We sat down with Therapixel CEO Matthieu Leclerc-Chalvet to learn more about Therapixel's mission, their product MammoScreen, and their thoughts on transparency in medical imaging artificial intelligence. Therapixel, which develops AI software to assist radiologists in the diagnosis of breast cancer, won the Digital Mammography DREAM challenge organized by IBM in 2017 and recently received a new 510K clearance.
Q: Can you tell us about Therapixel and its mission?
Matthieu Leclerc-Chalvet: Our mission is to reassure women faster. And this is because we're dealing with breast cancer screening, and we know that luckily, most of the time, like more than 99% of the time going to breast cancer screening means there's nothing to see, and there's nothing worrisome. Conversely, if there is something worrisome, it's better to catch it early, because we know that 90% of breast cancers can be cured if they are caught in time.
The whole purpose of our AI is to enable the radiologists who are conducting breast cancer screenings to be in a position to be certain enough to reassure the women that either there's nothing, or they will be taken care of because there's something, but it can be dealt with efficiently.
Q: What is your key product and how it can help radiologists?
MLC: We know breast cancer can be cured, and we know still that a number of cancers are being missed, and even cancers that are seeable on the screen are being missed – we know that because there are a number of statistics, and if you compare different types of programs for screening breast cancer in different countries, the detection rate is about 5 for 1,000 or 5.5 for 1,000 in the U.S., but other countries in Europe in particular, it can be up to 7 or 8 per 1,000. So you could wonder, why we catching more cancers in Europe than we are in the U.S.? One of the things is that there is single read in the U.S., whereas they are double reads in Europe. Two radiologists are reading every screen. Now, that's just catching cancer. The problem behind being efficient with breast cancer screening is that there's a cost-effective issue at the bottom of it, and breast cancer is becoming unsustainable.
There's a shortage of breast radiologists. There's burnout. There’s an aging population that brings more and more potential patients. And there's a lot of inequity in access to breast cancer screening even in countries where there should be access. So, as stated before, it's a screening problem. It's like finding a needle in a haystack – and we could have concentrated on finding the needle directly, but we thought it would be easier if we first burned the haystack. What we mean by that, is that we want to make sure that the radiologist who is the expert is concentrating his or her time and expertise on the cases that do require human expertise, and our vision beyond the mission is that we should try to eliminate the time wasted by radiologists reviewing cases that do not need review.
"[...] It's a screening problem. It's like finding a needle in a haystack – and we could have concentrated on finding the needle directly, but we thought it would be easier if we first burned the haystack."
So, this is why we design our product with a score. It’s what we call a superpower or a sidekick of the radiologist, if you will. To that end, we were the first to show a software that displays marks only in one in four cases. So, most of the time we don't alert the radiologist on, you know, stupid things. We just mark when there is something important, so 75% of the time they get a green score, and a green score means there's nothing to be worried about. If you are reassured, you can move forward. If you feel confident, you move forward.
Then all the time, we can alert them in particular if they are getting tired. We know a lot of radiologists read mammographies in batches of 50-100 at a time, and this is a human thing. You get tired, you get distracted by a call or something, and you might miss something at this point. So that's the safety net that we're bringing. So, this is where our score comes in. For more than 10 years we have provided tools for medical imaging, and what we want to bring is the right information at the right time. This is why we co-developed the product with radiologists trying to understand really what they wanted to have. Of course, radiologists were not asking for an AI, they were asking for certain things that the AI could provide. So that's the confusion that we hear sometimes, “oh, it's an AI product.” Well, we don't care if it's AI. What we care about is, what do we bring to the radiologist that helps him or her do a better job and faster. So that's the score. And, as you can see on this little animation, they use their standard screening environment, and when they need it, they can click on the little notification here to display the report from MammoScreen if they need it. If they don’t need it, it stays just as a notification, and they don't need to be bothered.
We have also been the first to take in consideration of as many images as we can. Particularly with combining the 2D and the 3D, but also the prior image. And this is a very important thing – this is something that every radiologist would do. If they see something, a mark or something they find suspicious on an image, they would be looking at the image from a prior screening exam and say, “Okay, was it there? It was there, and maybe it's not worsening if it has stayed the same.” If it wasn't there, it's certainly more worrisome because something has appeared in the meantime.
This is why our software has been, for the last couple of years, using priors to better the prediction. As we can see in this slide here, using the prior [images] on the right-hand side shows more certainty in in the scoring. You know we are more in the middle range of scoring when you don't use prior [images], and we push to a higher scoring with the use of prior [images]. Conversely, some marks could be deemed yellow without the prior, and could move into green if we know that we have the prior – so that's a very important thing.
Of course, all of this needs to be made vendor neutral, so we are multi-vendor. We treat images coming from multiple vendors, we are cloud processing, so we don't impose any heavy infrastructure on the customer, and we are very flexible with the integration, meaning that we are nearly a zero click workflow. Everything is sent automatically, and really, there's no additional work for the radiologist to get access to this important information.
"[...] We don't impose any heavy infrastructure on the customer, and we are very flexible with the integration, meaning that we are nearly a zero click workflow. Everything is sent automatically, and really, there's no additional work for the radiologist to get access to this important information."
What is also important is that we are able also to measure this gain in productivity that we claim, because constantly in the background, we're monitoring what's going on. And in this particular study we showed that we can save 27.89% of time, measured in real time, because we know what the radiologists are reading, and we can compare it to their baseline. Without surprise we can see that in the green cases, 75% of the time they can save a lot of time. But even in the red cases, even when it's cancer, or they are very suspicious of cancer, we can save a lot of time because they get reassured in their decision-making.
Q: Can you tell us about some unique strengths or benefits of your product or company?
MLC: A unique thing we use is a clear link between the recommendation from the software and why the software is making this recommendation. That's what we call the “no black box AI,” because when the recommendation is, you know, high suspicion, you should basically recall this case and explore further. We say specifically, well, this is because there's a score of 9 in this case on the left breast. And why is there a score of 9? Because there's a lesion with a score of 9 on the left breast. It seems very simple and very obvious, but this is unique.
We are the only solution that brings a clear, direct link with the same scoring system from the lesion to the breast to the case, and that means a lot for radiologists. They understand instantly, they build their trust in the software very quickly. They know the limitations as well, of the software. They know that the yellow zone is where the software is not completely sure, this is where they should pay attention. But when it's red or when it's green rapidly, they can trust it, and they know why they trust it, because they've seen what the software has seen.
So that's a very, very specific thing, this score. The other one I mentioned already is using the combination of all information available – the combination of 2D, 3D, and the prior, this is also a unique thing. This is what provides a much more reliable score when it comes to prediction.
Q: What are your thoughts on the Transparent-AI program?
MLC: We’ve been, for a few years, collaborating with the ACR to work more on the future of AI for breast cancer, and as you noted we've been working with radiologists to design a product. But we thought when it comes to the long-term roadmap, the more research-y type of things we have in mind, it would be better to involve the ACR from the start, because they are also the ones who are guiding the profession into the future. And we were very pleasantly surprised by the welcome attitude we had, and people really willing to discuss the future, even if it was sometimes bold and adventurous. But they welcome our ideas, and we could debate. That was really good.
And then, when we saw the Transparent-AI program coming in, we said, “yeah, we're in. Absolutely.” This is at the core of what we do. Almost in the same period of time, the French Society for Radiology came up with exactly the same program asking volunteers to file a number of detail information sheets for each product and put it on websites so radiologists who have a full access of information and always format it the same way. And so we were totally on board with this transparency initiative, I think it’s great. We’re proud to be part of the first cohort, so to speak, and on trade shows we proudly display that we partner with the transparency program. I think it’s very important for the more general acceptance and understanding of AI. I mean, people can continue to disagree with AI if they will, but being transparent is a very sure way to build trust.
Q: What can we expect from your company in the next few years? Are there any exciting developments on the horizon?
MLC: As I was mentioning a minute ago, we've been working with radiologists in the field as well as the ACR, on what should be the next generation of products. The first thing we noticed is, and also to comment on your number of clicks that the radiologist has to do during interpretation.
Currently, we noticed we were only dealing with image interpretation, and when you map out all the steps that have to be done by the radiologist for a complete cycle of screening, there's a lot of other things. That includes gathering information from different systems, doing a lot of other activities. So, we thought we could help and actually, radiologists quickly mentioned, “Well, that could be good. I like what the AI is doing for decision making, but it would be wonderful if the AI could help me also with the 13-15 clicks I have to do to collect all the other information. If you could collate everything, then I can concentrate on making my judgment on this case and facilitate the reporting.” So that's exactly what we've been working on. We just received a new 510K clearance, and we'll have very exciting news about covering a lot more of this workflow to help radiologists, including in reporting.
"We just received a new 510K clearance, and we'll have very exciting news about covering a lot more of this workflow to help radiologists, including in reporting."
But beyond that there is more on the horizon. We are working on continuing to use more data, a bigger variety of data to provide information. And you know, as mentioned, we use 2D, 3D and priors combined. We're looking into using other types of information for risk prediction or looking more into the future, helping the radiologists making the best decision, and more on the technology side, we've also started work on foundation models and new types of models for AI. I mean, AI is a generic word, it evolves almost every year, and we continue to do a lot of work to be there with the new generation of algorithms that are coming.
Q: Do you imagine that Therapixel will stay in the breast imaging space? Or do you think that you'll expand into different areas?
MLC: We are where we are, because we choose very consciously to do only breast. At the juncture back in 2017, we won a global challenge on mammography, and we were in in the top 5 for lung cancer. So, we could have pursued the two – we choose to pick the one where we were the best and to be the best. Part of our success is because we have concentrated so much on just one area. There's still a lot to do. I mean, we look at all these activities in the screening, we still have a lot to do so, for now I think we will stay on breast and continue to be the best possible in breast. Now, of course, all the knowledge we are accumulating doing this can be re-employed otherwise. Actually, the breast cancer work we're doing is the second life of Therapixel. We had a first life when we did something different that helped us do what we do now, and maybe who knows what we will do next? But one thing at a time. Also, from a geography point of view, we are extremely focused. We developed the product primarily for the U.S., and now we're starting to look into Europe. Because we think that focus is the best way to reach excellence.
Q: What do you envision for the future of AI in radiology?
MLC: It is very difficult. I mean, I've been working in AI for imaging for 10 years. So, it was Therapixel and another company before. It is really, really difficult to foresee anything beyond two to three years. I mean, the number of models and capabilities of AI continue to be surprising. Things that are the flavor of the day turn out to be not usable in the end. It is a very difficult exercise to do. I think one thing is true, that AI is not going away because AI is more and more recognized as a technology, and not as something that will swallow everyone.
I think radiologists will continue to use AI more and more, so there's a lot of collateral work to do in making sure, in the curriculum of radiologists, that they understand AI, and they can use it to the full extent. I don't like to make predictions, because, especially in the eye, it is quite difficult. But it’s here to stay. It will continue to be a matter of being excellent at what you're doing, being focused on what you're doing, and working with conditions – that's not something we can do in isolation.
Prudence and pragmatism in the approach, do your best, work with people, use your tools, don't do your stuff in your ivory tower, and be transparent in what you're doing. I think that's the key thing.
To see their full range of medical imaging AI products, visit Therapixel on AI Central.
This interview is part of a series interviewing medical imaging AI manufacturers on AI Central to help our members better understand the imaging AI marketplace. Since its inception in 2018, the ACR Data Science Institute® AI Central database has evolved from a short online list of FDA-cleared imaging AI products to the most complete and up-to-date online, searchable directory of commercially available imaging AI products in the United States. More than 200 software as a medical device (SaMD) FDA-cleared products have been curated by more than 100 manufacturers, and thousands of radiologists per month access the site in search of suitable AI solutions. Learn more on AICentral.org.