Mass. General Brigham’s Chief of Breast Imaging, Dr. Connie Lehman, MD, PhD talks about AI making breast cancer imaging a lot better.
Interview conducted by Ivanhoe Broadcast News in 2022.
How do you and your colleagues go about improving knowing what that risk is for women?
LEHMAN: It’s the most common conversation I have with patients after sharing with them their biopsy came back and it’s having breast cancer in it. The shock is associated with it’s almost not possible. I don’t have anyone in my family that has breast cancer. I don’t have a genetic mutation, I don’t have the gene for breast cancer, how can this be? That’s the most common experience because 85 percent of patients diagnosed with breast cancer have none of the risk factors except being a woman and getting older. And in fact, we’re finding more younger women are developing breast cancer. So even that part of the story is changing. So we can do so much more. First, we need to inform women of their risk. We need to get better at identifying those women who truly are at increased risk and reassure those women who are not at increased risk, because most women will never develop breast cancer.
You and your colleagues use artificial intelligence to push that way, to push the envelope, can you talk to me a little bit about that? Again, for our viewers, when you say artificial intelligence, they may not have a good concept of what you do. Can you explain in a nutshell what it is that you and your colleagues are working on?
LEHMAN: I cannot remember a time when I was more excited about a new domain of technology. It’s been amazing in radiology. We went from old analog films of the breast tissue to full fill digital mammography to now 3D mammography and even contrast enhanced mammography. The technology boom was amazing, but it was exceeding the human eye and the human brain’s ability to extract information from those beautiful images. And then we had artificial intelligence, computer vision, neural networks, deep learning, fast computers. That in other domains of computer vision were showing, we could have computers that exceed the human eye, the human brain’s capacity to pick up patterns, whether it’s patterns on a map, the work that’s been done in the military, in other areas, in health care, patterns of skin cancer from an image or a cancer that’s very subtle on a pathology slide. So computer vision is just exploding the field of radiology because now with all this amazing technology, we’re actually extracting more predictive information from images of the body to help guide best patient care.
What does that do? Does it mean much earlier detection or just showing which women maybe don’t have cancer now but may be at higher risk, or is it a little bit of both?
LEHMAN: With these tools, we can do both. We can both detect cancers earlier on the mammogram, we can see and find patterns and have areas flagged that we otherwise wouldn’t have noticed, but what I’m really excited about it as we can also guide women in understanding you may not have a cancer evident on your mammogram now, but you are at increased risk for the next 5-10 years. And so we want to talk to you about a more effective screening strategy for you. Now, we’re not going to do all of these other tests in women who are at low or average risk because the benefits wouldn’t outweigh the harms of all that additional testing. But in these women that are at very high risk, it can be lifesaving.
What does that equate to in terms of lives saved? You might not have a hard number, but are you able to speak to that a little bit?
LEHMAN: So we found that when you identify women at increased risk and you find their cancers earlier, we can save those women from, if the disease is detected late from being a breast cancer that actually cannot be survived to one that is completely cured, 99 percent cure rates when we detected early, it’s that much of a shift. But we do need to bring in more than just traditional mammography for those patients such as MRI or contrast enhanced mammography. And even more exciting, we have approaches where we can actually prevent the cancer from happening, and that’s where we want to get. Not just assessing risk and being better at early detection, but let’s reduce that risk, let’s prevent that cancer. And that’s a really exciting domain that we’re exploring now with AI.
How do we get to the point where it’s in a study phase to accessible to most women? Is there a timeframe and a way to do that in your mind?
LEHMAN: I think there is a really reasonable pathway that we need to follow. Right now, all of these tools are in the research phase. So we’re studying them, we’re evaluating them. We want to be very careful that our early enthusiasm is matched by the rigors of our scientific methods. But I do think the pathway is one where we then look to see how can we move out of the research domain and into clinical standard use. And that is typically going to be done with engagement of regulatory bodies, very careful study is it safe? What are the claims of it? Is it effective? Do we have ways to mitigate the risks? So for example if we have a risk tool that tells a woman that she’s at low risk, but she’s at high risk, how can that risk be mitigated of misinformation? No test is perfect. It’s like a screening mammogram. We’re delighted that the woman’s mammogram is negative, but we can’t tell her that she doesn’t have breast cancer. And I think our patients are actually when we engage in the right way and we educate in the right way, they understand that balance of benefits and risks. But what we’re really trying to do is having more accurate, more equitable, more precise methods to predict a woman’s future risk of breast cancer. And it really looks like from all the research that we’ve done that we’ve found a very robust pathway to do that using a mammogram.
We said this is still in the research phase, can you talk to me a little bit about what you and your colleagues are doing? Are you currently enrolling numbers of women or are you’re studying this and running through AI, running through computers, running through some of these high-tech to see what comes up on the other side?
LEHMAN: We haven’t yet rolled this out in a prospective clinical research trial. What we’ve been doing is taking registry data. So deidentified stored mammograms where we know the outcomes after that mammogram was taken and we know the AI risk scores. We’re able to simulate how accurate was this. Not only in our total population, but in really important subgroups of patients. Was this accurate in our patients who identified as African American or Black, is Hispanic or Asian? Because that was one of the domains that shocked me the most when I started studying better ways to predict a woman’s risk.
What about the inequities?
LEHMAN: Exactly.
Is there a time where AI could level the playing field?
LEHMAN: I am most excited about this domain. There have been some areas in artificial intelligence and computer vision where it has been very concerning that the model is being developed have very strong racial biases. So, for example, if a model was trained to identify skin cancer on all white patients, that is not probably going to work very well and hasn’t been shown to work well in patients who identify as patients of color. So we were concerned about that in our research. It turns out when we’re looking at mammograms and X-ray of the breast tissue, we don’t see those racial disparities. There’s something that the computer is able to detect in the patterns of the mammogram that are equitable across races. So even though our training data, which was over 200,000 exams, which were about 85 percent white patients and only 15 percent of other races and ethnicities, we still found that it performed at a very high level, the same as in white patients when we tested it in other patient populations, Black, African-American, Asian, Hispanic. So that was exciting to us that this was a tool that could correct the racial disparities we have in our breast cancer risk models. And those are very well-documented and very concerning. For example, at Mass General, we studied patients who had values from traditional risk scores. Traditional risk scores might ask a woman “When did you have your first child? When did you have your first period? Do you have family members with breast cancer?” And then they calculate her risk score. We found that white patients were two-and-a-half to three times more likely to be identified at increased risk compared to our Hispanic, Asian, and Black African-American patients. But there were no differences in cancer rates. So that was concerning because these risk models are used as gatekeepers to allowing women to have access to life-saving, supplemental imaging, prevention strategies, risk reduction strategies, or to not.
Is there an indication that AI can also help? Could AI cut down on those repeating times and more testing on the stress and the worry?
LEHMAN: Absolutely. In fact, AI can address and really hit the challenge of human variation. Sometimes we’ll talk about, Oh, the sensitivity of mammography, the ability of a mammogram to find a cancer is 85-90 percent but there is no performance of screening mammography, it’s in a human’s hands. So we have radiologists around the world that have very high rates of cancer detection and we have radiologists that are much lower. We have radiologists with very high rates of false positive assessments and radiologists with very low false positive rate. So the artificial intelligence and the computer vision techniques that we’re using can help more accurately identify a mammogram that’s likely to have cancer present now, and we can fine tune the AI to partner with the radiologists to meet that radiologists where they need help the most. So you can adjust the feedback to say, here’s someone that really struggles in seeing that cancers. Or you’re someone that really struggles with knowing this is normal tissue. You don’t need to do a biopsy. You don’t need to call this patient back.
Is there anything that you want to make sure that we highlighted that people know?
LEHMAN: Is there something to highlight a statement? I want to say like something really big. In many ways, the pandemic opened our eyes to both the challenges, but also the opportunities we have to move the field of healthcare forward. In my domain, we certainly saw increased inequities in how we deliver care, how we assess risk, how we ensure that every patient that we are taking care of has access to the same high-quality care. And artificial intelligence and the technology of AI can address those challenges, can help us have a better health care system. We have to make smart decisions with it. We have to be careful about how we do our research, but I couldn’t be more excited about the impact that I see it can have. With artificial intelligence and these models that we create, it’s the same mammogram. A woman still goes in, has a mammogram, but then we evaluate that mammogram. Not only the human the expert radiologist is evaluating it, but also the model is evaluated and providing more information than just the simple human reading alone.
END OF INTERVIEW
This information is intended for additional research purposes only. It is not to be used as a prescription or advice from Ivanhoe Broadcast News, Inc. or any medical professional interviewed. Ivanhoe Broadcast News, Inc. assumes no responsibility for the depth or accuracy of physician statements. Procedures or medicines apply to different people and medical factors; always consult your physician on medical matters.
If you would like more information, please contact:
Katie Marquedant
(617) 726-0337
Sign up for a free weekly e-mail on Medical Breakthroughs called First to Know by clicking here
