Aleksandra Zuraw: [00:01:11] Hello today my guest is Sylvain Berlemont the CEO and the founder of Keen Eye. Hi Sylvain, how are you?
Sylvain Berlemont: [00:01:19] Hi, Aleksandra. How are you? Thanks for the invitations. Wonderful to chat with you.
Aleksandra: [00:01:24] Good, good. So, tell the guests, tell our listeners about yourself, about your background, and then later about Keen Eye.
Sylvain: [00:01:33] Sure. My name is Sylvain Berlemont. I’m the founder and CEO of Keen Eye. In terms of background, I’ve got a double background in both science and technology. I first graduated from a French computer science school. And then I made my Ph.D. in biomedical image analysis at Institute Pasteur, also in Paris, and then from France where I originally come from, I moved to the US for about five years first in The Scripps Research Institute in San Diego as a post-doctoral fellow. And then I moved to a Harvard med school in Boston.
Aleksandra: [00:02:08] And what kind of research did you do there?
Sylvain: [00:02:10] Cell [00:02:11] biology, essentially, we were in what we used to call the computational cell biology labs, where essentially you have half biologists and half mathematicians/computer scientist/physicians. And the overall is to develop AI and algorithms in order to make some basic research scientific questions moving forward.
Aleksandra: [00:02:35] And your degree in computer science. Is it in computer vision or did you start being interested in computer vision on the way?
Sylvain: [00:02:42] Oh yeah. I think it’s a fantastic research area. When I didn’t finish yet my master’s degree I was already interested in computer vision and there were at that time really two outcomes for if you are interested in computer vision, either life science or military and I was more towards life science.
Aleksandra: [00:03:07] Oh really? I didn’t know. There was this kind of pathway. [00:03:11] So what led you from that research positions in the US to found Keen Eye?
Sylvain: [00:03:19] Yeah. When I was still in Harvard, I met some connections with the industry in the US. The first company I made contact with was Fluidigm Corporation. It’s a San Francisco-based company. They are doing some microfluidic devices for the life sector. And they asked me for consultancy, for some image analysis platform to help them analyze their single-cell devices. And then when I moved back to France, after my post-doc I came up with another company asking me pretty much the same, but in totally different applications and so on.
[00:03:52] And then I said why on earth I’m gonna redo from scratch for every customer I have. And then I started to see patterns and turn a service company into a product company and then to a software as a service company. And this is what is a Keen Eye today.
Aleksandra: [00:04:10] So you [00:04:11] basically had the work and clients before you even started the company.
Sylvain: [00:04:16] Yeah. And that was really exciting.
Aleksandra: [00:04:18] So I’m curious about the name, why Keen Eye?
Sylvain: [00:04:23] Yeah. That’s obviously related to the core business of the company. Our business is to find insights in biomedical images and sometimes insights that are very hard to find by eye or manually. This is like, finding a needle in a haystack.
[00:04:38] And this is I believe in the company, we have a keen eye for such thing.
Aleksandra: [00:04:43] Very memorable name. So are you also located in France where’s your main location?
Sylvain: [00:04:51] Yeah, so our headquarter is in Paris, France, and we are a team spread a little bit across Europe. So we have people in the UK, , people in Germany, in Italy, but the majority of the people are in France remote of course for such a four our times of look down.
[00:05:10] But yes,
Aleksandra: [00:05:11] Now, everything is remote. So what is that you exactly do at Keen Eye? You said delivering insights from biomedical imaging, but how is it structured? How does it work? What’s your platform?
Sylvain: [00:05:26] Yeah. At Keen Eye, we are delivering AI tools for the life science industry. By AI tools it means we are providing some algorithms and platforms in order to analyze better biomedical images.
[00:05:41] And so this is very large as we, we do have image analysis everywhere in diagnostics, but also in the drug industry. Where you want, for instance, to rely on imaging to quantify your biomarker you are relying on image two better stratify patients or things along those lines.
[00:05:59] There is a ton of things to do in this activity. And we are providing tools in that area.
Aleksandra: [00:06:06] And it’s tissue image analysis, mainly, you say biomedical images. So [00:06:11] practically you would be able to do everything. You do focus on pathology, right?
Sylvain: [00:06:14] We can technically do everything whether it’s a bunch of pixels at work, whether it’s in cell culture, tissue, whole slide tissue, and so on and so forth. So technically the platform is allowed to manage everything. But of course, we are focusing on tissue imaging because it represents a lot of challenges in the industry, but also for the patients.
Aleksandra: [00:06:39] You said it’s a software as a service.
[00:06:42] How does that work? How does it work with your clients?
Sylvain: [00:06:46] So we have software as a service platform, which means that you use that platform through your web-based web browser. So, as you open any browser you have on your basic PC and you can manipulate gigapixels images and do some heavy computation on those images.
[00:07:08] Just staying on your basic PC and [00:07:11] just staying on your office at your desk. So, this is the software as a service spark. We do provide the platform through different types of installations, whether it’s on the cloud, our cloud, the cloud of our customers, or we can also do standalone installations on site installation, but the user experience is the same.
[00:07:31] It’s through a web browser.
Aleksandra: [00:07:34] Okay. So no need to have some special computers, some special hardware with the GPUs, it’s on your end.
Sylvain: [00:07:41] Exactly. Exactly. And this is one of the things that is really associated with a product vision. I remember when I was a post-doc I was manipulating and my colleague biologist, there were many, some scientific software where actually you need to have a PhD to just have a look on the user manual because the interface it’s just crowded with buttons, parameters and things like that. The way we think that the product is willing to have it easy to use interface. Without [00:08:11] forgetting about the power of the algorithm, you can have to a very easy to use interface. And this is what steer the product vision of our platform.
Aleksandra: [00:08:23] So it’s AI, it’s deep learning, probably. Do you do supervise deep learning with the customer’s input as a ground truth or annotations? Or how does that work?
[00:08:35] Sylvain: [00:08:35] So it is. It is deep learning you’re right. And we use any kind of technique that is necessary to bring the solutions to our customers.
[00:08:42] So we can use transfer learning. We can use various different types of deep learning architectures. We can use strongly supervised machine learning, where we have annotations from the pathologist that outline tumors, or fibrosis areas and so on and so forth, but we can also do those more popular techniques where we do weekly supervised machine learning, where we don’t rely too much on the [00:09:11] annotations of the pathologists. We just take the outcome, the patient outcome, for instance, abnormal tissue, disease, not disease, something like that on the global slide. And we let the machine find out what are in the image, the region that contribute to the prediction.
Aleksandra: [00:09:28] Okay. So, both strongly and weekly supervised. Okay. I guess you’ve covered everything. So, what is your mission with this? What do you want to achieve with your company?
Sylvain: [00:09:43] Yeah, our mission is really to support the pharma industry, by being able to accelerate the delivery of the right drug to the right patients, through the use of image analysis.
[00:09:56] And as I said, the image analysis is everywhere, all along the path of the drug development. Whether it is to quantify your biomarker or at the end of the late stage , to select the right patients. And so what [00:10:11] we provide is really this: it’s AI tools to accelerate the drug development. Drug development as is can take 10 years.
[00:10:19] Sometime vaccine can take a less than a year,
Aleksandra: [00:10:23] Hopefully.
Sylvain: [00:10:25] Yes, hopefully. But on average, it’s a very long and very hard process. And one of the ways to accelerate this process is to remove any type of bias you can have and bring more objectivity and bring more quantification results.
[00:10:44] And this is what image analysis is allowed to bring is really to bring standardized, accurate and reproducible results.
Aleksandra: [00:10:52] And during your journey what was something that you wish you had known before or something that was unexpected that you had to change your mind about or rethink or approach in a different way. And now it seems obvious, but at the time you wish you had known that.
Sylvain: [00:11:11] Oh a lot of different things, but if I have to take just one, one, one thing is really focus. Focus is super, super important. At the beginning I remember the beginning of the company, you jump on any opportunity you have.
[00:11:25] You have a lot of people that, that see your technology fantastic. That can enable a lot of things. And at the very beginning, you have a two or three people company. You, you just grab the business where it is. And so you take all the opportunities, but in the long run it is it is really time-consuming and you really have to have a vision of focus and to really steer and operate the company to the right directions. So I wish I knew that from the very beginning, it would have maybe saved me a one or two years.
Aleksandra: [00:11:59] But do you think you could have known that from the very beginning when you were just starting or you have to go through that, like jumping on every opportunity, like you say, to figure out your own process.
[00:12:10] Sometimes [00:12:11] people know exactly what they want to do, but sometimes it’s just part of the process. So, what’s your thought about this?
Sylvain: [00:12:17] I, I think I think there is, it’s a mixed, there is part of the process that you have to go through it. You have to go off the beaten path and, have the trial-and-error process to really find the right way to go.
[00:12:32] After that, then I cannot only speak for myself. This is my first company. So obviously I guess for my 10th company, I may learn from experience,
Aleksandra: [00:12:43] Exactly where to go.
Sylvain: [00:12:45] Exactly. But I think also every activity, every business, and especially every team, teammates you’re working with, they are very specific and special.
[00:12:58] And one thing, which is important. And I see that every day is that you have to really listen to what you have in order to do the right things coming with too many certainties [00:13:11] sometime prevents you to see some opportunities that are existing.
Aleksandra: [00:13:18] You say it’s your first company.
[00:13:19] When did you launch the company? The company, and later, when did you launch the product, your software as a service?
Sylvain: [00:13:27] Yeah we set up the company, I set up the company in 2015 then so we were only one employee. Now. We are about 50 people. And we started on research use only applications to launch the product in 2017 and then we we started launching the product in 2019 for clinically grade applications that are used within clinical trials.
Aleksandra: [00:13:53] So who are your customers at the moment? Who do you provide your services to?
Sylvain: [00:14:00] Yeah, so the majority of our customers comes from the pharma industry.
[00:14:04] So they are either small, a small pharma, big pharma or biotech. So the, [00:14:11] all the questions they have is: okay, I’ve got my data and I need to know whether that data makes sense. In term of I don’t know the level of expression of a certain biomarker I want to quantify that and not have a manual way to do that. So, they come with that request. And so we develop, validate algorithms and then run all the simple analysis for them. So usually our customers, they are, yeah, pharma companies.
Aleksandra: [00:14:44] So are the questions customer-driven or the algorithms driven by the questions the customer have, or do you have also solutions that you can already offer?
Sylvain: [00:14:55] Yeah, so that’s a great question because I need to also give a little bit of specificity from the company. It’s a key differentiator, I believe in Keen Eye. We don’t believe in “off-the-shelf products”. Or “off-the-shelf algorithms” in life science, it’s a very [00:15:11] complex and challenging landscape.
[00:15:13] Any, all the customers they have the, their way to stain their slides, they all have their specificities. So, it’s very hard to say, okay, I’ve got my PD-L1 expression algorithms, and it works off the shelf and it works for every single slide in earth, on earth that’s not true.
[00:15:32] So if companies say that, claim that I don’t believe this is the right way. So at Keen Eye, what we do is we always provide custom solutions. By custom solution, it means that not only we provide algorithms, but we also provide a way to state the performance of the algorithms against the data of the customers.
[00:15:56] So we always provide that. So, we provide the algorithm, and we say, yes, the specificity is around a 95%. And the accuracy is 90% or something like that. But custom solution takes time to deliver. And here [00:16:11] comes, I think one of those specificities of Keen Eye is that we have developed an internal process to develop very fast algorithms.
[00:16:20] So, it allows us to go from weeks or months of development and validation times to weeks or day or weeks that, that really shrank the development times very fast. And this is due to I think two things, a good process, pre-trained algorithms, and also I believe good team the teammates I have to cheer for them.
Aleksandra: [00:16:44] So one question, because you can do transfer learning with the nets that are developed on the natural images. Are there any pre-trained networks on pathology images already? Can you even do that? Is that an option?
[00:17:01] Yes, we do transfer learning a lot.
Sylvain: [00:17:04] So to go from one question to another with of course, some retraining to really target specifically the new [00:17:11] questions, but we rely massively on transfer learning. And actually we succeed in going from one organ to another successfully, but we also find out that we could go from one therapeutic area to another.
[00:17:26] So for instance we did have some experience in ophthalmology. And then we transfer some of models that may be on ophthalmology into oncology. And so we speed up the development process by a large amount of time just by doing that.
Aleksandra: [00:17:42] So basically the more models you develop, the bigger the pool is to draw from, to do transfer learning, and to work even faster.
Sylvain: [00:17:53] Exactly. And I believe that we pre-train models right now. They know quite well what tissue is, what a tumor is because they realize on a lot of experience of training.
Aleksandra: [00:18:05] So why pathology, why digital pathology? You mentioned a little bit that this [00:18:11] is important, but why in particular? What was the thing that made you go into pathology? You could have done radiology, you could have done ophthalmology, which you did a little bit. Anything else? Why pathology?
Sylvain: [00:18:26] I think pathology is a great field, a great field.
Aleksandra: [00:18:29] I totally agree.
Sylvain: [00:18:30] It’s a great thing. It’s heavily related to one of the biggest I would say a therapeutic area, which is oncology.
[00:18:37] And oncology is a highly complex, scientifically complex, and challenging field which we would like, all of us, to have an impact on that. I think pathology really encloses technology challenge, the scientific challenge in great hopes for patients. I think we choose pathology also because there is a lot of, there is a lot to do.
[00:19:03] Radiology that it’s a more mature market. I would say if we talked about markets. They all [00:19:11] have consolidated image format and things like that since several years, several decades right now and pathology is a little bit newer. Although we are talking about digital pathology from a long time for now, but there is a lot to, to do, and there is also a lot to explore and discover, and this is what excites us.
Aleksandra: [00:19:30] How do you work with pathologists for development of your algorithms?
Sylvain: [00:19:35] Yeah, so we do have in-house pathologists that help us to design models and to validate models. We also work with in combination with other pathologists, outside pathologists, for instance, pathologists from our customers, our clients.
[00:19:53] So we, we do for instance multiple annotations. A pool of annotations from multiple pathologists in order to have consensus training that’s something we, we do a lot.
Aleksandra: [00:20:03] Okay. I think this is also important. It’s rarely the case that you have access to so many pathologists that you can [00:20:11] do consensus training. Usually, it’s one person or.
Sylvain: [00:20:15] I actually I think to do consensus training we all agree that this is, I’m not a pathologist by training and, but I’m just quoting pathologists. They do acknowledge the fact that pathologies visual expertise is very subjective can be very subjective and they would love collectively to be able to go more towards standardizations and reproducibility. This is something that I think they all want that. I think there is really two, two ways you can achieve that. The first thing is obviously to get access to pathologists that are specialized in the right indications, in lung, in breast cancer, in other, all the disease. But also, it’s also a technological challenge because you know that the pathologist you may need for certain scientific questions, they are not on the same room. They are [00:21:11] not close to you or in the same office they are spread, scattered. Many in your country, in different countries, et cetera, et cetera.
[00:21:19] So you have to have a SaaS platform. This is mandatory in order to be able to reach out to those pathologists so that they can easily annotate give their own expertise that you can take and transfer into an algorithm.
Aleksandra: [00:21:35] I think this is so important and it would help us so much if we had more standardization, like you say, it is pretty subjective.
[00:21:44] And I experienced it every time., I try to figure out an annotation strategy for an image analysis project. First you figure out, okay, what are the classes you want to annotate? And then. If you want to have more people annotating, you have to convey to them what you want to have annotated.
[00:22:04] And sometimes they’re more experts than you in the field and they disagree. So it is [00:22:11] important to, to agree. And it would be great if we had standardized guidelines that you could just send out and say, this is what we’re annotating for this particular project for this particular problem. Because the number of problems is so vast .
[00:22:26] I don’t think it’s possible in the nearest future, but at least for the common things. And I think that would be a good development and that’s what I strive for when they develop annotation trainings.
Sylvain: [00:22:39] Yeah. And I think having consensus training does not mean you, you lower down everything or you average everything, or you killed specificity of each pathologist.
[00:22:51] This is, I think it would be wrong to go in that directions. But with AI and with especially deep learning, you can integrate that specificity, that viability into a model that will try to take into account this large variability and these disagreement rates. We still don’t want to see [00:23:11] disagreement as a value.
[00:23:13] This is what will allow us to train and teach the algorithms the right way.
Aleksandra: [00:23:20] How do you innovate? It’s a funny question in this context, because this whole discipline is innovative, but we’re just at the beginning of this AI for pathology. So how do you look for the next best thing?
[00:23:35] How do you try them out and how do you invite innovation into the company?
Sylvain: [00:23:40] Yeah, I think that’s a great question. That’s a routine question or daily ask the question. I would say everything that we want to Keen Eye is trying to answer these questions on a daily basis. I think at Keen Eye we do see innovation as a mixed. Innovation is not only technical.
[00:23:58] Obviously the technology is great in AI, et cetera, et cetera. But he’s also organizational. If you are only bringing in algorithms, it does not make a product. [00:24:11] So there is a huge gap between having an algorithm, having an AI model and having the right product that will be used on a daily basis.
[00:24:20] There is a gap and at Keen Eye we’re trying to fill that gap, meaning that we develop AI model, but we also developed the tool that we’ll be able to host these algorithms so that it will be massively used through pathology labs.
Aleksandra: [00:24:37] What is your thought on a computer scientist and pathologists interactions and collaboration?
[00:24:45] How was it that the beginning when you started and how is it now, was there some kind of evolution, and how it was this for you?
Sylvain: [00:24:52] Yeah, I actually, I’m surprised how fast the mindset of both computer data scientists and pathologists changed over the course of the last years. It’s really positive.
[00:25:06] I’m positively a stroke because I think [00:25:11] now it’s a consensus that at least for the majority of the pathologist they acknowledge the capability of AI that could be brought to the community. There, there was a famous quote about AI will not replace pathologists, but pathologist who do not use AI will be replaced by those who are, something like that.
[00:25:34] But that’s a, of course completely true. And I do see now a lot of pathologists from hospitals that desire to, to really be engaged into AI project. There is a a movement now that a pathologist or hospitals are coming toward companies, such as us to really be able to work on new technologies and new ways to get better, to get a more accurate diagnosis or to just [00:26:11] be able to see something that was impossible to see before.
Aleksandra: [00:26:14] I’m super happy to hear that. When I started in this digital pathology image analysis field, I felt the gap a lot. And this is where I started the blog. And one thing that they noticed than -this notion of over promising and under delivering with AI and the pathologists who were hoping for so much, and they had the reality check and then they would withdraw themselves from the whole field saying, ah, it’s not worth it. Anyway. Now I think there’s not so much of that. People on both ends have more realistic expectations. They did accept that it’s still narrow AI. You have to narrow down your problems and focus on one specific thing.
[00:26:58] It’s not gonna solve everything. And for every problem, you probably have to develop a new solution, at least at the moment. So I’m happy to see this as well.
Sylvain: [00:27:09] Yeah. And I think the last [00:27:11] couple of years both ends just learn from the others. I think I’ve attended to great presentation from pathologists trying to educate other pathologists in AI. And that was fantastic AI course, very technical, very mathematically, very technical. And I said, wow, the level is great. That’s awesome. And I think now they all know, as you said, they have realistic expectations. What is good with AI as well is that they do have an active contribution to the design of the model, and they know that this is not something that is, data scientists hidden in the garage that will do something and just delivering back to the pathologist an algorithm. It’s not like that.
[00:27:56] And then they notice that there is a very close interaction. And this is the key I think to success. Close interactions from the visual expertise of [00:28:11] a pathologist, with their knowns that they are surrounded with and the data scientists that will be able to deliver quality models.
Aleksandra: [00:28:23] Definitely. This is an advantage of, eh, AI for pathologists that they can actively participate. Before, as I experienced it as well, there was with the classical image analysis, classical computer vision with the feature definition that pathologist, like you say, would need the PhD in computer science to even operate the software.
[00:28:45] Now they give examples and I think this also closed the gap tremendously. And they can see the performance visually before, as I experienced it, it was a couple of screenshots to show, Oh, it is possible to do that, but you were rarely able to review the whole slide. Now it’s standard. If you don’t show the performance on the whole slide, the pathologist or whoever is [00:29:11] doing the project is not going to trust.
[00:29:13] And then neither will the trust if they don’t see the performance metrics, like you said, you deliver performance metrics. I think it’s crucial component of algorithm development. It’s never going to be a hundred percent accurate, sensitive and specific. And you have to know if you’re willing to compromise on one of these parameters or any other parameter you can choose.
[00:29:34] I don’t know, whatever performance metric is appropriate for project. I think this is a big differentiator as well, but it’s not going to be a differentiator for long because people will not accept if they will not have this delivered. So, you’re definitely in the forefront, but the others will have to follow as well.
[00:29:54] Is there anything else that you would like to tell our listeners?
Sylvain: [00:29:58] I think we got; we covered a lot of things. Thank you very much. It was a, it was fun. And I think it was thorough. I thank you very much.
Aleksandra: [00:30:05] Before we go let the listeners know where they can find you online.
Sylvain: [00:30:09] Yes. So [00:30:11] it’s www.keeneye.ai, this is our website, and they can find all the information about the company.
Aleksandra: [00:30:18] I will definitely include it in the show notes and thank you very much and have a great day.
Sylvain: [00:30:23] Thank you very much. Bye, bye.
