run tests to see what’s wrong with you and we will be able to make those so much more powerful than they are today and that will lead I think to generally better health in the population. So, that’s what makes us wake up in the morning and me personally as well, and then I hope that that’s where we’ll be.

[0:15:26.7] HC: Well, I look forward to following you and see where this goes. This has been great. Erez, I appreciate your insights today. I think this will be valuable to many listeners. Where can people find out more about you online?

[0:15:37.9] EN: You can go to www.scopiolabs.com, there’s a lot of information there about our product and also about open careers in the company.

[0:15:47.0] HC: Perfect, thanks for joining me today.

[0:15:49.2] EN: Thank you so much. Thank you for having me.

[0:15:51.4] HC: All right everyone, thanks for listening. I’m Heather Couture and I hope you join me again next time for Impact AI.

[END OF INTERVIEW]

[0:16:01.3] HC: Thank you for listening to Impact AI. If you enjoyed this episode, please subscribe and share with a friend, and if you’d like to learn more about computer vision applications for people in planetary health, you can sign up for my newsletter at pixelscientia.com/newsletter.

[END]