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]

[0:00:03] 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 and planetary health. You can sign up at pixelscientia.com/newsletter.

[INTERVIEW]

[0:00:34] HC: Today, I’m joined by guest Carlos Ciller, Co-Founder and CEO of RetinAI, to talk about medical data for eye diseases. Carlos, welcome to the show.

[0:00:43] CC: Thanks for having me. It’s a pleasure to be here.

[0:00:45] HC: Carlos, could you share a bit about your background and how that led you to create RetinAI?

[0:00:50] CC: Yeah, absolutely. I’m originally Spanish, moved to Switzerland over a decade ago and I’m an engineer by training. I always wanted to do something that was meaningful in the space of healthcare. It was my master’s studies. I basically focused on neurodegenerative disorders. Then there was this opportunity, specifically in the section of machine learning and medical imaging between different switch institutions, working with patients with eye cancer. Then I thought it was a meaningful and impactful Ph.D. to work on.

Very early on in my Ph.D., I realized that I wasn’t likely going to become a professor, because it’s actually really challenging to do that. I still wanted to make an impact and build things that were going to make the life of people better. Over the course of my Ph.D., I was very lucky to meet my co-founders, Sandro and Stefanos. We did our Ph.D.s together at the University of Bern in the ARTORG Center for Biomedical Engineering.

Then after spending part of my Ph.D. in the UK and another startup that was being developed in the lab and seeing that it was possible, I came back and had a conversation with them. We had already had some experience together, participating in startup weekends, working in hackathons and basically, working together under very stressful conditions. We decided to give it a try and there was no better area to focus on, the area where we were doing our Ph.D.s in, and that was in the intersection of artificial intelligence, technology, and machine learning. Here, we have now RetinAI a couple of years later. That was seven years ago and growing and very happy about the decisions we took.

[0:02:40] HC: What does RetinAI do?

[0:02:42] CC: We are a software company. We are software that is enabling the right decisions sooner in healthcare and that, of course, goes a long way. We basically build our flagship product, it’s called Discovery, a transformative AI medical image and data management platform that enables you to organize data, enrich that data with AI algorithms, and introduce some level of automation and optimization. Basically, we build software, we deploy Discovery. It’s a cloud-based solution. Then we commercialize this solution to primarily, pharmaceutical and life sciences companies that they are using it to assess the development of firstly, human track candidates, or bring a new track to the market.

Coordinate, management, and oversight of clinical studies, or to enable the research needed to elevate the quality of care on patient suffering from eye diseases. Basically, we are a holistic solution provider with the Discovery platform, it’s certified as a medical device for the European and the US market. Then you have different AI modules on top that enable you to further enrich that data that you have in the platform.

Some of that are imaging-based. Others are combining multiple sources of input, imaging, pure data that could be generated, clinical data, and demographics information. We can use this to provide different types of insights. More screening classification, like segmentation, or quantifications have been able to do a segmentation for instance of biomarkers that are specifically linked to a disease, or even to predict the progression of a specific patient under certain conditions, with and without treatment within specific populations, which goes in the direction of predictive and precision medicine.

[0:04:39] HC: You mentioned image data, clinical data, a few different modalities there. Could you elaborate on what types of data these are and how you use the different types?

[0:04:48] CC: Yes. As of today, we are primarily focused on the ophthalmology space, or diseases that can be monitored through the eye. We see now neurodegenerative disorders and vascular conditions, we see a lot of publications that are coming out, where optical coherence, tomography, or optical coherence tomography (OCT) is being used to understand the health of your eyes, as well as the health of the rest of the body. We use primarily OCT, color photos, and photography, which has a lot of vascular information, so you can look at the microvascularization of your retina and then derive information about your health conditions, or even your kidneys, or even your brain.

There is a lot of information in this type of image modalities, as well as optical coherence tomography, OCT, and geography, which is looking as well at the flow in the form of a 3D cube of a very small region of the eye. Very similar to what you would get by getting an MRI, but for a very small region of the eye. Those are the main imaging modalities that we work on. Then the typical clinical data that you could collect from an electronic health record. So, specific gender information, age, comorbidities, this type of information. That’s what we are using.

[0:06:07] HC: I imagine, there’s a lot of diversity with respect to different imaging devices, technicians, patient characteristics, perhaps. How do you train models that are robust across these different variations?

[0:06:19] CC: Most of our models are actually device-independent. There are two ways. One of the secret sauces of the company is that around 40% to 50% of the team has actually a very strong academic training, specifically in the ophthalmology space. We know how to model, or how to be able to capture the different variations that you have across different devices. Ophthalmology is a very heterogeneous space. Every device is different, is producing different types of images. When we are building our models, we always try to make sure that we have this diversity in terms of devices, that we are able to generate either artificially, or with the proper data set, the variations that will be inherent to those devices, or different types of patient populations.

The good thing is that you can encode this information during training. I would say, one of the key expertise of the company. Then, of course, we have plenty of collaborations with researchers in Europe, as well as in North America, with whom we work and have data collaboration agreements, to be able to, in a way, work together towards building these models, making them robust, making sure that we get the right diversity. We are also able to test them before we roll them out at scale.

[0:07:45] HC: The focus is mainly on collecting diverse datasets, or simulating some of the variations. Are there other strategies more on the algorithmic side? Or is it focus here as mostly on data in order to capture the variations?

[0:07:58] CC: Machine learning is just but a small part of the process. There are computer vision techniques that you can use that are necessary in order to be able to perform a good training. You also need to curate data. That’s a very important one, because you can have millions of datasets. But if these millions of datasets are not properly aligned, not properly annotated, then you will not have a good enough model for the quality that you expect as a result. In a way, what we have seen after years of working in this field is that quantity is not the most important aspect, that you can just encode quantity, or create simulations, or you can literally simulate those validations.

Quality is the most important aspect. If you work on quality, you will create stronger models, at least in the specific space, and with the knowledge that we have, that has been the result we have seen. Everything we come up with new findings in this area, and we try to publish them, because we believe it’s also important to bring this information back to the community.

[0:09:03] HC: Another access in which things can change is over time, and sometimes you can’t foresee these changes. How do you ensure that your models continue to perform well over time?

[0:09:13] CC: There are two approaches here. On the one side, we have most of our certified models. It’s a frozen model, so you have frozen weights that you need to submit for regulatory approval, and this has different mechanisms, or different processes, depending on whether you’re applying to the European Union, where we have certified models, or the United States. One of the approaches is that we have a fixed model that is frozen. Weights are always the same, and we don’t update this model frequently.

Another approach is the continuous learning. Today, we are following both of them. It really depends on the objective. What you want to make sure, because while the frozen, you can use to capture variations on the population, and you can just follow standard processes for data creation, diversity of data, diversity of devices, if you are working with multiple devices. There is a way to ensure, or to increase the security, or the reliability of the model, the robustness of the model.

Now, if you have a continuous learning model, you need to define different mechanisms in order to provide these type of guarantees. In our experience, having a frozen model can help, so you could make it specific to certain populations, under specific conditions on certain devices. This is something that the FDA is going to ask for in many cases. Then you can expand the scope of those models with post-market clinical follow-ups, or post-market surveillance, or different real-world evidence studies that you can have on the same model with a broader population. It really depends on your strategy.

I would say that you need to be able during training to create models that are robust, that provide results that are really repeatable so that you have the same patient coming again, taking another scan, and being able to show very similar if not exactly the same results. You need to keep these learnings during the process of building such a model, because otherwise, you may just come up with huge mistakes that are going to set you back. Luckily for us, I believe that having this knowledge in the field has helped, and being surrounded by very talented individuals and collaborating with some of the research centers we collaborate with are very knowledgeable and happy working in the field as well for many years. We are happy to have them as licensed supporters in building these models.

[0:11:49] HC: So, you mentioned regulatory processes. One of the effects of this is that you have to keep your model weight static and the process for updating them requires going through the regulatory process again. Are there other ways that regulatory processes influence the way you develop machine learning models that might be different if you didn’t need approval from someone?

[0:12:11] CC: Yeah, absolutely. I think that you are absolutely right. We need to have, or at least for us, from the beginning, we realized that we had to have a very good process in place. Good thing is that sometimes it’s similar to the process that you would follow for proper academic practices to develop good quality research.

There has been a lot of effort in that direction. For instance, there was a paper that came out, I think, in 2018/2019, with the consort AI, spirit AI guidelines to develop clinical studies for clinical trial protocols. For us, it was really relieving to see that we had been working on models and building models that followed certain criteria, and 95% of the things that were in this paper published in The Lancet by Xiaoxuan Liu, I recall. I think we were more or less there.

The most important thing is that you need to educate your team. This is a necessary process that you need to use to develop your models. When you are in the early stages of research, maybe you can just work around to see what is working, and what is not bad. When you have a certain level of confidence that something could be used one day on a patient population, physicians could rely one day or some of your customers could rely one day on the output of those models, and that could influence the life of a patient. You need to be very careful with it.

I think regulatory bodies are catching up with that, especially the US, FDA Is doing a good job on that front. In the past, there were mistakes. For instance, I’m sure you recall this issue that IBM and Watson had in the past with radiology, where they were getting some very specific quality results in European centers, and then they moved to the US, and then there was a drop in quality. I think the regulations that we have today, and some of the guidelines and support material provided by regulatory agencies and some of the leaders in academic space are precisely to help you not commit the same mistakes that others committed in the past.

Of course, it is not the same as a physical device. Some of the regulations that we have, ISO 13485 is coming out from regulations for medical devices, or specifically for individual diagnosis, depending on how you look at it. It’s not the same when you are building software. Regulations need to adapt to the new times. I think it’s important to, from the beginning, keep always these regulatory processes in place, because you will eventually find them along the way. If you are serious about it, if you want to bring a good, reliable, robust product to the market.

[0:14:47] HC: Your team has published a number of research articles. What business benefits have you seen from publishing your work?

[0:14:52] CC: I think it’s huge. First, on the one side, I feel that we have a moral duty. If we find something that is going to benefit the life of millions of people, then you can just put it out there. I think coming from a research background myself and being a Ph.D., I think that it’s important to share your research, and you can still make a good company out of sharing your own research, and letting others build on top of what you are building. I want to advance innovation. I want to work with others to advance their innovation. At the same time, you will get the endorsement of people from your community.

Having peer-reviewed publications, where there is a process, where multiple reviewers are going to look at your data, going to look at your results, your model, your process, the datasets that you are using, this is going to help you refine the quality of the work that you are putting out there. For me, it’s really important. I think we have more than 40 publications to date in different journals, and Nature scientific reports, especially in ophthalmology. I cannot even recall all the publications that we have, but I’m proud of these publications, and also, I also believe it’s the right approach.

When somebody asks you if you have done the proper research behind a model, you can always back it up with a peer-reviewed journal, whether the community has already validated those results for you, and this is going to help as well when you need to convince customers, physicians who are going to be sometimes the end-users of your solutions, as well as regulatory bodies.

[0:16:17] HC: AI has been in the headlines a lot lately with generative models, large language models like ChatGPT, and text-to-image models, like Stable Diffusion. How did the latest generative AI advancements influence what you’re working on?

[0:16:30] CC: I think it’s a huge opportunity. I cannot believe that the field is moving this fast, and it’s an ever-increasing development speed. I still believe that we are just scratching the surface of what is possible with generative AI, and we need to be very careful. As from what I stand by, I think that technology and AI, or generative AI is just a technology. It’s not a product. A product may use different types of technology in the background. Being able to build the right wrapper around the technology that is available is the nuance that is going to make a product successful or not.

For us, we are already applying generative AI technology in our flagship product, Discovery, mostly using it to be able to identify and navigate the sea of data points that we have in the background of structured data. It’s mostly building an application layer on top of this generative AI technology. I feel, as well, and having done a Ph.D. in this field, it’s a hard-to-swallow pill, but AI is just a means to an end. From where we stand, AI is just the technology that is going to enable you to make better decisions, is going to enable you to access your information in a more properly distilled way or encapsulate all the information from our clinical study in a model. I think that that’s some important learnings that we had today. For me, AI is just a means to an end.

In the past, you could just go out there and raise funds saying, “I am an AI company,” just with that. I think that the hype for generative AI is closing very quickly. There are only a couple of companies that can use that hype, because of their background, or because of where they are coming from. Then afterward, you need to show results. Showing results means building a product that people are using and people are willing to pay for.

I think, I’m very excited. I think for us, it’s helping us because we have already our data management solution and it’s in power with generative AI. But I’m looking forward to see how the field is going to move generative AI in the next two years. At the same time, when you are working with patients and patient data, you need to do everything from a proper regulatory frame perspective. I think generative AI is going to take some time to be validated. Those are my thoughts.

[0:19:01] HC: Is there any advice you could offer to other leaders of AI-powered startups?

[0:19:06] CC: One of the most important learnings for me was that AI is just a means to an end. You still need to deliver a product. I think you should let the market define the product. You should rush to bring it into an MVP that you can put in the hands of the people that you believe early on are going to be your customers, because this may change. For instance, for us, when we started the company, we thought we were going to go in one specific way, distributing only our AI models, but then we realized that we had to build a platform as a whole, because the real problem was not the AI part. The real problem that we were solving was the same problem that has been around in the healthcare space for a very long time, is data management and organization.

We are down decades of digital data being generated more than ever before every day, and we don’t still have tools that are removing those silos. I would say, that follow exactly the customer that you want to serve, build something that you can very quickly put in their hands and listen to them, and also listen to what they are not saying. You can just have very quick, iterative feedback loops, so you can just refine that product. That’s what I would tell to my past self. Another thing is to start with a niche, or an area where you are the best, or the product that you’re building is the best, and expand from there. Because trying to take over the entire market, it may work for fundraising, but then when you need to sell a product and you need to compete out there, you need to have a very solid value proposition. That would be pieces of advice for any new potential leader that is willing to build a stack that is already in a startup in the future.

[0:20:47] HC: Finally, where do you see the impact of RetinAI in three to five years?

[0:20:51] CC: I think that we are very well-positioned to go deeper in the eyecare space, or using AI as a window to the rest of the body as a whole. We have been preaching this message for the last couple of years because your eye has a lot of information, especially structural information, and vascular information. It’s almost the perfect screening mechanism to be able to identify early signs of Parkinson’s disease, multiple sclerosis, and even Alzheimer.

Some of the certified biomarkers that we have are linked to early signs of those diseases. I’m very excited to see that the technology that we have built and those biomarkers that we are able to monitor through the eye are going to enable us to expand into new therapeutic areas in the near future, especially neurodegenerative disorders and vascular conditions. We are going to expand into those areas, into different mechanisms, and pathways, one, using the eye as a window and building a library of certified biomarkers that will enable a different type of screening, bringing down management and monitoring of your own health to the level where even you and me, we could go to a pharmacy and be able to monitor our own eye health and also, Discovery as the solutions provider across therapeutic areas.

We are working today with many different pharma companies and clinic groups around the world. We see an opportunity to use the same technology stack to expand into new therapeutic areas. I’m really excited to see that the idea that we started a couple of years ago, which was bringing innovation to the people who need it the most, and for us, everything was about the patient. Even if we took a bit of a longer walk across pharma clinics, but then we are going back to the people who we believe are going to benefit the most from this technology, which is the patient. Really looking forward to it.

Lastly, I would say that we believe in the four P’s of healthcare. Prevention, participative, personalized, and precision. Among them, we believe that in the future, AI technology is going to be used to elevate the quality of care of treatments that are going to be put in the hands of patients, and we believe that we are going to have an important role in opening that field. Looking forward to digital precision medicine and the possibility of using software to elevate the quality of care of decisions that are today happening in routine medical care, or that the patients are even taken on their own. Hopefully, we will have an important role in that future, too.

[0:23:41] HC: This has been great, Carlos. I appreciate your insights today. I think this will be valuable to many listeners. Where can people find out more about you online?

[0:23:49] CC: I would say that LinkedIn, our website, or simply drop me a line on Twitter. I’m always happy to follow up and happy to see if there is anybody who wants to follow up or has an interest, I’m happy to talk.

[0:24:03] HC: Perfect. I’ll link to all of those in the show notes. Thanks for joining me today.

[0:24:07] CC: Thank you, Heather. Thank you, everyone, for listening.

[0:24:10] 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]

[0:24:19] 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]