Aleksandra: [00:01:21] Welcome Jeroen. How are you today?
Jeroen: I’m good. It’s evening in the Netherlands, so I had a long day, but I’m good. I’m looking forward to it.
Aleksandra: Fantastic.
And now I usually let my guests introduce themself, but I have to say a couple of things about you. I am super excited for you to be my guest because this year he earned a spot on the pathologist power list. And for those of you who don’t know what that is. There is a magazine “the Pathologists” obviously for pathologists or for the pathology community, and every year they publish the power list.
So those pathologists and more and more people on this list are not really pathologists but are involved in the computational digital pathology space. Jeroen is one of them, and I’m super proud of having one of the power people in the pathology space with me. And now I let you introduce yourself Jeroen.
Jeroen: Okay, thanks Aleks. My name is Jeroen van der Laak, I’m currently a professor of computational [00:02:21] pathology at the Radboud University Medical Center at Nijmegen. I have a background as a computer scientist, but that’s really a long time ago. I studied computer science. I think I finished like 33 years ago, and ever since I have been working at the Harvard University Medical Center, and actually all this time in the pathology department.
And…
Aleksandra: How many years?
Jeroen: Yeah, it must be I don’t keep count anymore at some point but I think I started it in ’91. So it’s 31 years now.
Aleksandra: Okay.
Jeroen: Very early in the field of, we didn’t call that computational pathology back then. But I, I started doing image analysis in pathology on pathology images very early way before we had AI as we see it now.
We had host slide imaging that didn’t exist yet. And I worked in the field for a long time, and of course we all know the last say 10 years, maybe 15 years, this has really. So the field used to be very kind of research oriented. So we could build all kinds of nice image analysis applications and [00:03:21] then, the user would have a microscope with, on top of that, a camera, and they could grab a single image, and then we would perform some image analysis of that image.
And that would be used quite a bit for research projects. Did a lot of projects on angiogenesis. Studying micro vessels in all kinds of animal models to see the, to study the function of of vessels in tumors. We did a lot of stuff on ploidy analysis, DNA ploidy analysis, which I think completely disappeared from the world.
That used to be a big thing. But none of that was really used in practice. Not in, diagnostics. So ploidy analysis was used a bit. We had some stuff in psychology that was used. But only the last, Yeah, I think five to 10 years we really are approaching the level of actually. We are reaching and even surpassing the level of pathologists for certain tasks, which means it’s becoming much more exciting now.
We used to be a niche area, where our computer nerds would be in pathology labs and be hidden away in the basement and writing software codes [00:04:21] that pathologists didn’t know much.
Aleksandra: They still don’t know too much about the…
Jeroen: But now, we have keynote speakers at pathology conferences that just talk about ai, which is, was unthinkable 15 years ago.
Aleksandra: And actually I take it back because you actually have some pathologists who have dual background, who have computer science background. Few of us they are like the real bridge. But yeah, I don’t code.
Jeroen: Yeah, absolutely. Sure. But it used to be a rare species. And, I think it’s great that this has improved and now yeah, I think there are many opportunities for us to really use the AI and help pathologists in doing their work and in the end, maybe treat our patients better because we have better diagnostics.
So for me, it’s great to see this happen. And I’ve seen a lot of changes in our field over the 30 years I’ve been here of and it’s great that we reach this point, I think.
Aleksandra: So you say like around five years ago, how long ago did the focus switch from research to actually practical [00:05:21] applications? And how do you approach differently, like in, in your workflow, in your way of working together? What are the differences?
Jeroen: I think there are two as described a lot. There are of course two major breakthroughs. One is whole slide imaging if you don’t have digital images, then well your AI is not going to do a lot for you.
And of course that starting to happen, I guess even something like 20 years ago, that we start to motorize microscopes and Grab slides, field by field, and then after three hours of scanning, you would’ve one image.
Aleksandra: Did you have something like that in your lab?
Jeroen: We did, yes, actually we we had very early prototypes, but it would never scaled. It would be nice. You would’ve very cool images and of course your hardware was not capable of dealing with all the data. So you would have nice images, but you couldn’t really use that. I think that’s one great advancement that we now have whole side images and these beautiful scanners that digitize a slide in two minutes, one minute, I don’t know.
And you have very high quality [00:06:21] images. Still digital pathology is not completely used all over. There are city main labs that don’t. And I think that’s one of the things that has to happen in the coming years that we go to digital pathology in, in every lab. Because that’s, I think that’s, that will be mandatory, of course, to use the AI.
And the other thing that happened, of course, is that we started to, at some point, deep learning came along. People start to build these deep neural networks. And that of course has given us the, to compute possibilities and compute power.
Cause you can run these very efficiently on, on GPUs. Yeah, that combined with whole site imaging is the great combination that now makes it possible to analyze whole site images in couple of minutes and do great things with them, with really very strong analysis capabilities.
Aleksandra: The second question was, how did the way of working with your collaborators change from like the research application when I assume you’ve been giving images and you were doing something like that[00:07:21] to now that you actually have to incorporate the clinical information, or the clinical context of a patient, not just some phenomenon that you’re researching.
Did you notice a different way? Was there a shift? What is now more important than then? To have it run smoothly?
Jeroen: I think so I used, and I think a lot of people in, let’s say that, that did the same kind of work, used to work a lot with really very research oriented people, so not directly the diagnostic we would really do very specific studies, on 20 images or 30 images or 40 images.
So it was really very research oriented. And the publication was the output, right? We would write publications. Now we tend to work much more with clinicians directly and do reader studies and say to a clinician, Okay, you do breast cancer grading, Cool. Do you want to do a study? You grade 100 cases for us and then we wait three months and then you grade the same cases, but now you get an AI help.
And now we see what happens. Do you [00:08:21] become more accurate? Do you become faster? How do you feel about it? So the collaborations in essence to studies are completely different. We are now really testing stuff in diagnostic practice which is a whole different application. It also means for a lot of the collaborations, So we used to develop, if you go back to the publications of 20 years ago, people would have great publications and they would have 25 tumors in their publication.
If you now submit a paper and you say, I built AI on 25 cases…
Aleksandra: Don’t think that I got published, will you?
Jeroen: They don’t even look at it. So now you have, you need to have two and a half thousand cases or something like that to really make people feel is relevant, what you’re doing. So the work that we are doing is also shifted a lot to, a data collection and we did a podcast together on Big Picture.
I think it’s a great example.
Aleksandra: That’s a super huge initiative of collecting data to do this at scale.
Jeroen: And that’s what we see, right? We are focusing more, and using more and more very large data sets to make algorithms that are very robust and that can really deal [00:09:21] with all kinds of heterogeneities.
So there, there is a very strong shift on establishing collaborations with people that can provide eye quality data. And of course, best would be to not only out the images, but also have some molecular data to it and have some clinical data to it, and treatment data and outcome data. So the hunt for data, if you wish, is much more important than it used to be.
And that really has shifted, at least for me. The work that I do has shifted completely.
Aleksandra: So , you’re after 31 years, you’re still in the field and I assume you’re staying right or are you thinking of switching careers now?
Jeroen: No, I think I like it too much. I always liked that you I never liked, maybe I shouldn’t say this, the people working in radiology, I like the people, but I never understood why they like the images because, And nowadays it’s much better, but when, back in the days, they would look at mammograms, which were very fuzzy, black and white images, and you had [00:10:21] to find some spot, a vague spot. And that could be cancer, it could not be cancer. Whereas if you go to pathology, it’s beautiful. You see everything, you see every detail. Early colon. I think I fell in love with pathology images long time ago.
And the, I always, I study computer science. I like to write software and the combination of software and AI models today that can deal with these images is still, it still fascinates me. I think it’s still, it’s a beautiful combination of techniques and yeah, I think I’ll stay.
Aleksandra: Okay. Good. That brings me to my next question. Just a apart even from that, we’re talking about digital pathology. To me, 30 years in one field is a huge measure of success because 30 years in any industry is a lot of shift, a lot of pivoting, a lot of I don’t know, even inter-institutional changes that you survived. You thrived. You are a renowned scientist. So this is [00:11:21] definitely a big success in the field.
What is success for you in the field as such? Not necessarily a scientist, but to successfully contribute to the field of digital pathology at whichever level you happen to contribute.
Jeroen: I think it’s a combination of a lot of things. So I think of course, it’s always about working with the right people, right? So establishing a good team having good people around you, you can’t do all of it yourself. Actually, if you look at it, you can do very little yourself. You need a lot of people to help you with a lot of things.
So I think it’s one of the very important things is to have the proper people around. And I think that’s one of the things we really pay a lot of attention to. And I think one of the great things to see then is that people really yeah, they develop themselves within we have a lot of PhD students and it’s great to see that people learn so much in four years of PhD.
They develop themselves in all kinds of different fields, right? It’s not just science, but there is so much more. So I think for me it’s, that’s always [00:12:21] been one of the great stimulants for my work really to see, to work with a lot of people and see so many talented people, find ways to, to use their talent, to build cool things, do great research, but also, build these days, build solutions that in the end will help people, right? We hope that we can do better diagnostics and have better cure for people or maybe don’t overtreat so many people. So I think that’s, for me, that’s a great value that we add. I personally also feel that I’ve been in the pathology field for a long time and I think the previous blog post you joked, How could you stand being with pathologists for 30 years?
For me, that’s of course, that’s a huge value now because there’s a lot of AI groups, a lot of AI researchers. But to be successful in pathology as an AI researcher, you have to understand pathology very well. And you have to understand not just pathology, but also what is the place of pathology and the whole healthcare system, right?
What is the role of pathology for an individual patient? What, and that brings you to the[00:13:21] questions that you have to answer with AI. You, can build AI for a lot of things, but many of those are not really helpful for patients in the end. So to, it’s really important to understand pathologists and stand pathology, and I see as one, one of my that’s not missions.
That sounds very heavy, but I like to.
Aleksandra: Everybody has a mission. Nowaday my blog even has a mission, which is bridging the gap between computer pathologists and computer scientists. So.
Jeroen: That is yours.
Aleksandra: Yes. That’s my mission.
Jeroen: Okay. You’re going, doing a good job I think. So, actually it’s a bit, I guess my, our missions align well in that sense because I like to talk to pathologists and to go to pathology conferences and give talks and try to take away the mystic, the magic around AI and just make it very tangible and just show people, Okay, guys, this is what it is.
And see this as a tool that you can use to make your life. Your job better, that helps you. So don’t be scared of [00:14:21] it. Don’t see it as a challenge or threat, but see it as a tool. And I try to do that by going to a lot of conferences and really bringing the AI back to very simple terms like I hope the people in today in the different sessions will do.
Aleksandra: Yes.
Jeroen: Cause that’s just what it is, right? It’s just a bit of mathematical equations. And of course it’s great. It’s very powerful. It’s beautiful, but in the end it’s not magic. It’s computers and bits and bites. And that is what I hope to convey to people that look at it, the pathologist especially, look at it and see what it brings you and help us shape it in the best way that it will help you as a pathologist rather than reject it and say this is not going to replace me. No, it’s not. We don’t want to replace the pathologist.
Aleksandra: So you say you to contribute you need to understand the pathology workflow, the place of pathology in healthcare. What about pathologists? What do they need to understand? If I was starting in the field, what would you need to teach me [00:15:21] or what kind of resource would you need to give me so that I could join your group and contribute well and be on the same page without knowing how to code without being an AI expert myself.
Jeroen: I think we are in, in, in my research group there are several pathology residents that do that.
Yeah. And so these these are residents and actually at this point, there are three residents as a PhD student in my group and one as a post doc.
And they all do very useful work in, not, in, not in coding. They don’t code or very little, but of course there is much more to it, I think about the applications. They collect cases, they revise cases, they use the AI, they set up the reader studies, they talk to the pathologists. There is a lot of value also for pathologists in developing these techniques.
And I think as a pathologist, it’s good if you at least have a feel for how we develop these AI algorithms, right? What do we do? How do we train them? And that gives you also a feel for what can you expect from the output of algorithm, right? What, [00:16:21] when we at work, when can it fail? How can we improve it if it fails?
So just to gather. Feel the tuition for what is the thing doing? What is the black box actually doing? How do you feel it? Because that also gives you a better understanding of how you interpret what comes out of it and how you can help to improve it and also see the limitations maybe.
Aleksandra: That’s an important thing to say, to see the limitation and to make people not discard a good tool because it’s a limited tool. Like any tool is a limited tool. Any tool has a certain application. There’s no one tool that you can use for everything. Ah, maybe smartphone, it’s pretty close to that, but, like tools have an intended use.
And if you’re expecting something out of the scope, out of the intended use from the tool that you’re being given and discarding the tool for that, then that’s a misunderstanding. You’re missing out on a power that you could harness to provide [00:17:21] better patient care. So I think this is super important as well.
Jeroen: Yeah.
Aleksandra: So how big is your group now?
Jeroen: The group is, I think I didn’t count it but I think it’s 30 people, maybe a bit more. Yeah, so it’s a large group
Aleksandra: So you have 30 people since, like, when did you started, how many did you have and how many, like when did grow?
Jeroen: The growth is really recent, so I think yeah, I could look it up. Probably seven years ago I had five people, something like that. And I should say the group is, so, the group is led by three people.
Francesco Ciompi is in the group, Geert Litjens is in the Group, and I am. And the three of us are leading the group.
So it also means that we apply for grants. We work with companies. So all of us have to have our own PGD students and also try to get funding because you can only have a group if you have funding. And in our case, most of the funding is, yeah, it’s just you have to competi, you have to write grants and competitive schemes and try to get funding from the Dutch government or the Dutch Cancer [00:18:21] Society or the European Union or whatever.
So we can only have this large group because we have three people that really apply for funding and have their own PhDs and their own research lines. So basically it’s say three research lines combined in one group. And otherwise it would not be possible to have that many people.
Of course having three lines is great because there was a lot of synergy between the lines. We share a lot of, the code we use, the techniques, we use the infrastructure so yeah, that really helps a lot to have a good solid group.
Aleksandra: So how do you choose successful candidates? Now we’re talking more about like personality traits. What, like, how are you can you be successful as somebody entering the field? Do they need to have some. Do you have to have some background knowledge or it doesn’t really matter. It’s more like how you acquire knowledge. What are your criteria for if I was applying next semester to your group, what would you be looking for?
Jeroen: That’s a good question. I think one important thing is people [00:19:21] should fit in with the team. So they should have yeah, we, we should like them and we should feel that they fit with the rest of the group. Have a good team spirit. We really liked people to collaborate, work together, so they should really be open to working with other people. I think that’s really important. And I think it’s, I like.
Aleksandra: I think this is a super important characteristic for the field of digital pathology because it’s such a multidisciplinary field.
Jeroen: Yes. It’s, yeah.
Aleksandra: That you just have to be openminded. This is like, you don’t have to be in any, like in every research field because many are very narrow and you can just really dive deep into your expertise here you have to have spokes to all the things that you’re touching.
Jeroen: I think it’s really important. We, once had a candidate and I always remember that, that I think he was a very brilliant AI researcher. And he applied to us and we talked to him about pathology, medical images, and somebody said I don’t really care about the images if you give me the images to data and I can do my AI [00:20:21] and I don’t care what’s in the images. I am, I just want to build AI.
First that’s a no-go, right? We, are a part of a hospital. We are part of a pathology department. So people should really realize what they’re doing. It’s not just images, right? We are, we have to, to be able to tell our colleagues, pathologists, why we are in that department. It’s not an AI department. It is a pathology department. So I think people should really be open to that as well. They should feel that what they’re doing, they should understand what the images mean, that there is a story behind every image, right? If it’s cancer.
Aleksandra: Do not say that in the interview when you interviewer with your Jeroen. Touch it. Don’t care about the patients and what’s on the images.
Jeroen: No, it’s, it, don’t do it if you interview with me, that’s, it’s not as, as good thing to say.
Aleksandra: Basically, don’t enter the field if you don’t care. There are plenty of easier images to analyze.
Jeroen: I think at least for us, it’s really important that you realize what you’re doing and what the, yeah. What value you’re adding. And I think, I really like people to be very open [00:21:21] minded also in the sense that they are open to learn. The PGD, especially if it’s PGD candidates. It’s also a learning trajectory, so you should be, yeah, you should be open to learn new things and experience new things and be open minded. And I guess next to that, of course yeah, you should have people that are good in AI. If it’s about an AI position, yeah, they should be good in, in their field as well. But I think that the other characteristics are very important as well. Yeah.
Aleksandra: Tell me in couple of sentences, where do you see us going in the next 10 years as a digital pathology community, as digital pathology, as an area of research digital / computational pathology. Where are we going? What’s gonna be the next break?
Jeroen: The next breakthrough is a tough one. Of course.
Predicting your futurists is not easy. I think, yeah, I guess it’ll become much more mainstay, right? Like the molecular pathology has really become a part of pathology labs and you can’t think of a pathology lab without it. I think AI will [00:22:21] find the same place where also pathologists are much more involve.
We have, quality management systems in our pathology labs for AI tools. And I think people will really start using AI for specific applications in the next whatever, five years or so. So I think that will be the, yeah, it will just become much more regular and probably we will also at some point hit the point where we say this hype cycle thing. At some point we will also say, Okay, so that’s AI and that’s what it can do. Okay, good.
Aleksandra: Can use it. It’s ok.
Jeroen: It’s not what they promise us. They promise us it would do, it would solve everything. And now this is only what it can do. Probably that’s also going to be happening at some point, but I think it will really help us in a lot of ways. I think it will become, it’ll make us more efficient for sure but also I really think that if we do it well and we do the proper studies,
It should help us become more accurate in certain diagnostics and there with also have better predictors. If you want to plan your [00:23:21] treatment, you need to have very strong data, a lot of data and you have to have very accurate data.
I think a lot of things we do now in grading for instance, is rough, right? You have a great 1, 2, 3. I think…
Aleksandra: Yeah, that’s any semi-quantitative of grade, it’s like that, so if we can get more granular…
Jeroen: We should be able to get more information out of the images and make that together with molecular pathology, maybe combine it with radiology to really better stratify patients. But I think a lot of study has still to be performed before we can do that. But I’m confident that we are in a good way and yeah, these studies will come and we will get up.
Aleksandra: Thank you so much for giving us the overview. For talking to us about the beginning.
Okay, Thanks.
Jeroen: Thank you.
