Want to take quizzes and track your credits?
According to a 2011 study in Transactions of the American Clinical and Climatological Association1, in 1950, researchers predicted it would take about 50 years for all available medical knowledge to double. But in 2020, estimates peg it at 73 days. It's not humanly possible for medical professionals to keep up with the influx of constant new information about health conditions, treatments and medical technology which is why AI and machine learning pose such growth opportunities in health care. Mark Michalski, MD, executive director of the Center for Clinical Data Science at Massachusetts General and Brigham Hospital discusses the history of AI and how it's being used today. He also examines how it might change radiology and how radiology can leverage this new technology to provide even more value and better patient care.
Sign in to take quiz and track your certificates
RLI Taking the Lead podcast series explores the unique career journeys of radiology's most influential leaders to provide practical insight into how to structure a career in leadership and find success across a spectrum of clinical environments and organizations. Learn more
American College of Radiology, through its Radiology Leadership Institute (RLI), offers this podcast as one of a series of educational discussions with radiology leaders. The podcasts reflect the perspectives of the individual leaders, not of ACR or RLI. ACR disclaims liability for any acts or omissions that occur based on these discussions. Listeners may download the transcript for their own learning and share with their colleagues in their practices and departments. However, they may not copy and redistribute any portion of the podcast content for any commercial purpose.
Scott: Hello, and welcome to part two of our discussion on artificial intelligence. I am Scott Pearce, with the American College of Radiology. And for our discussion today, we are very happy to have with us Dr. David Louis, the pathologist and chief at Massachusetts General Hospital. As radiology is beginning to grapple with many of the implications of what AI and machine learning may mean to the profession, pathology offers an outstanding use case of how radiology can approach these challenges, and possible [00:00:30] ways forward. Dr. Louis, thank you so much for taking time out of your busy day to speak with us.
Dr. Louis: Glad to chat.
Scott: While radiology is recently entering into the AI realm and beginning to consider what the future of its impact might be, you and your specialty have actually been working on similar approaches for several years. Can you please speak about some of the history of computer-aided diagnosis in pathology?
Dr. Louis: In pathology, we've been using computer-assisted diagnoses for many years. But [00:01:00] we, like radiology, are only beginning now to get into the area of artificial intelligence. But pathology has been doing a few things over the years that are relevant to how it approaches artificial intelligence and its contribution to diagnosis. Firstly, it's important to realize that in the history of pathology, the laboratories have for many years done direct reporting of [00:01:30] results. This is not necessarily true in the anatomical pathology sphere, where we look at slides in order to make diagnoses. But in the clinical pathology or laboratory medicine sphere, we do blood test analyses and other fluids, and those results are often reported directly to physicians after going through internal quality assurance standards. [00:02:00] And so, there's a precedent for automation and automatic reporting happening in our field without a pathologist directly intervening in test resulting in particular instances. The second aspect of what's happened in pathology that's relevant to the emergence or the possible emergence of AI is that we've been [00:02:30] using image recognition technologies, all computational, for a variety of applications over the past decade or so.
And these are applications that are increasing in use and are all running on FDA approved devices right now. Examples of those would include the screening of Pap smears in the liquid-based format, the screening of [00:03:00] urinalysis, and the screening of blood smears. All of these are screening approaches, where the computer is looking through at single cells, whether they are cervical cells for Pap smears, or blood cells, or cells and casts in urine. The computer is looking through, comparing that with databases and telling the pathologist or laboratory technologists that [00:03:30] certain samples don't have any suspicious cells in them and others do have suspicious cells. In this manner, we can basically auto-report the cases that have no abnormality seen in them and the computers do a very good job. They are set to be highly sensitive, rather than entirely specific, but they do a very good job of not missing [00:04:00] abnormal situations. And then passing those abnormal situations over to a pathologist or a cytotechnologist for further evaluation. So in the area of screening technologies, we are already using those approaches.
Scott: I recently read an article from 2013, where you talked about pathology and data crunching, and how you were using computers to analyze recurring patterns and trends to be able to make diagnosis much earlier. [00:04:30] Do you have some examples of how this has been implemented?
Dr. Louis: No, I think there are many other instances. So in pathology, there are three areas that are being pushed forward by these types of approaches. The first are image-based, things, like, image-based situations, like, we were just talking about, where we have an image, like, a radiologist has an image and you're using some type of digital image recognition [00:05:00] software to come up with an answer. The secondary in pathology that we use computation for diagnostics is in laboratory results coming out of the clinical laboratories. So, for example, all of the blood counts or the glucose results, and I can give you examples of those. And then the third area is bioinformatics, where we now are sitting on large biological [00:05:30] data sets, such as next-generation sequencing results from cancer or germline sequencing, where we need to go through complex analytic pipelines to move from the raw sequencing data to identifying abnormalities in the DNA sequence, that are thought to be pathogenic. So each one of those has examples associated with it. [00:06:00]
And there are some interesting examples I can give you. For example, in CBC, is the complete blood counts, you can do mathematical modeling as John Higgins has done in our department to demonstrate that you can use a lot of the data that is already sitting in CBCs to do additional analyses and predictions. So, John has a model where he's able [00:06:30] to predict the onset of anemias in patients, a couple of months in advance of when the standard looking at the CBC by a doctor can do the same thing. He's then taken that data to go on and personalize how you can use knowledge about how different people change their blood counts over time to look at personalizing hemoglobin A1c [00:07:00] target measurements. So all of that is involving mathematical modeling, it's not artificial intelligence per se. But you can imagine then taking those same examples and using AI to do the next set of analyses.
Scott: How do you see the movement in your specialty from computational screening approaches to the actual use of machine learning in AI applications?
Dr. Louis: Well, this, like, in radiology is really only beginning right now. [00:07:30] But we've started to explore AI type approaches in all three of those areas that I just told you about, whether it's digital imaging, laboratory data, prediction, or bioinformatics. For example, at the present time, we're starting to utilize a system that uses machine learning to look at the DNA sequence from cancer [00:08:00] diagnoses and to predict which types of mutations are more likely to be causative mutations. And therefore, more likely to be ones that you would target for treatment if a targeted therapy were available, than ones that are less likely to be causative. And so in any one of those areas, you can begin to think how you could use AI types of approaches to do it at the next level. [00:08:30]
Scott: Were there fears or conversations amongst your pathologists, colleagues that wondered if computational approaches to diagnoses and eventually AI were going to spell the end of the specialty? And if so, how did the profession as a whole address those fears?
Dr. Louis: There have been suggestions made that AI will replace pathologists, but they've largely been alarmist types of suggestions, and none of them have been realistic [00:09:00] in any way. And there have been very, on the other hand, realistic discussions in the literature of how the AI technology might augment what we do, rather than replace us. But you have to also remember that, in pathology, technology has changed the field so many times over the decades, that pathologists, I think, are much more amenable to thinking [00:09:30] about how technology might change what we do. If you think about the difference in how a clinical laboratory doing bloodwork works today on large automated production lines versus the way it worked with the numerous technologists and handheld essays, in the past, it's something that pathologists have realized can help as technology moves ahead. [00:10:00] So the challenge is not to become alarmed by it but figure out the best way you can capitalize on technological advances.
Scott: I know that Mass General and MIT have jointly committed research funds to support the collaboration in the data science space for purposes of early diagnosis. Can you speak a little bit more about these types of projects and some of the work that's being done?
Dr. Louis: Both MGH and MIT have worked in a variety of different ways to move these fields [00:10:30] forward. And they range from individual collaborations between investigators and clinicians at both institutions, to more organized and centralized approaches. There is something called the, “Grand challenge and diagnostics,” that's going on right now, that asks folks at MGH and at MIT to look at the use of novel devices, including wearable devices, [00:11:00] as well as the use of big data in lab data crunching, and other types of data that can be used for these types of approaches. So it's definitely something that both institutions are interested in. And there has been a fair amount of progress.
Scott: Based upon your experiences and what you've learned over the past few years with new computational approaches, if you could go back three years and give yourself some advice, what would it be?
Dr. Louis: I think there are many [00:11:30] take-home messages. One is that, things take a long time, and these types of developments will undoubtedly take a long time to bring to any sort of practical fruition. Secondly, in order to move any of this forward, you need the input of experts to frame the right questions. And it's the pathologists and the radiologists that are going to be able to frame, what are [00:12:00] the most important questions that are clinically relevant and that can be answered in a realistic way? We're not yet at the point of global approaches to diagnosis using AI, in my opinion. We're more at the point of being able to use AI to suggest specific questions. Most of the AI approaches that have been shown today, [00:12:30] answer a specific yes or no question. They don't approach diagnoses in the same global way as a pathologist or radiologist does, looking at data or looking at images. So I think that's a key difference and has to be kept in mind. And the last point I would make is to say to radiologists, as well as pathologists, “Enjoy it. It's an exciting time to be in the field [00:13:00] and try to be a part of it because it really could change the way we do things and change them for the better.”
Scott: Great. Thank you very much for taking some time, Dr. Louis, to speak with us today. I know that a lot of radiologists agree that it is a very exciting time to be in the field and very many of us are looking forward to seeing what these new technologies will mean to the specialty and to the future of medicine as a whole. So thank you again for taking some time. We appreciate it.
Dr. Louis: You're welcome.
Scott: Thank you [00:13:30] again for joining us for another episode of the RLI “Leadership Insider” podcast. We appreciate you listening. And for more information about the AI case study session that will be taking place at the RLI Leadership Summit with Dr. Louis, please visit radiologyleaders.org/leadership-summit. Thank you very much.
Mark Michalski, MD, executive director of the Center for Clinical Data Science at Massachusetts General and Brigham Hospital
You currently have no searches saved.