Artificial Intelligence vs. Cancer

Dr. Diane Reidy-Lagunes:

Artificial intelligence. We seem to be hearing about it everywhere.

News Montage:

“We're talking about one of the hottest topics, artificial intelligence.” 

“Artificial intelligence is finding its place in all sorts of scientific fields, and perhaps none holds more life-saving promise than healthcare.”

Dr. Diane Reidy-Lagunes:

We're already seeing the power of A.I. in medicine. It's helping to spot abnormalities beyond what the naked eye can see. And it may prove to be a significant game changer in cancer care with the promise of detecting cancers early; as well as in cancer discovery, leading to more successful treatments. But does artificial intelligence live up to the hype and what are the concerns? Let's talk about it.

Hello, I'm Dr. Diane Reidy-Lagunes from Memorial Sloan Kettering Cancer Center and welcome to Cancer Straight Talk. We're bringing together national experts and patients fighting these diseases to have evidence-based conversations. Our mission is to educate and empower you and your family members to make the right decisions and live happier and healthier lives. For more information on the topics discussed here, or to send us your questions, please visit us at mskcc.org/podcast.

Today we are joined by Dr. Sohrab Shah and Dr. Larry Schwartz. Sohrab is a PhD scientist who leads MSK’s computational oncology program. His work centers on computational data to understand cancer evolution and how cancer spreads and develops resistance to treatment. In summary, his work is using A.I. to foster cancer discoveries and hopefully identify new cancer treatments. Larry is the Chair of the Department of Radiology here at MSK. He's known for his innovative use of new technology and cancer care. His current work centers around the use of artificial intelligence in the world of radiology, and his research is helping to create and use more A.I. supported tools for cancer imaging, including cancer detection and early cancer response. Thank you both for joining me today.

Dr. Sohrab Shah:

Thanks for having us.

Dr. Lawrence Schwartz:

Thanks.

Dr. Diane Reidy-Lagunes:

So we are super excited for this pod because it feels like everywhere, A.I. is the sort of buzzword that's going on. In most basic terms, could you just help us identify exactly what is artificial intelligence?

Dr. Sohrab Shah:

Yes. It's really the science of making intelligent computer programs, and that can mean different things to different people. One example is to learn patterns in data and that can be used to determine whether an image contains a certain object, for example, a cat versus a dog. And that is very powerful because it allows us to make decisions based on information and data.

You can imagine in cancer, we have a lot of information that's encoded in digital form and we might want to, for example, take a scan of the cells of a tumor under a microscope. A.I. can be deployed to determine whether there are cancer cells in that tissue or not.

A.I. has been a powerful tool that allows us to really synthesize information, and that helps with decision-making. It helps with making predictions. It helps with statistical analysis of data as well, and is now obviously making a huge impact on society.

Dr. Diane Reidy-Lagunes:

And Larry, do you have an understanding of A.I.'s potential as it relates to medicine, in particular cancer care?

Dr. Lawrence Schwartz:

I think there's a great potential for it. Certainly in the imaging field we've actually had A.I. applications for close to a decade. For instance, in many settings a healthy woman's mammogram is second-read with an A.I. technology to make sure that nothing was missed. But we're certainly just at the beginning of the journey to understand what A.I. could do. I think in most health systems, 40% of the data generated is medical imaging, is radiology. Then you have the other 60%, so putting that all together, it's very difficult for an individual to do. A smart machine could really help us.

Dr. Diane Reidy-Lagunes:

Before we dive a little deeper into understanding how the smart machine is going to help us cure cancer, hopefully one day, there's a lot of concern about it too, right? Larry, can you talk to that piece in terms of like what there might be concerns about in regard to patient privacy, and/or any other kind of concerns using this type of technology to help our patients?

Dr. Lawrence Schwartz:

There certainly are concerns now, but there are ways to protect those concerns and really truly protect patient privacy. In the medical imaging world, for instance, every patient scan is identified by their name, a medical record number, a date – frequently their birth date – and their own referring physician. We're able to strip off and eliminate all of that information before we actually analyze the image. In fact, that's an absolute requirement. We actually use A.I. to check, to make sure that there is no privacy information on the image itself.

We actually have to take it one step further. If you think about a patient that may have a CT scan or an MRI of the brain or the neck, there's information about the face there, right? Theoretically now, you could reconstruct the face and may be able to identify the person. We have ways to deconstruct that ahead of time. And it's actually a requirement of ours that we do actually deconstruct that so that the facial recognition could never, ever possibly be used on that data set.

So there really are a lot of ways to protect. And then of course, regulators decide how and when A.I. can appropriately be used. I do think that our regulators are getting a lot smarter about this.

Dr. Diane Reidy-Lagunes:

Absolutely. Sohrab, can you share with us a little bit about how those data sets are created in the world of scientific discovery and genetics, and in particular, the potential to collaborate with others to inform that your data set is robust and clean and valid?

Dr. Sohrab Shah:

One of the major challenges in our field is that in many institutions like ours, we have data sets within our own four walls. We might train a model based on that data set, but it doesn't generalize well to a data set that was generated in a different institution. There are some barriers there because of the privacy issues of sharing data across institutions. So this is a challenge that we have to systemically address whereby institutions should try to collaborate on this in order to really achieve the promise to improve and enhance cancer care for the population writ large. I would really hope that many institutions can get together and start to share data in an anonymous way that protects patient privacy.

Dr. Diane Reidy-Lagunes:

So let's get to the applications once we have those data sets that we feel are clean enough for which we want to learn from. Sohrab, can you explain to us your fascinating work, particularly in ovarian cancer, and how you are using those data sets to better understand the science behind it?

Dr. Sohrab Shah:

One of the powerful attributes of taking these computational models is actually to be able to synthesize different types of data. Cancer is fundamentally a disease of the genome where ultimately mutations that accrue in the DNA of certain cells will result in a transformation of that cell's normal behavior to a cancerous behavior. But that's only part of the story. If we only measure the genome, we will only be able to learn so much.

What's fascinating about how our field is progressing is that the number of different forms of digital information, including radiology, that Dr. Schwartz is working on; or digital pathology, which is essentially a microscopic view of the tumor that can be digitized and become a piece of data incorporated into a model.

So what we've been really focusing on is trying to understand: How do these different forms of data fit together to make a comprehensive view of the tumor, or even the patient at large? And this is really where some of the A.I. models and progress in computer science are shining a light on a holistic view of a tumor, or even a holistic view of a patient. If we can start to integrate what can we learn from radiology, pathology, genomics, and even the clinical record, then we have a real chance of learning from every patient that we see. And this is all data that's procured typically in the provision of care. It's an incredible opportunity to bring these different sources of data together using A.I. models to make better predictions, for example, of whether a patient might respond to a given therapy. It's taking personalized medicine to a next level.

Our research is really at the very beginning stages of this, but we've shown some really, really exciting proof of principle that if we execute this model, pulling different sources of information together, we get a much better predictive capacity of whether a patient will respond to a therapy or not.

Dr. Diane Reidy-Lagunes:

Based on that A.I. and based on their genetics and their radiology, all together.

Dr. Sohrab Shah:

Exactly.

Dr. Diane Reidy-Lagunes:

Larry, talk to us a little bit more about the applications of A.I. and radiology today, and where you're expecting the ball to go.

Dr. Lawrence Schwartz

I'll give a few examples of it. One example would be in the acquisition of a scan. Many patients have follow-up scans – CT scans, MRI, PET scans – and they take a long time. In fact, they used to take up to an hour to do each of them, and the reason is that it takes that long to collect all of the information, all of the data, in the scan. Utilizing A.I. technology, now we're actually able to collect that data more rapidly. So nowadays if you come to Memorial, your scans take half the time that they used to a few years ago, and we'll continue to cut that time down.

Another application is in detecting lesions. We do that now visually to identify areas of abnormality. The reason why patients really shouldn't be afraid of this is that the radiologist oversees everything the computer identifies. It just makes us a smarter radiologist. But really the big advance is, not only to do it faster and do it equal to what the radiologist could do, but actually to do it better. There are patterns we've learned that may indicate malignancy or may indicate a lesion is more likely to be benign, especially as it changes over time. So as we apply the A.I. algorithm to that, and then have the radiologist interpret it, we could say with more certainly, “That's a tumor.” Or what's even better is to say, “That's not a tumor and you don't need any follow up. You can go about your life and you're going to be absolutely fine.”

Dr. Diane Reidy-Lagunes:

And that's a beautiful thing, because I can say firsthand our patients get so nervous about these things. We say, “Come back in 12 weeks,” and they're worrying about it for 12 weeks.

Dr. Lawrence Schwartz:

Right.

Dr. Diane Reidy-Lagunes:

I think that's really important clinically. Currently, as you both know, there are real world applications of A.I. in the use of dermatology. I had a chance to speak with Dr. Allen Halpern, a dermatologist here at MSK who specializes in early detection of skin cancer, particularly melanoma. I asked him about how A.I. was being used in the field of dermatology and how it's made a difference. Let's take a listen:

Dr. Allen Halpern 

A.I. is now touching absolutely everything, right? And since dermatology is a visual specialty, the real exciting part for us is using image based A.I. to help us do a better job diagnosing skin conditions and especially skin cancer.

Dr. Diane Reidy-Lagunes: 

Could you explain to us a little bit about some of those consortiums and those big databases, such as the International Skin Imaging Collaboration, that you've played such a pivotal part in?

Dr. Allen Halpern 

As you're all probably beginning to hear and understand, the way A.I. works is it needs huge data sets to train on, and then it just gets smarter and smarter based on those data sets. And in dermatology, that's a bit of a challenge because the A.I. really needs pictures that have appropriate, accurate labels. What is it? What is this rash? What is this spot? And the reality is that those images, when they exist in medical centers, are all siloed and there's a lot of privacy issues around them because these are people without their clothes on. So we created the International Skin Imaging Collaboration back in 2014 to create an archive of images that A.I. can train on, that could be in the public domain without any privacy concerns, and for which we have really accurate diagnoses. And then we've been busy engaging the computer scientists to use this resource and helping move the field along, including creating standards for A.I. for dermatology.

Dr. Diane Reidy-Lagunes: 

So clearly, some huge potential in helping us identify these lesions.

Dr. Diane Reidy-Lagunes:

Sohrab, one of the struggles with A.I. is the data and ensuring that it is inclusive and that there's equity in the interpretation of it so that that all patient populations are included. Can you speak to that a little bit?

Dr. Sohrab Shah:

Yeah, I think this is a critical element, especially as we move to deploying A.I. in a healthcare setting. We wanted to make sure that the tools are functioning well for all patients, and the risk is that models that are trained can only make predictions on the data that it's seen. So if the input data is not representative of the population as a whole or is not representing rare populations, then those models will not be as powerful for those patients. So there's a real concern about equitable application and equitable care.

I think what we as a field have to do is make sure that the models that are actually deployed in clinical practice, if and when that happens, pay very close attention to how those models were trained, what populations were represented, and which populations were not represented, so that we have a better idea of where it's going to derive benefit and where it may not derive benefit. As we move towards clinical implementation of these models, this is really a critical element that we must consider.

Dr. Diane Reidy-Lagunes:

I want to go back to the future of A.I. and cancer discovery. You talked about the critical importance of trying to merge data in different academic centers. I think we're getting there, albeit very slowly, in terms of even just the clinical data. Now we have similar electronic medical records in many centers that may be at least a piece to get us there. Where do you see A.I. in the next five years in terms of understanding the science behind that cancer? Because we talk a lot on this pod about cancer being thousands of diseases. Even when we're just calling it breast cancer, it's really not. There are hundreds of cancers that we call breast cancer. So I'm just curious if you think that A.I. could help us better understand the different types of cancers that we have?

Dr. Sohrab Shah:

Absolutely. One thing we haven't mentioned in this discussion is really the emergence of large language models. Your audience will be very familiar with the idea of ChatGPT. These types of models are called generative A.I. and the way that they work is by ingesting huge amounts of data. When I say huge amounts of data, I mean learning patterns of all of Wikipedia or all of the internet. The way these models work is with a prompt, and so you can ask a question and based on that question, the models will generate an answer that is derived from this language that it learns. So to someone who's consuming that response, it really reads as though a human could have written that sentence.

The reason I bring this up is because I think the text information that you're talking about in the electronic medical records are a great example of this. I think we're just at the very beginning of being able to now deploy these large language models on clinical record information. And I do believe that, combined with molecular data, we can very powerfully extract what treatment that patient was given from the medical record. And based on the genetics, the mutations that are present in that tumor, the treatment that the patient was given, and the response of the tumor to that treatment, we can really start to learn the associations of what mutations might be predictive of response to a given therapy. And that's just one small example.

You can take that further to what patterns are present in radiology or what patterns are present in the digital pathology. And by pulling all these different forms of information together in what we call multimodal integration, the text information that's present in clinical records will become critical to making these predictions. In the next 10 years, I'm sure we'll see major progress in that area.

Dr. Diane Reidy-Lagunes:

Yeah, I mean, it could really flip our whole paradigm upside down. Right now, if I care for patients with colon cancer, every patient irrespective of their genetic analysis – more or less in first- and second- and third-line setting – are all going on treatment. Then we see who responds and then look backwards to say, “Okay, what were their genetics?” to see who may or may not have responded. But what you're saying is we have the potential to look at all the data and then start to really define their treatment based on what the data's telling us to give to that patient, which is very powerful and very exciting.

Larry, along those lines, can you share with us a little bit about the potential to use A.I. to know earlier rather than later if a treatment is working? I mean, right now, if I have a patient receiving therapy, generally we wait at least 8 to 10 weeks before we know if it worked or not, and the scan just shows if it shrank or if it stayed the same or if it grew. Any potential for A.I. in that space?

Dr. Lawrence Schwartz:

The reason we had to wait 6, 8, 10 weeks is we were simply measuring tumor size. But now with A.I., we're able to detect and correlate with outcomes the features of the tumor that occur much earlier – a week, two weeks, three weeks – quite easily. And not only that, but then we can map out these over time.

You talk about big data: You take a large number of patients and add multiple scans per patient and multiple time points per patient. It really does take an A.I. engine to sort through this and understand the disease trajectory, how much treatment the patient should get, when they should get the treatment, when they are actually responding, and unfortunately, when the tumors are progressing. Can we tell that earlier so we could actually switch the treatments? To your point, it really takes an understanding of a large cohort of patients that have been on therapy, and to analyze how well they've done or how haven’t done.

Dr. Diane Reidy-Lagunes:

Amazing.

Dr. Diane Reidy-Lagunes:

Any final thoughts, Larry?

Dr. Lawrence Schwartz:

It's an exciting time, but it’s also really a hopeful time for patients. I think we will be able to make these changes and advance this technology. One final thought: We've talked about data sharing for a long time, and it certainly has been important. I would say it's imperative now because we have a technology that we could advance so far with.

Dr. Diane Reidy-Lagunes:

Sohrab?

Dr. Sohrab Shah:

When we look back 10 years from now, we will look at the years 2022 and 2023 as an inflection point in the use of A.I. in society writ large. But it will also be looked at as a time where we could start to adopt these models using A.I. for good, if you will. That means deploying these as tools in our arsenal to treat cancer patients more effectively. We're in the beginning of this revolution, and it's very exciting to think about the impact of that in the future.

Dr. Diane Reidy-Lagunes:

Totally agree. Thank you both so much for joining me today.

Dr. Lawrence Schwartz:

Thank you.

Dr. Sohrab Shah:

Thanks for having me.

Dr. Diane Reidy-Lagunes:

Thank you for listening to Cancer Straight Talk from Memorial Sloan Kettering Cancer Center. For more information or to send us your questions, please visit us at mskcc.org/podcast. Help others find this helpful resource by rating and reviewing it on Apple Podcasts or wherever you listen. Any products mentioned on the show are not official endorsements by Memorial Sloan Kettering. These episodes are for you but are not intended to be a medical substitute. Please remember to consult your doctor with any questions you have regarding medical conditions. I'm Dr. Diane Reidy-Lagunes. Onward and upward.