There are some technologies that enhance human efforts and abilities and other technologies that make such a drastic impact, they revolutionize protocol and entire ways of thinking. AI in the healthcare field is one such technology. Scientists—including BCRF investigators—are working on ways to harness AI to improve how medical professionals interpret mammograms, and finesse and better personalize existing risk prediction models, and tackle disparities in screening and risk assessment.
That’s what Dr. Regina Barzilay is working to uncover. A BCRF investigator since 2022, she is a School of Engineering Distinguished Professor for AI and Health in the Department of Electrical Engineering and Computer Science and a member of the Computer Science and Artificial Intelligence Laboratory at MIT. She has received numerous fellowships, including a MacArthur “genius grant,” was awarded the first Association of the Advancement of Artificial Intelligence Squirrel AI Award for Artificial Intelligence for the Benefit of Humanity, and, in 2022, was elected to the American Academy of Arts and Sciences.
Chris Riback: Dr. Barzilay, thank you for joining me. I appreciate your time.
Dr. Regina Barzilay: My pleasure.
Chris Riback: I just want to make sure that I’ve dialed into the right podcast recording. You’re an MIT School of Engineering Distinguished Professor for AI and Health in the Department of Electrical Engineering and Computer Science. We’re here to discuss breast cancer, correct? I didn’t dial into the wrong Zoom recording, correct?
Dr. Regina Barzilay: No, but you missed my other important title at MIT. I’m also AI lead for Jameel Clinic, which is a Center for Machine Learning and Health at MIT.
Chris Riback: Yes, it is a fascinating mix and intersection of skills and interests. I must confess that of all of your incredible fun facts, MacArthur Fellowship, the first recipient of the million dollar award, from the Association of the Advancement of Artificial Intelligence, I think the one that I found most amazing was that IEEE Intelligent systems named you to their AI “10 to Watch” list in 2006. I mean, most of us had no idea there was such a thing as AI back then and you were top 10. You’ve been on the cutting edge all your life, haven’t you?
Dr. Regina Barzilay: So it’s actually really funny when you’re talking about AI that at those times in 2006 it was not very popular term because people kind of thought about AI as all technologies, it didn’t work, so the list was not maybe the coolest list one can be in. So I was of course very flattered to be on the list, but for general audience, AI didn’t really mean much and it’s really kind of fascinating to see that previously I would say I work in this area and people say, “What exactly do you do?” Right now, I don’t even need to explain because everybody knows. So this was a big change throughout my career.
Chris Riback: ChatGPT is the best thing that’s ever happened to you?
Dr. Regina Barzilay: So I actually don’t use ChatGPT, but it’s really exciting to see how everybody else is using ChatGPT and how people are excited to see and it’s really spectacular to see this amazing progress that happened because when we are looking back at history with the kind of major technological breakthrough you can imagine, people always give the analogy of AI and electricity, but it’s true that if you’re thinking in 19th century when people didn’t have central electricity, when factories were powered by steam engines and other things, so how there was a translation that happened that slowly changed our lives. So it’s really interesting to be able to observe it through kind of my lifetime that from systems that could barely do anything. When I took my first class in processing, I didn’t even see any systems that can do anything productive to observing machine translation. This part was really super amazing to me how machines can really produce such beautiful translations, how the transcription of the speech improved over time. So I still remember those days when it was totally unusable technology. So it’s amazing to see it.
Chris Riback: Yes. And it’s got to be very reassuring to know that the rest of the world has started to come along to where you’ve been all the time where one has been, yes, I guess it must give a sense if one has an interest that has for so long, probably at dinner parties or a cocktail parties, although you probably hang out with much more interesting people than I do, but go and say, “Wait, wait, wait, what do you do? What is that?” And now the rest of us have not caught up to you by any stretch, but we’re at least in the same time zone as you are. So that’s got to be somewhat rewarding. So the connection between AI and breast cancer screening, what is it? And as I understand from reading about you, this work evolved from your own personal experience, didn’t it?
Dr. Regina Barzilay: Yes. So one of the areas where we actually don’t see AI at all when we experience it as a patient is a healthcare. I was just looking at Boston Globe, I live here, so in Boston Globe and there was this article about 13 users of AI by Boston citizens. People were saying different things. Somebody was using ChatGPT to write class descriptions, somebody was using it for planning meals. It was really stunning to see that nobody ever used it in any way related to healthcare. And it’s a very representative. There are lots of different measurements which are much more systematic than the story in Boston Globe that shows that there is truly no AI in healthcare. But if you’re thinking what AI is designed to do, is actually to do predictions and to do prediction about something where we have uncertainty because as humans it’s very hard for us to take data which comes from many different sources, put it together and kind of provide probability distribution over the outcomes.
This is not how our mind works. And also even if you’re very experienced doctor or expert, maybe you’ve seen in your life 50,000 patients. It’s a lot, maybe 100,000 patients, but you can send millions of patients, tens of millions of patients to the machine. And machine has this unique capacity to combine the data and to make these predictions. That’s how it’s optimized. And in terms of breast cancer, the specific area where I started, I started in initial language processing. That’s what I got my tenure and got all the many of their awards when I was younger faculty is work on the processing and where machine learning was always a part of how these tools are developed, ChatGPT-like tools.
In 2014, I was diagnosed with breast cancer. And this stunning thing was that there was really no AI. There is still no AI actually when you go to be treated, but there was none. And I wanted to see when I came out of my own breast cancer, I wanted to do something because I felt it’s really unfair to women that the technology in 2014 or ’15 was still not a year of AI for the crowds. But still this technology is not really is part of cancer care. And among many different ideas that I had, some of them were really stupid. But one of the ideas that I had was really to understand: What is your risk? Because I could never perceive myself being high-risk. Nobody in my family had breast cancer. I was athletic, relatively speaking. I was trying to eat well. I never drank or smoked. So based on the traditional risk factors I was not it. So, I just always brushed it because I don’t have family history or other factors, so why would I be worried?
And the question that I wanted to see is how you can predict based on the first mammogram that a woman takes or any subsequent mammogram, how likely is it she’s going to develop breast cancer? And this is a question in which to some degree clinicians are trying to answer without AI, and this is something which is called density, which shows you how much white on the mammogram you have. And every woman who does a mammogram in this country in the United States would get a letter that if she has dense breast that would tell her that she’s in increased risk of breast cancer and that her cancer may be missed. This is actually a new federal regulation. But if we are looking at this measurement, it’s a very faulty measurement because above 40 percent of women have dense breasts. So clearly not 40 percent of women likely get breast cancer or not even half of that.
So the question that we ask: Can you in a more accurate way to translate the pattern of a woman that we can see that she doesn’t have breast cancer? Can we take whatever is there and predict whether she’s going to develop breast cancer in the future? And that’s what we worked on using the power of machine learning to teach the machine to take the images of patients which look to us and look to the doctor cancer free and kind of assess what will be their trajectory in the next few years in terms of breast cancer.
Chris Riback: So describe for me if you would please, how does it work? What is MIRAI, M-I-R-A-I? I read a wonderful quote from you that with AI and the work that you’re doing and with MIRAI, we’re not just looking for the cancer, we’re looking at the soil that allows the seed of cancer to grow. So take me through the detail. How does it work?
Dr. Regina Barzilay: The person who did MIRAI for his PhD is my student Adam Yala, who is now a professor at Berkeley UCSF. And he came up with this name MIRAI—if I’m not mistaken is future in Japanese. And the idea is alluded to. So people were trying through this density pattern, trying to predict what is there that happens in the tissue before we can actually see cancer? And people try to do it using kind of just analysis, like looking as humans and try to say, “Okay, this is a bunch of women who didn’t have cancer. That’s how their mammogram looked few years before. And this is the women who did get cancer. That’s how their mammogram looked like.” But it’s a very imprecise. So instead what we’ve done, we gave to the machine images of women, which we knew what happened to them in the next five years.
So the machine would kind of see the images of somebody who doesn’t have cancer in the next five years. It would see images of somebody who would get cancer within two years and images of somebody who is going to get cancer in four years or in five years. And by providing this kind of pairs of image and the outcome, machine can correlate particular features of the image with certain label. And it’s not very different from your iPhone that is trained to identify your face when you want to unlock it. So in this case, instead of training it on your face, you train it just on the image and the outcome. And there are many kind of techniques that you can do to do the better job. But this is a basic idea. You just apply a version of computer vision algorithm that is trained on the patient outcomes and then once it’s trained, once it’s seen hundreds of thousands of examples like this, they can actually go take a new image that hasn’t seen before and give you the likelihood of developing breast cancer in the next few years.
Chris Riback: I know that what you were just describing was the retrospective part of your research and proving out the viability of MIRAI. I’ll ask you in a moment about the prospective trial that you are doing, but given what you just described, would it have caught your breast cancer?
Dr. Regina Barzilay: Yes, so it was very interesting. So after I started working on it, I had my mammograms. It took three mammograms before I was diagnosed and we can actually look back and in my case it’s actually even visible. It’s very tiny, it’s like ambiguous. But if you see the sequence, you see there’s a particular spot that increases over time and something like MIRAI would flag me as a high-risk patient. So the question is what do you do if you are high risk? So you can suggest to this patient to do MRI or maybe do ultrasound or do something. It’s not going to tell you, it can say that it’s already sitting there, but it just kind of flag that says you need to go and do follow up. And I just want to say one of the things that we did, and it was interesting because it was done just before the pandemic.
So very early in the pandemic, one thing that we did for MIRAI was validated across many different populations. We went in the U.S. to Novant [Health] in the Carolinas. We went to Emory University Hospital, we went to Taiwan at the beginning of pandemic to day one I think in January or February in Sweden, in Israel just to ensure that this model actually delivers good results in different population. And the reason you have to do it because human cannot validate it. When human is looking at it, they don’t know whether the patient has or has not. So you need to ensure that the model works well across different groups.
Chris Riback: And I was going to ask you about that because as the rest of us are becoming more educated around AI, obviously one of the things that we all hear about is the risk of bias in various AI executions. It’s no surprise given what you do and the awards you’ve won and the years that you’ve spent thinking and worrying about AI that you factored in or worked in the individual differences that can occur based on culture or race or historical access to healthcare or any other range of social issues. That obviously has to be table stakes in terms of creating an algorithm or a program like what you created.
Dr. Regina Barzilay: I think that for breast cancer when you’re developing AI for this patient certification software, it’s really important to ensure that it works across different groups of patients. For instance, we know that African-American patients develop breast cancer much earlier. It’s more aggressive cancer. One of the challenges––and they talked to many patients, many patients who some of them by chance discovered it or they already had tumor which was growing, and it was before 40 before they even start screening. So first of all, we want to ensure that it works for these patients who really need help. But second of all, we can imagine that looking forward that we can do the first mammogram not at age 40, maybe we can do it at age 30 and then say these patients really look kind of safe and secure and they can come back at 40, but this patient really do not look safe and secure and they maybe should be coming in more frequently.
Because right now the program is that for all of us, unless you have BRCA, which is only 15 percent of breast cancer patients who have this condition. For the rest of us, the recommendation is you come every year since starting at age 40 or there are some people who argue about different frequency in Europe every two years, but we need to have a screening regime which is personalized to your individual risk rather than just look at your age and your genetic status and so on. So I think it’s key for these tools to be really equitable because this is the only way they can make a difference at the population level.
Chris Riback: What’s been the reaction within the healthcare community, within the breast cancer community?
Dr. Regina Barzilay: So it’s very interesting. So when we started working, it took us a while because my papers from breast cancer were the first papers I wrote kind of venturing into medical domain; before I wrote all my papers in computer science venues. It took us a while, but then we start publishing and Adam Yala recently attended a breast density conference in Hawaii, and he related to me an interesting conversation. He said a big portion of papers there were actually trying to check how MIRAI works in different population in different settings. So it’s good to see that people are trying it and independently validating it. But what he said, which was kind of interesting, that when we first published our paper, it was I think at a 2019 radiology conference said that they didn’t believe our results because it looks so good.
Chris Riback: Wow.
Dr. Regina Barzilay: And now that they validated it and there are more and more results coming it’s changing, but it’s not easy because now we are after, again, paper was first in 2018 to today we’re in 2023. We are in a better shape in terms of people accepting it, people talk about it. It’s a known thing, but still it’s not part of the routine clinical care and it’s not unique to only breast cancer. We see very few AI tools in our routine. This is an interesting question, which is maybe not necessarily a pure scientific question, but more implementational clinical question. How do you take it, some things that work in many research studies, how do you really translate and make it part of patient care? That’s the question that we are all thinking about now.
Chris Riback: And to your point, if I understood this part of my research correctly, you’re also applying your work now to lung cancer risk.
Dr. Regina Barzilay: Yes, absolutely. So we actually did the paper and it actually works even better on lung cancer because you use CT scans. CT scans, it’s a higher-resolution modality than mammograms. And it works reasonably well. And when it predict that the patient has cancer, it also has a chance it’ll tell you on which lung it’ll be located. So what it tells combined, thes two pieces of evidence that we’ve collected is that maybe the malignancy should grow to sufficient size and have some significant impact on the tissue so that human eye can distinguish it. So, when we are thinking about risk, when we’re talking about BRCA gene, a girl may not even really have breasts, she can be very young, but she still has a higher risk of developing breast cancer. Talking here really about risk. Here we may be talking about the patients where the cancer already grows in their breast or in their lung. It’s just machine can predict it much earlier or identify that the patient goes this way before human eye can discriminate it.
Chris Riback: Tell me about you. You grew up in a couple of different countries I believe, or at least spent a few of your initial years in Moldova before leaving. Tell me about you, your journey and was it always science for you? Was there another passion or always computer science and always technology? And I know you’ve got electrical engineering and probably 75 other highly intricate interests and degrees in your background. But was it always that side of your brain or was there ever a possibility that you were going to go and do poetry or creative writing instead?
Dr. Regina Barzilay: I wouldn’t go that far. So I did my undergraduate, I always was interested in math. So I did my undergraduate in math and then I was a teacher in high school in Israel. I worked while I was still doing my undergrad, I was working in high school as a teacher and then I went to study masters in computer science and then eventually I get there. So I’ve never thought of first of all doing creative writing or being a poet and also about the medicine. It was kind of clear to me that it’s not my cup of tea because I’m afraid of blood. But it’s really surprising to see how much now I go to it through this line of research, but luckily I can just do my computer science without observing through patient care.
Chris Riback: Did growing up in multiple cultures, and I guess maybe you don’t view it as multiple cultures. I believe you spent your first five years in Moldova before then growing up in Israel and going to university in Israel. I think then before coming to the US for graduate school. Is there an aspect of that part of your life that you think has affected your perspective and applies itself to the work you do today? Or do you think irrespective of that background you’d be doing what you’re doing today?
Dr. Regina Barzilay: I think that one kind of lesson that I see moving through different cultures. So I was 20 actually when I moved to Israel. Moving through different cultures and kind of observing personally how the Soviet Union fell apart and when they became independent and we moved to Israel and then I came to the States. There are lots of things that we take for granted in our daily life just because it’s something you don’t even question. When you are moving between different cultures, actually some things that in one society are norm. In other society, they’re totally not normal. In some ways, it gives you more courage to try things that others didn’t try. When I started working in this field, when I was literally going to people, who I barely knew at [Massachusetts General Hospital] and saying, “I would like to walk with you, I’m a computer scientist and I’m a breast cancer survivor, can you please work with me?”
And there were so many rejections and people say you don’t know what are you doing and go back to your NLP [natural language processing]. And I think that this robustness of you kind of go to different places and you just stand up and try another way that maybe a bit helped me to go through this challenging period—and it was quite challenging. I just want to say how much I really appreciate support of BCRF because for us to be able to move fast and to really bring it to patients, because this is my key motivation for doing this research. It should really change the outcome for patients. BCRF funding enables us to do these prospective studies where we actually are applying it in clinic, seeing how it changes the outcome and observing the results of prospective studies would be a motivation for daily changing the clinical standards, assuming that the result’s as good as we hope for. So this was another piece of success that having an organization that can take your idea however crazy the idea may sound and provide you this support was really essential for us.
Chris Riback: Yes, that is among the many things, but it is a key aspect of the work that BCRF does. To close out Dr. Barzilay, of course, we are all extremely grateful for the historical work that you have done. We are all equally impatient about the future. So what’s next?
Dr. Regina Barzilay: Oh, thank you for asking question. So while we are working, and Adam and I are working, Adam actually more leading this work now in the translation, there is a question that again as a patient I am really, really curious about and that what motivates my new project. So as you know, many breast cancer survivors are taking tamoxifen, which is a drug, and now the recommendation is 10 years of tamoxifen. And depending on the patient, they experience various side effects. Some of them are immediately apparent, some of them are less apparent. There are again clinical studies that shows that if you take it instead of five years, by 10 years, it decreases your chance of recurrence. There are several questions. Is it really benefiting everybody? Because when I actually was starting my treatment, the standard was five years, now it is 10 years.
And the answer is we don’t really have a good answer to this question. We don’t know. Can the patient stop? Maybe somebody’s ready for 15 years, maybe somebody is now five years. There are some different tests like the Breast Cancer Index [developed by BCRF investigator Dr. Dennis Sgroi] and others that try to answer these questions, but there’s still not a standard of care. So, the questions that I’m trying to answer, and we’re just now in very early stages: Can you detect based on your tissue if tamoxifen is still benefiting you? Do you need to take it? And the bigger question is, and I’ve recently discovered by doing a lot of reading that actually tamoxifen penetrates blood brain barrier and it has a long range of effects on your brain. It’s well-documented in scientific literature. And again, the question is can we create another version of the drugs that maybe don’t do it or then they can what does it do to based on your individual makeup because another part of my research today is primarily drug discovery.
So can we create another drug which will be as effective, but at the same time maybe with less undesired side effects? So that’s what I’m thinking right now.
Chris Riback: You are genetically incapable of slowing down is the lesson I’m taking from you.
Dr. Regina Barzilay: Hopefully, hopefully.
Chris Riback: Hopefully, for sure. Well, I look forward to, and I’ll put in my request now, for the opportunity to get to talk with you again as you advance on that new effort. We thank you for your historical effort, your continuing efforts on MIRAI, and thank you for taking the time with me today.
Dr. Regina Barzilay: Thank you. Thank you very much.
Please remember BCRF in your will planning. Learn More
