Exploring the Horizon — Episode 7
Interview with Prof. Bram van Ginneken at ECR 2019
Introduction
This interview was made during the ECR 2019 meeting in Vienna, one of the largest radiology meetings in the world, with over 30.000 attendees. It is the second of a series of interviews that was organised by the European Society of Medical Imaging Informatics (EuSoMII). These interviews can also be listened to on the EuSoMII on AIR podcast.
Prof. dr. Bram van Ginneken is Professor of Medical Image Analysis at Radboud University Medical Center and chairs the Diagnostic Image Analysis Group. He is founder of Thirona, a company that develops software and provides services for medical image analysis. He studied Physics at Eindhoven University of Technology and Utrecht University. In 2001, he obtained his Ph.D. at the Image Sciences Institute with a thesis on Computer-Aided Diagnosis in Chest Radiography. He has (co-)authored over 200 publications in international journals. He pioneered the concept of challenges in medical image analysis and is founder of the grand-challenge.org platform.
Interview
Erik: Hello friends, here we are again at the European congress of radiology in Vienna. This is Erik Ranschaert from the European Society of Medical Imaging Informatics (EuSoMII). Today I was able to catch Bram van Ginneken, who is professor at the Radboud University in Nijmegen (UMCN) and chairman of the Diagnostic Image Analysis Group (DIAG). Bram van Ginneken is really a pioneer in computerized analysis of medical image analysis and in developing algorithms based upon machine learning and deep learning. So, this is a very interesting person to talk to, and please listen to this interview. If you like it, you can share it with your friends. Thank you.
Erik: Hi, this is the second part of our live podcast at the ECR in Vienna. Today I have invited to Bram van Ginneken to talk to. Hi, Bram. I’m happy that you’re here. Maybe you can introduce yourself briefly.
Prof. Bram van Ginneken: I’m a physicist by training and I’m currently a professor of medical image analysis at Radboud University Medical Centre in Nijmegen, and I lead a group of about 60 researchers in deep learning image analysis for various applications in radiology, pathology, ophthalmology, and also radiotherapy.
Erik: That’s quite a large team, 60 researches. It’s this kind of unique in the Netherlands or are there other institutions that you know of with this capacity?
Prof. Bram van Ginneken: You have a lot of large groups in medical imaging, but I think this group is the largest with a focus on deep learning image analysis.
Erik: The largest part of image analysis is for radiology, or could you maybe give us a better idea about that?
Prof. Bram van Ginneken: The fastest-growing part is definitely pathology. We started in 2013 with 1 researcher, and that part is now 23 people. That’s growing very rapidly because pathology is going digital at the moment. But the largest part is still radiology.
Erik: Just out of interest, what is the main source of funding for the research that’s being done in this field?
Prof. Bram van Ginneken: We work a lot with companies because we have a very practical approach. We build algorithms that analyze images and we want to integrate these algorithms into the clinical workflow. To have them really used on patients, you have to work with companies to achieve that. But I think that’s about 30%, maybe a bit more, of our total funding, and the rest comes from grants from the hospital.
Erik: Also grants from the European Union?
Prof. Bram van Ginneken: Yes, of course.
Erik: Bram, how do you think AI can create the most value for radiologists?
Prof. Bram van Ginneken: That’s a difficult question.
Erik: [laugh] I know that’s difficult for you, but maybe that’s why I’m asking you.
Prof. Bram van Ginneken: I think that where we are right now, is that we can build algorithms that can analyze medical images with the performance of human experts. That’s really a major breakthrough, because I’m in this field for over 20 years and the first 15 years basically it was very difficult to write computer algorithms that were at the level, or close to the level of a radiologist. Then the value for radiology of those algorithms was quite limited. Now we can make algorithms that really perform very well, so the value for radiologist is that certain tasks can be performed at radiology level. But this is also the big limitation of the system that we have today, namely that they only address a single narrowly defined task. For many of these tasks there’s still no software available. I expect in the next 10 years there will be many more of these algorithms coming on the market, so radiologists can try to integrate them into their workflow. That will save them time and make them more accurate. How to actually do that is also a tremendous challenge, which goes beyond what a scientific research group is dealing with.
Radiologists can try to integrate algorithms into their workflow. That will save them time and make them more accurate. How to actually do that is also a tremendous challenge, which goes beyond what a scientific research group is dealing with.
Erik: How do you define the clinical use cases? You’re developing algorithms, so you have to make selections for the purposes of such algorithms. How do you make that selection?
Prof. Bram van Ginneken: You basically have to look for areas where you see that a lot of work is being done on a very specific narrowly defined task. You talk to radiologist and you ask them what kind of tools they would like, and then you identify certain areas, and we then write a grant or work with a company to build a tool for that. So that can be basically anything. It also really depends on the interest of a particular researcher. It’s often a single radiologist you’re working with who says: “This is what we should work on. I can provide the data and I have actually time to do research.” This is a difficult thing for radiologists usually. So then can you start with the project, and if it’s successful, the project grows.
Erik: You are working in different fields of medical imaging, it’s not only with radiologists. Where can you see the most positive response of, let’s say, the users or the specialists?
Prof. Bram van Ginneken: I think that, if you make software that works, everybody who could use a software is super happy with it. There’s a tremendous demand for software that works. I basically have people coming into my office almost every day saying: “I’m here, I have this data, I want to analyze it. Do you have a software for that, or can you make that?” And I often have to explain that it takes a bit more time than saying, “Okay, give me the data. In two weeks, it will be finished.”
If you have software that works, you have happy users. That includes radiologists, but it definitely also includes referring physicians. We work with pulmonologists and they are super happy with the reports that we produce when we analyze the scans. Some even say: “Oh, then I don’t need the radiology report any more”. So all doctors can be happy users of this software.
Erik: This may be an important question. From which direction do you see most of the demand or the greatest need for this kind of applications? Is it really from radiologists or is it more from other specialties?
Prof. Bram van Ginneken: Yeah, that is the pull or push question, right? I think both. I do see a distinction between software that would handle a task that is currently done by humans and could make it easier and faster and maybe more accurate to do this particular task. There’s also a lot of software that has been developed that claims to do something that we are not doing now. Some new imaging biomarker, new quantification software. I think that will specifically appeal to the referring physician who just wants to know “Can I have a better way to measure response or to predict response to therapy?” If there is some computer program that can do that, they will be very interested.
Erik: Do you see, and this is maybe the most crucial element of my question, other specialists that would like you to develop algorithms that are used to analyze images instead of radiologists?
Prof. Bram van Ginneken: Yes, of course. A big example now is immunotherapy in oncology. This is very expensive medication that happens to work very well in a small subset of the patients that received that medication. I think everybody is asking themselves: “Can we have better ways to predict this immunotherapy, will this very expensive drug work in this particular patient?” If you have tools for that, and you can prove that they work, then an oncologist will be super interested in using that.
I think everybody is asking themselves: “Can we have better ways to predict this immunotherapy, will this very expensive drug work in this particular patient?” If you have tools for that, and you can prove that they work, then an oncologist will be super interested in using that.
Erik: So it would be actually very interesting for radiologist to have a look at how clinicians are using these data or what type of information they wish.
Prof. Bram van Ginneken: Yes. And what you see of course is that radiologists spend more and more time in tumor boards. It’s actually a whole team that makes treatment decisions and makes the follow-up decisions during treatment. The crucial information could come from radiological imaging, but it could also come from genetic analysis or from a combination of these elements. These are very difficult and ambitious projects, so actually it might be a whole team of doctors asking for a particular solution for a patient group.
Erik: So, radiologists should learn how to look at imaging findings in a more holistic approach, and look at the type of information that is desirable for the purpose of such meetings such as the so called “theranostics”, for making decisions in the treatment of the patient?
Prof. Bram van Ginneken: I think that holds equally well for the other members in the tumor boards. Compared to when we started working with pathology, I now see much better that there is a lot of interesting things you can do in the interplay of pathology and radiology. But these are difficult projects to set up. Now I’m talking about developing something new, some new predictive tools, and these are long projects that take like 5 to 10 years, at least. On the other hand, there are also a lot of smaller tools which could replace a particular task, and which could end up as a button in your workstation, which could just speed up your daily workflow a little bit.
It’s important to distinguish these two types of tools. Maybe there are not only two, maybe they are three or four types of tools, because in the ongoing discussion people talk about AI as if it’s one thing and about what it will do to radiology. But that is simplistic, right? Because it’s a technology that affects the work of doctors at many different levels.
In the ongoing discussion people talk about AI as if it’s one thing, and about what it will do to radiology. But that is simplistic, right? Because it’s a technology that affects the work of doctors at many different levels.
Erik: There is a quest for medical data to develop these algorithms. Most of the research is being done in academic institutions. Do you think other radiologists should also be encouraged to engage more actively in for example collecting data and making these data available? If you think so, how should they do it? What would be the best way?
Prof. Bram van Ginneken: I’m not sure if most of the research is done in academic groups. I think companies also do a large amount of research. For example, I said that our pathology group grew from a few people to 23 people right now. Google also has 23 people working on digital pathology. If Google sees potential, they can very quickly scale this up to a much larger number of people. Right now, they stick to this amount of people. But I think that in general it’s a very difficult for academia to compete in AI with companies that have more resources, that have also more computational resources. But you are right, what the companies are lacking is the access to the medical data.
As a radiologist you could contribute data to both academic research and company research, or maybe to a European or National data repository for research, to which both academic groups and companies could have access. Yeah, you said it’s a quest. I think that’s the right word. Everybody’s thinking about how to solve this problem.
Erik: How to setup such a platform?
Prof. Bram van Ginneken: Yes, many companies are trying to do this of course and scientific societies are trying to do this and an individual research groups are trying to do this.
Erik: Exactly.
As a radiologist you could contribute data to both academic research and company research, or maybe to a European or National data repository, to which both academic groups and companies could have access.
Prof. Bram van Ginneken: But I think that starting with challenges really was a positive thing. These competitions for algorithms started 10 years ago. There was public data available but not so much. And if you look today, there are so many public datasets already available. Maybe they not always have a huge size so that you can really make the best AI tool possible, but the situation has much improved and it’s actually so much public data already out there. But it’s difficult for people, for researchers to find the right public data.
Erik: But then the question is also what the quality is of those data? How can you verify this? I think this is a very difficult issue.
Prof. Bram van Ginneken: Yes. They say about data science: “90% of the work is curating the data, 10% is the algorithms”. Maybe that’s an overstatement, but it is definitely a very large part of the work to get a well curated data set, yes.
Erik: Maybe my last question Bram. How do you think that radiologists should engage? What should they do to apply artificial intelligence? Should they just buy the applications that are available, or should they just become more active as well?
Prof. Bram van Ginneken: What I think they should do is team up together, and realize that they are the users of this software, and they should seriously test the algorithms that are available. For that you also need well curated data sets…
Erik: Some validation scheme?
Prof. Bram van Ginneken: …and they should not just buy software from the first company that happens to walk in into their hospital.
Radiologists should team up together, and realize that they are the users of this software, and they should seriously test the algorithms that are available. For that purpose you also need well curated data sets.
Erik: So, they should develop standards for validation of the algorithms?
Prof. Bram van Ginneken: Yes. And that’s an ongoing process, because if you test that now, the year later you have to test it again. Technology evolves. CT scans are different 3 years from now than they were 3 years ago. There should be an ongoing effort to validate these algorithms. Right now, algorithms are validated for example by the FDA. That’s a largely secret process, where the company who developed the product actually is responsible for organizing the evaluation. In the end, you get a stamp from the FDA. But as a user, as a radiologist, you have no idea who this software really is. Instead of testing it with 5 or 10 cases in your own hospital, and doing it at every hospital, this should be a more coordinated effort.
Erik: The ideal solution would be a neutral platform with the public dataset that can be used for validation for example?
Prof. Bram van Ginneken: Yes, but if you want to use it for validation, you shouldn’t make it public. Part of it has to be kept secret as well.
Erik: Yes, of course. Thank you very much for this interesting interview Bram. I wish you good luck with your future research. I hope to learn from it and to hear from it more and more.
Prof. Bram van Ginneken: Thank you very much.
