How AI Is Changing Cancer Detection and Treatment (Full Transcript)

[00:00:03] Speaker 1: I mean, we should be asking Regina, she may have even studied this. Have you studied this, Regina?

[00:00:09] Speaker 2: I did have cuts, but it was before the AI time. But I will tell you what, actually, now we're trying to understand that there is some content, as you know, we've seen from the letter, when the cuts are excited and some which is not. And you can imagine that if we want to fully automate the generation of movies, we can just have an AI system that, you know, in maths, trains on the cuts and their reactions, and then creates a customized content, which would really make the cut captivated for hours.

[00:00:45] Speaker 1: It may not be good for the cut, you know, health. I can see why you're on the Time Magazine 100 list. Look at that. Welcome to AI Decoded. This week, we are focusing on one of the most urgent, tangible uses of artificial intelligence, healthcare. We're about to speak to a scientist who is already using AI to detect and treat cancer. She is Dr. Regina Basile, an MIT professor who herself was diagnosed with breast cancer in 2014 and has used that experience to target her research towards prevention. The AI model she and a team built, named Mirai, is now able to detect a patient's risk of developing breast cancer within five years. It is quite an extraordinary breakthrough. To date, it's been used on over 2 million mammograms in 48 hospitals and across 22 countries. In fact, she was recently recognized by Time Magazine as one of the top 100 most influential people in artificial intelligence. Dr. Basile, welcome to the program.

[00:01:57] Speaker 2: Hey, I'm very happy to be here.

[00:01:59] Speaker 1: Also with us, of course, our regular co-host, Dr. Stephanie Hare and the BBC's AI correspondent, Marc Szyslak. Dr. Basile, you've said that your cancer diagnosis changed the direction of your research. Take us back to that moment. Why did you suddenly come to the realization that AI could play a much bigger role in preventing cancers like yours?

[00:02:20] Speaker 2: So, first of all, I need to go back to 2014. In 2014, AI was still not a word that, you know, everybody knew and everybody used. It was something for LA audience, which was more kind of a scientific abstract notion, not a concrete technology that we're using every day. And when I came as a patient to Massachusetts General Hospital, it is located just across the bridge from MIT. It's like 20 minutes walk. You know, I was really surprised as a patient what a huge difference there is between what we do at MIT and what I knew about technology in all other areas of life. And what I observed in the hospital, which is considered one of the top hospitals in the United States. None of the information technologies that is routinely used everywhere was not used in patient care. And this was particularly surprising in the case of cancer care because there are so many questions as to, you know, what will happen. And whatever question you ask your treating team, the answer would be, you know, there was this clinical trial that happened 10 years ago. And this is a list of outcomes. And you say, but what's going to happen to me, which is a normal prediction problem. And the answer would be, you know, go read the study and decide. So there is a lot of uncertainty in all the stages of diagnosis and treatment. And it was clear to me that AI can really help here.

[00:03:53] Speaker 1: So then you come to this AI model called Mirai, if I'm right. Is that the Japanese word for future? Was that the intention?

[00:04:02] Speaker 2: Yes, indeed. My student, Daniela, who is currently a professor at Berkeley and UCSF, come up with this name. He really thought of it as a future diagnostics.

[00:04:14] Speaker 1: OK, well, it's very futuristic in what it does. You say that it can estimate breast cancer risk years before the symptoms appear. What signals is the model seeing that clinicians traditionally can't or don't see?

[00:04:30] Speaker 2: So let me first, before we even talk about AI, bring you to a point that majority of radiologists and oncologists agree. There are a lot of studies that confirm it. Is it actually when a patient is diagnosed with cancer? Often when we look back at the previous image, the signs of cancer, maybe the cancer itself was already there. It was just ambiguous. It was too hard to diagnose. And it's not necessarily we can think of it as a mistake, because if you ever seen a mammogram or any other medical image, there are a lot of white areas. So like I am in doubt, let's go and pull, because it means we need to do a biopsy, which is a surgery, which have side effects, which cost, which takes a patient out of their ordinary life. So at the end, you know, clinicians can only tell you should go for biopsy when they're really sure that there is a very high probability. Because it's very hard for human eye to discriminate these very subtle changes. So before it becomes big enough that we can really see it very clearly, nobody will go and send a patient to biopsy. And what machine can do, machine can identify very, very subtle changes in color, in texture and other things. So in many ways, it's much better equipped than a human eye to see these very subtle changes. And if you train it correctly, it will be able to say ahead of time that the changes are already underway.

[00:06:01] Speaker 3: Regina, if I may come in for just a moment, I'm really curious about this. So, first of all, we spoke a couple of months ago, you and I, and it was a big question about the way that health care differs around the world. So if prevention and early detection is key, do we need to, at the very first thing, before we even throw AI at the problem, get some consensus around what age is the best age for women to be having mammograms? When should we start that?

[00:06:30] Speaker 2: So I actually think that the question when we should start and the questions what do we, you know, how we go after that, they're very much intertwined. Because imagine like in, and there's a huge variance in terms of how different country approach screening, because it's typically a national policy as we know in the UK. It's every three years for, you know, general population. You know, in the United States, it's every year according to one professional society and every two years according to another, but insurance have to pay you every year. In Israel, it's every two years. So there is a huge discrepancy in what's going on. And there is another, and the big question is when do we start? So one scenario that we can imagine is scenarios that we screen everybody, let's say at age 40, which is an expense indeed, because we're going to be screening the whole population, but we're going to say, you know, for 97% of women, they can come back at age 50, or whatever, they will determine the next screening piece, and another remaining 3%, which are deemed high risk, they actually are going to be screened much earlier. Today, it doesn't happen. I'm really happy that you asked me this question, because I just recently looked at the numbers in Israel. And in Israel, the recommended screening is at age 50. And it turns out that 25% of breast cancers happen to women before age 50. And these women, because they're not screened, actually their detection stage is much higher than those counterparts who are screened, which is really a terrible thing, but at the same time, it will be a very big expense. And also, you know, a lot of unnecessary burden to young women to go and screen themselves, you know, every...

[00:08:30] Speaker 1: Just to interject there, is this the point then where the technology really makes a giant step forward? Because if there is a cost imperative to this, and you can analyze mammograms that much quicker and that much more efficiently, and cheaper, presumably, then you could start the process that much earlier, surely?

[00:08:51] Speaker 2: The point is not to start it earlier. The point is just to say that vast majority of population doesn't need to start very early screening. They can come back at 50. But there will be very few that do need to have something special, and they need to be screened much earlier.

[00:09:07] Speaker 4: Dr. Barzilay, can I just jump in there? There's another important strand to this work. I think it was highlighted by the National Academy of Medicine last year. You developed AI models that predict how effective flu vaccines are likely to be forecasting, what the spread is going to be into the next season. Now, this is crucially important because we see every single winter our health services come under enormous strain as a result of these sorts of illnesses, which limits their ability to provide services. Do you think it's important, it's crucially important to see these kind of innovations?

[00:09:49] Speaker 2: Thank you for bringing up the flu vaccine. It's very interesting because this year, if you are following the news, actually the selection was really not a great selection, and literally every day we're reading about hospitalization burden and deaths, especially in children. Actually, the problem that we solved in our paper is slightly different, because at a time when WHO decides which strain to use, they look at a set of circulating strains and then say, this is their current frequency, which one should I select to be the main one for which I'm going to have a vaccine? So we can tell them how the frequency of the strain is going to change in six months, because maybe who is leading the race today is not going to be the one who is there in six months. So we kind of help them to tell how this competition between strains is going to evolve.

[00:10:45] Speaker 1: What is the AI actually doing? What is the AI doing in detecting that?

[00:10:49] Speaker 2: The AI in this particular case is doing two things. It kind of looks at frequency of the strains, like frequency of how much of it, and it very much relates of how strong the strain is, how likely it's going to continue to grow. And you can correlate, you know, the sequence characteristics, kind of the characteristics on the biological sense of the sequence and its ability to be strong, to kind of change over time. So we model this progression as a function, kind of a protein properties. It actually looks at the proteins and make predictions. But I would tell you the problem that we didn't solve, and we could not have addressed the problem that currently we are all suffering, which is total mismatch, because what happened at the time when WHO was looking at a set of strain, the current strain that actually is now dominating the market was not even in the running. So this is another question. Can you actually predict some things at a time when they're making decision is not even there? So this situation happened. But many times, you know, all the strains were there and they just bet on the wrong course, if I may say so. And, you know, again, it's not like AI is perfect, but it can assist you in this betting process by providing another source of information, complementary to many other sources of information that WHO experts are looking at.

[00:12:17] Speaker 4: But does that limit, does that sort of illustrate limitations in the predictive modeling and in the modeling that when novel strains appear, it's just it can't predict them because it can only basically predict stuff based on the existing training data that it has?

[00:12:34] Speaker 2: It's a slightly different question. It depends what do you formulate. The specific question we asked is, give me the strain, give me their frequency. I will tell you how it would look like in six months. This is a problem that we could solve with existing data. There is a separate question for which you are right. We don't have much data because it's not a very frequent occurrence. Can you tell me who else may appear there? But this is extremely, extremely important question, because when we're thinking about all the zoonotic viruses, when we're thinking about the previous pandemic and others, all of a sudden this kind of dark horse became like the main winner. So I think that, you know, we demonstrated the first step. You can do this prediction in a reliable way, but there are many big questions that need to be answered, is like who and what is going to be the next threat in the same way as we had COVID?

[00:13:35] Speaker 1: If you're able to predict how a disease might change in a particular way, are you applying that same technique to cancers? And are you therefore able to tailor a specific treatment to an individual patient? Yeah.

[00:13:51] Speaker 2: So, yes, you can see, you know, we're looking at different modalities. When we started with, you know, we work on the big images on mammograms, we moved to proteins with really a protein language model and work in that space in differential equations. Now we're addressing a question which, you know, as me who, you know, as a cancer survivor plays really, really a big role. Because, you know, when I went through my treatment, of course, you know, when you go to the cancer ward, you feel miserable no matter what. But, you know, I've seen a lot of patients who, you know, were untreatable. Those are metastatic patients. And one of the biggest challenges with metastatic patients is that, you know, there is no good treatment. There is huge variability in what you would give to them, depending on which hospital they would go. Because, you know, at the initial stages, when cancer is treated, you can bug the patient and say you're going to get this, bugging the patient, say you're going to get treatment A or treatment B. At a time when the patient become metastatic, this is so very different, so unique, changes in a very personalized way. It is very hard to do this mapping. And the specific questions that we started working on, I am happy to say that there are already clinical trials in the United States where patients can enroll in breast, colon, and lung cancer, where the machine learning is used, given, you know, the pathology slide of the patient, the sequencing information, to identify which treatment is more likely to be successful on this patient. So kind of uniting the guesswork that today oncologists need to do.

[00:15:30] Speaker 1: Wow.

[00:15:31] Speaker 3: Regina, are we getting to a point, do you think, where we might be able to have a greater opportunity for personalized medicine? So if AI is truly able to stop treating us as a gigantic cohort of humanity and actually look at you, the individual with your scans, your data, and how that compares to humanity, is that where you see this going, or is it going somewhere else?

[00:15:57] Speaker 2: So scientifically, I totally see we are going there. And it's going there like from two directions. On one direction, we have much better capacity today to identify your trajectory, to kind of predict your future. In some ways, we're not truly predicting the future because it's already there. It's just human eye cannot see it. But we can, you know, identify this trajectory and stratify the patients. On the other hand, you know, again, talking about the big breakthroughs that came from UK, you know, the discovery, the alpha fold. And, you know, at MIT, we did both. Now it's in London. It's a company. It's in London. But there's a lot of work on the molecular modeling that enables you to really characterize disease on the molecular level and produce new drugs. So on one hand, you could say this patient is likely to go this way. And in theory, we would be able to generate something that is going to treat this patient. I'm not saying it's going to happen next year, but we can imagine the scenario in which you can really kind of personalize.

[00:17:02] Speaker 1: Do you think that speeds up the bottleneck that there is in drug development? Because we know from from conception to production, there are enormous amount of years in between. Do you think that time frame starts to shrink if that's possible?

[00:17:17] Speaker 2: I think it already does start to shrink because, you know, before you need to do humongous screening processes. And there was a lot of failure. And we know that majority, of course, goes to fail attempts to the clinical trials that happen to fail, you know, in the late stages when already humongous amount of money and time was spent. So I think AI can really help to make it faster. Now, can it make it faster to to the extent that, you know, it will take us, you know, one month to produce new drug? We are not quite there yet. But I do think there is a meaningful speed up across many different parts of drug discovery, starting from understanding disease mechanism, selecting the molecule and running effectively clinical trials.

[00:18:11] Speaker 1: Just a final thought. You're right at the cutting edge of this. Paint a picture for people of where cancer treatment will be in 10 years time.

[00:18:24] Speaker 2: So what I hope, what I hope will happen is we will have tests, which will be the same test that you can do in the general practitioner. It's not going to be an MRI images. Maybe you're going to be doing some simple blood test. If you regularly go, you're screened and you may be identified at risk in certain conditions. And at the same time, very, very early on. And the machine would be able to propose you a set of modification in your lifestyle, maybe diet, maybe some lighter changes that you can make that would make you kind of be in a healthier zone. And if in some cases we cannot do it, we hope that at this point you can identify a treatment as early as possible and less kind of toxic, because, you know, cancer treatments are terribly toxic, that will more likely to result in a better outcome. So on one hand, it's going to be much earlier detection. Second, understanding how your general conditions in your life are contributing to the development of the disease. And the third one, a personalized, non-toxic treatment with high efficacy. This is a hope.

[00:19:43] Speaker 1: And survival rates.

[00:19:45] Speaker 2: And again, I really hope that even if we solve the first part of early detection, because we know how to treat the cancer early, we can dramatically decrease deaths. And also, I really hope that the projects we are working on and many other scientists around the world, including the UK, that we would be able to do much better matching with existing drugs to make sure that patients can live better and longer, and also come up with a new drug that can really cure the advanced disease.

[00:20:19] Speaker 1: Wow, that is exciting news. It's good to have a good news story on AI Decoded, isn't it? And that is certainly a very good one, which we can all get around. Thank you very much indeed for that. Listen, what I want to do just in the short time we have left on the program is look a little bit at some of the questions that you've been sending in. We've been getting some fantastic questions from viewers on the email link aidecoded.bbc.co.uk. So, I want to take a few minutes to get to some of the best of those. Here's one from Veronica, who asks, Do governments need to step in and make it compulsory for schools to teach AI, including its biases, how to spot fake content online, and how to distinguish what is AI generated from what is real? Stephanie, why don't you pick that one up?

[00:21:09] Speaker 3: I'm a bit nervous about this because some governments are actually using AI for misinformation and disinformation purposes. So, looking for ethical guidance from them is perhaps not the solution we would love and expect. I think instead what we could see rather is push from the ground up. So, parents wanting it, teachers wanting it, the students themselves wanting it, and creating space for that in the curriculum where it's being brought in from the bottom up and also laterally, horizontally asking each other. I don't know if I would want it coming from government on down. Even countries like the United States, which is ostensibly a liberal democracy, the White House is one of the biggest users of AI misinformation and slop. So, just be careful what you wish for on that one. But I think we can all agree more education and critical thinking skills, particularly for young people, but also, frankly, some of us who are not so young is great.

[00:22:07] Speaker 1: All right. Here's another question from Trudy, who is taking us truly left field here. She's written in with this. My cats seem to prefer cartoons to live action TV or films. And I've even noticed my cat reacting to a clip of an AI generated film that you showed. Is there something about animation or AI generated visuals that animals perceive differently?

[00:22:32] Speaker 4: This is something I'm hoping, well, I'm hoping here that we can recruit the audience to help us out a little bit with this one. I am very much a cat person. Unfortunately, I have allergies, quite severe allergies, so I can't have a cat at home. But after seeing this email, I thought, OK, let's conduct an experiment here. So we made a couple of clips for domestic animals, for cats and for dogs. So I think we can play those clips right now. Hopefully you've seen those clips. You can you can go back and view them again on iPlayer if you need to. Show them to your animals. See if they respond to them. And I would love to find out from the audience what their animals responses were to those clips. They could even film it.

[00:23:32] Speaker 1: This is citizen science. I show my dog things on my iPhone that I've recorded. And I know the dog has no idea. There's no reaction, none whatsoever. So I'm sort of intrigued as to why AI generated film might be stirring some sort of interest. What do you think, Stephanie?

[00:23:55] Speaker 3: I mean, this is the black box. Who knows how it works? No, I'm just teasing. I think it may also depend on the colors.

[00:24:01] Speaker 1: We should be asking the scientists. I mean, we should be asking Regina. She may have even studied this. Have you studied this, Regina?

[00:24:09] Speaker 2: I did have cats, but it was before the AI time. But I will tell you what, actually, now we're trying to understand there is some content, as we've seen from the letter, when the cats are excited and some which is not. And you can imagine that if we want to fully automate the generation of movies, we can just have an AI system that, in mass, trains on the cats and their reactions and then creates a customized content, which would really make the cats captivated for hours. It may not be good for the cat, you know, health.

[00:24:47] Speaker 1: I can see why you're on the Time Magazine 100 list. Look at that.

[00:24:52] Speaker 4: I'm fascinated to see what the results of this are. Because those two clips were completely generated by a prompt, and I also asked the model what it thought cats and dogs might want to see as well. So everything that we're seeing here, the images that are created, not just entirely AI generated, but the prompt itself was basically AI generated too.

[00:25:16] Speaker 1: Yeah, I can't get the remote control in our house as it is. I'm not giving the dog a reason to take it as well. I mean, I'll never get the football on. Listen, that's all we've got time for this week. Keep the questions coming about cats and AI, if that's where you want to go. We love hearing what has grabbed your attention. Thank you to Regina. Thank you to Stephanie and Mark, as ever. Hope you'll join us next week for more AI Decoded. Thanks for watching. AI Decoded