How AI is Reshaping Medical Imagery with MIT CSAIL Professor Polina Golland

What this episode is about

The discussion centers on how artificial intelligence is transforming the field of radiology and medical imaging, countering the initial fears that AI would replace radiologists. MIT professor Polina Golland highlights the evolving roles within the profession, the potential for AI-driven tools to enhance diagnostics, and parallels with other occupations that adapted rather than disappeared in the face of technological advancements.

Key takeaways

• Radiology's resurgence is a reminder: tech evolution creates new roles, not just replacements.
• Just as spreadsheets didn't phase out accountants, AI technologies will enhance rather than obliterate medical careers.
• AI's exponential growth may feel slow now, but wait a few years for a radiology revolution.
• Transforming medical images into numbers can revolutionize diagnostics and patient care decision-making.
• Innovative tech in medicine isn't about eliminating roles; it’s about redefining and elevating them.

Who speaks in this episode?

• Kara Miller (Host) · 15 min speaking
• Polina Golland (Guest) · 22 min speaking

What is discussed, and when

1. 0:00 – 5:00: heart failure, Artificial Intelligence, Polina Goland
2. 5:00 – 10:00: heart failure
3. 10:00 – 15:00: heart failure
4. 15:00 – 20:00: Artificial Intelligence
5. 20:00 – 25:00: Children's Hospital
6. 25:00 – 30:00: Artificial Intelligence
7. 30:00 – 35:00: MRI
8. 35:00 – 36:17: Polina Goland, CSAIL

Notable quotes from "How AI is Reshaping Medical Imagery with MIT CSAIL Professor Polina Golland"

• [KARA] Welcome to MIT's Computer Science and Artificial Intelligence Labs Alliances podcast.

  about MIT · at 0:00

• [POLINA] We are in the exponential growth at the very flat part of the exponential.

  about Artificial Intelligence · at 1:07

• She's a professor of electrical engineering and computer science and a principal investigator in CSAIL.

  about Polina Goland · at 3:15

• So the problem that was posed to us is can we convert the image to a number that tells us how severe the heart failure is.

  about heart failure · at 4:33

• My colleagues in radiology at Children's Hospital, with whom I collaborate a lot, already have applications that measure, for example, limb length for kids who might need surgery to correct for discrepancy in limb length,.

  about Children's Hospital · at 22:48

• And now, when the patient goes into an MRI, sometimes 1,000 2D images are acquired, together comprise dozens of 3D scans.

  about MRI · at 32:46

Who and what is mentioned in this episode?

People

• Polina Goland — Person · Main guest · 3 mentions · first at 0:54
  MIT EECS professor and CSAIL principal investigator explaining how AI is reshaping radiology and medical imaging, using heart-failure assessment from chest X-rays and broader workflow/translation challenges as key examples.

Companies & orgs

• MIT — Company · Host institution · 5 mentions
  The podcast is hosted by MIT's Computer Science and Artificial Intelligence Lab (CSAIL), highlighting its research and contributions.
• CSAIL — Company · Research lab · 1 mention · first at 35:40
  MIT's Computer Science and Artificial Intelligence Lab, where significant research on AI in medicine is conducted.
  She's a professor of electrical engineering and computer science and a principal investigator in CSAIL.
• Children's Hospital — Company · Collaborator · 2 mentions · first at 22:48
  Mentioned as a collaborator in implementing AI tools in radiology.

Concepts & topics

• Artificial Intelligence — Concept · Core topic · 5 mentions · first at 0:00
  Central theme of the episode, discussing the impact and future of AI in radiology and medicine.
• heart failure — Concept · Case study · 8 mentions · first at 2:02
  Used as an example to discuss how AI can assist in diagnosing and treating medical conditions.
• MRI — Concept · Technology · 2 mentions · first at 31:40
  Mentioned as a key imaging technology in radiology that could be enhanced by AI.

Full transcript with timestamps

0:00 [KARA] Welcome to MIT's Computer Science and Artificial Intelligence Labs Alliances podcast. I'm Kara Miller. About a decade ago, it seemed like radiology might be dying as a profession. Why train radiologists, the thinking went, if computers were just going to take their jobs? Fast forward to 2025, turns out radiology is growing as a profession. [POLINA] I think saying we shouldn't train radiologists because that will be replaced is like somebody looking at personal computer invention and spreadsheets and saying, we shouldn't train accountants.

0:45 They will be useless. And yet we have accountants decades after spreadsheets were invented. Because the profession itself changed. [KARA] MIT computer scientist Polina Goland talks about why that prediction went wrong, what it tells us about jobs in general, and how she is seeing AI reshape medicine. [POLINA] We are in the exponential growth at the very flat part of the exponential. As in the numbers are very small, but the uptake is extremely fast. So I suspect that in a few years, we will have explosion of tools that are useful to radiologists.

1:24 [KARA] That conversation is coming up in just a minute. But first, CSAIL Alliances is the industry gateway to MIT's Computer Science and Artificial Intelligence Lab. Members get access to CSAIL research, talent, conferences, and courses. If you want to learn more, visit csail.mit.edu/podcast. That's CSAIL, C-S-A-I-L mit.edu/podcast. For a computer scientist, Polina Goland has had to learn a lot about heart failure. [POLINA] So heart failure is a chronic disease where the heart is not quite operating at its perfect rhythm or capacity.

2:15 And so if the left ventricle of the heart is not pumping the blood properly, then there is a build-up of pressure, and the fluid seeps from the capillaries in the lungs into the lungs. And the patients feel as if they can't breathe. [KARA] So let's say you show up at the ER. What are the indications that you might have heart failure? Well, because people with heart failure can have this fluid build up in the lungs, one of the most important things for doctors to do is find out the state of a patient's lungs. [POLINA] If somebody comes in, and they're short of breath or there is any indication that there might be a problem with their increased blood pressure or anything like that, they're likely to get chest X-ray because they are suspected for heart failure.

3:06 And then the physician might look at the chest x-ray to understand whether indeed heart failure has caused the problem. [KARA] And that's where Goland comes in. She's a professor of electrical engineering and computer science and a principal investigator in CSAIL. Interpreting those pictures, it's not simple, and deciding what comes next for the patient isn't either. [POLINA] So this is the engineering part about heart failure. There is a continuous spectrum of how severe it could be. And the physician treating the patient might decide to prescribe increased dose of medication and send them home if they think that this episode will resolve on its own with changed medication dose, or they can be admitted into the emergency room, or they could be admitted into the hospital if the sense is that they will need more continuous and longer intervention to stabilize them.

4:03 [KARA] But how does a doctor decide what to do. How serious is serious enough for these various options in front of them? [POLINA] The physician has access to many different pieces of information, the results of blood tests, vital signs, and so on. And one crucial bit of information is buried in the image. And so the physician could look at the image and understand what's going on. However, that's very qualitative and cannot be integrated in a quantitative way with all the other information. So the problem that was posed to us is can we convert the image to a number that tells us how severe the heart failure is.

4:43 [KARA] So in some ways, your blood pressure is a number. Your temp is a number. Lots of things are numbers. But this picture is an outlier. There's no obvious data you can get from this picture. [POLINA] Exactly. And so in addition to changing potentially, the intervention of that specific patient, converting images into numbers can help physicians to also hypothesize new interventions and then check them or run clinical trials for new interventions much better, because then they have numbers that they can manipulate to understand whether the intervention A is better than intervention B. If the information is in the images, you can't just directly access that information.

5:28 [KARA] So when you started working on this problem, which by the way was when? When did you start working on this? [POLINA] About 10 years ago. [KARA] OK. If I had stopped you as you understood the problem, and you were like, OK, let's start working on this. And I had said, in your ideal world, if you could really make a contribution here, what would you contribute? What would the outcome be? What would you have said? [POLINA] So one thing that happens in all of these projects, and I'll answer your question as soon as I finish the story.

6:00 What happens in every one of these projects is we set out to solve a specific problem, and along the way, we realized that the problem might be different. And oftentimes, the goal changes even as we develop the computational solution. So when we started, we thought that basically, the physicians will have an indicator on their screen of the trajectory of the patient with respect to all of their vital signs, including x-ray. And so we can't practically do this with x-rays because it would take way too long to arrange the x-rays on the screen, and it would take way too long for the physician to flip through all the existing x-rays.

6:36 But if we could convert it to a number, we could just show a plot that shows the physician how the patient reacted to medications in the past is a simple line that goes up and down as the health status of the patient it resolves. But as we worked on the problem, we realized that there are many more applications to this. And one interesting application is actually figuring out that seems to be getting even more traction now than the original one that we started with is figuring out when to release the patient from the hospital.

7:08 Because when those patients who are admitted to the hospital are brought to a stable condition, the physician has to decide what day is the day that the patient will go home, and it's important to not release them too soon because oftentimes, if the patient is not stabilized enough, they will bounce back into the hospital. And that's very traumatic for the patient health-wise and very expensive for the hospital because they don't actually get reimbursement on these very quick bounce back patients. [KARA] So do you feel like, when you think about yourself now and yourself 10 years ago, do you think this problem turned out to be much harder than you thought it would be?

7:49 [POLINA] No. I think it turned out to be as hard as I thought it would be. [KARA] Got it. [POLINA] So we usually seek problems that we cannot solve easily. Like that's what makes it exciting. When I hear a technical problem where I think, I have a hunch that computation can help, but I don't quite know how to solve it immediately. And we start working on it. And frankly, the first couple of years, it's just getting to understand the medical aspects of it and the data before we can make meaningful contributions. [KARA] You said that you would like to do something for radiologists that is similar to kind of code generation for software engineers, which by the way, we see proliferating.

8:30 Is that still the hope, and are we getting any closer to it? [POLINA] Absolutely. And so I think there are lots of applications where the question can be formulated as very specific yes or no, or measure me this, where we're getting to the point where the tools that are developed in academia and in startups are getting to the point where they're clinically useful, and they can release the radiologists from doing a lot of mind-numbing, sort of boring work, and focus on the complex cases where the patient comes in. And it might be heart failure.

9:04 It might be something different. And they look like that patient from two days ago, and maybe thinking about it a little bit longer might be useful for the radiologist. And by removing the burden of making simple measurements, answering binary questions, I think the current tools have lots of promise in that regard. [KARA] So let's talk a little bit where I want to talk about human weaknesses and some of the computational weaknesses. But let's talk about human weaknesses for a second. What are human weaknesses or just or even weaknesses in the medical system that you've come across that you feel like, yes, computing can be helpful with this?

9:41 And it can be obvious things or things that most ordinary people have not thought about. [POLINA] Well, to start with, I spent most of my time with computers. So I actually think of radiologists as superhuman, superheroes. So I think they're extremely-- the good ones are extremely good at seeing nuances in these images and suggest possible causes of the symptoms that they're observing and so on. So I'm in awe of their ability to synthesize insight. I think the biggest problem with humans is that they get tired. And so if we have overwhelming amount of work that we can reduce for them, and they can focus on actually where the synthesis is needed.

10:25 In my view, that's where the perfect integration of computation and human will happen. [KARA] Do you think at all about doctors, either being overconfident on a certain case or even being under-confident, like needing a kind of helping hand with this case because they're not quite sure what's going on? [POLINA] Oh, absolutely. So one of the nuances, recently, we've done a project that is offshoot of what we are talking about. But effectively, it was on calibrating uncertainty in radiology reports. And so to give you a little bit background, oftentimes, radiologists will write things like likely pneumonia or possible heart failure.

11:11 And the question is what do they mean by that? And so we've done work to convert those kind of expressions into more quantitative measures of uncertainty. And what we discovered as part of this work is that if the radiologist is confident in the answer-- so for the cases where the disease is very likely present in the image, they tend to be overconfident relative to the ground truth. And for cases where they need to state absence of something, they tend to be under-confident. So what we found is that the same person can act different with different level of confidence, depending on a setup in which they are put.

11:55 And so I think it's very fascinating because the radiologists transmit this in uncertainty, and then treating physicians act on that information. And so right now, we have an ongoing project on trying to calibrate the two sides of that dialogue, so that they actually see eye to eye on what it means for the radiologist or for large language model to say, likely present. [KARA] I mean, and you point to something I think really significant there, which is that people are people, and they're different. And when they write about something, they write differently.

12:32 And categorizing writing can be hard. It's not binary. It's not a number. It's hard to compare likely, and possibly, and maybe. And these are like fuzzy kind of gray scale words, right? [POLINA] That's right. And in many different areas, there have been attempts to constrain radiologists and physicians to more quantitative scales. And as far as I know, it hasn't been entirely successful in complex cases. So for example, detecting pathologies such as breast cancer has been quantified a long time ago, and the radiologists enter a score that represents possibility of malignant growth.

13:19 But in many areas that are complex, such as heart failure, these kind of attempts have not been successful. And I think it's because we need the nuance of natural language to express the nuance of the disease. [KARA] So you're you, and you're trying to figure out how you categorize neatly these messy words. How does somebody who's on the computational side deal with like these almost infinite variations in what somebody can say about a picture that they see? [POLINA] Well, that's the interesting part. Basically, the recent evolution of large language models has finally brought text into the same quantitative domain that we have been working to bring images in.

14:06 Images and texts contain a lot of nuance. And effectively, they are the hardest to work with because the information is spread out kind of across the bits. In images, it's pixels. In text, it's among different words. And so distilling information from images or text is a fascinating area. So my focus has been mostly on images. But I think the recent developments with the large language models suggest that we are getting close to being able to distill information from text equally efficiently. [KARA] Just a further question on that picture thing.

14:40 I wonder to what degree radiologists agree with each other. I know at one point, you had, I think, four radiologists looking at something like 150 images, and they had to say, how severe is this problem that they're seeing. How often do radiologists disagree? [POLINA] Quite often. So there are lots of studies that show that inter-physician agreement is in many different areas, I've seen numbers ranging from 60% to 80%. So if we say 3/4 or so. It was actually reflected in this study that we did of 150 images. We weren't trying to study disagreement.

15:22 We just wanted ground truth, golden labels on a small data set that we could use for testing our methods. And we discovered that there's quite a bit of disagreement. But what's interesting to me, again, is that-- well, so computer scientists love jumping at something like that and saying, well, if we could be consistent, if we could replace this with consistent judgment by a computer, it would be better. But if you look at nuance of disagreement, a different picture emerges. A lot of times, they disagree. Because we force them again to produce an answer in a way that is unnatural.

15:56 So they might disagree, for example, by one unit of our artificial scale of severity of the disease. So somebody might say that the particular image represents moderate amount of edema, which is fluid in the lungs. And somebody else might say, oh, no. This is mild. But really, it's because the image is borderline. And they are disagreeing because we force them to produce answers in these constrained kind of answers. But we saw quite a bit of that phenomenon. [KARA] And that kind of goes to what you were saying, that this exists on a spectrum rather than being in two buckets or three buckets in a very precise way.

16:35 [POLINA] Exactly. And so they're trying to express that nuance in writing about this because it turns out that people are not that great at putting things in buckets when they're continuous. [KARA] If you just back up for a minute. When you think about this project and other things you've done, I wonder what you've learned from working with doctors and working with the health care system that maybe ordinary lay people who don't have-- I mean, you are really connected to people. You talk with them. You collaborate with them.

17:07 What have you learned that ordinary people might not think about when they think about, transforming the medical system with AI, for example? [POLINA] I think, well, so one of them is my undying admiration for physicians. They're acting in such a limited amount of information. And for many of these disorders, we don't know the underlying mechanism of the disease. So effectively, medicine is forced into being this empirical science looking at very sparse information and making inferences and then recommending treatments.

17:43 So the ability of physicians to synthesize is unmatched. One thing that I learned that I think would be useful to think about is more about the health care system rather than the physicians themselves. And it is that for any algorithm, to make any kind of difference, it has to fit into the clinical workflow. So our model, before we go into the hospital, is that a patient comes in, a physician looks at them and prescribes medication or intervention, and they stay in a hospital, or they are released. But in reality, they might come in contact with many, many different practitioners, nurses, physicians different specialists.

18:27 And there's-- basically, treating a patient is like a choreography. And if the algorithm doesn't fit into the choreography and it forces people who participate in this choreography, to change substantially what they're doing without visible benefit to the patient, that will never get adopted. [KARA] Right. [POLINA] And so I think for computation to change health care, it has to fit into the current practice and then perhaps change it over time. But I think it would be a nonstarter to say, oh, I have a solution that will just solve it without understanding where in that process of treating the patient it would fit.

19:09 [KARA] And do you think that at the moment, there is a kind of mismatch there, where people either starting companies or with ideas on the computational side or funding companies, or this is so exciting. We can revolutionize medicine. And they do not fundamentally understand the constraints under which the medical system operates. [POLINA] For sure. But it's OK. So I have two kind of responses to that. One of them, this happens in consumer market all the time also, where companies get started based on an idea, and then you have to change what the product is going to be, so that people are actually interested in using the product.

19:51 So in health care, it's trickier because access is tricky. But there is definitely a learning process in every startup that starts. Wearing my other hat, I actually love working with startups as an advisor, to take some of these algorithms and put them in practice in real products that affect how health care operates, and I would say that each one of them had to go through two or three different business models. And oftentimes, when you say business model, you think about how you're going to make money. But to me, each one of them also substantially changed how the algorithm would fit into what the physicians are doing now.

20:32 [KARA] Has your model been implemented in any hospitals at this point? [POLINA] I hope not. [KARA] It's not ready yet. OK. [POLINA] First of all, it's not ready. Second, there is a huge amount of effort that is required to take an algorithm and make it a product. [KARA] OK. [POLINA] Hospitals are not set up to take research algorithms and run them. SorryÑ our collaborators have taken algorithms and ran them in experimental radiology setups, to give them a spin and to try them. But to fully deploy it, you need a company that will actually translate an algorithm into a product that fits into the ecosphere.

21:12 And I definitely see that happening. [KARA] Yeah, yeah. I was going to ask, is this be two years down the line, 10 years? What do you think? [POLINA] I definitely see very active set of startups right now, taking the algorithms that I and my colleagues have been developing and putting them in practice. It's just the question is 10 years from now, when you look back, which of these startups will have implemented them in the hospital system, and which ones will disappear? We don't know. [KARA] And when you think about what we were talking about in terms of looking at chest x-rays and the work you've been doing, I assume you hope that someday it will be like doctors will be using it.

21:51 Is that right? [POLINA] Or some variant of it, yes, absolutely. [KARA] So you do hope that it's sort of moving towards a product someday that will exist in hospitals. [POLINA] Absolutely. [KARA] OK. And do you think maybe-- in an optimistic scenario, would that be true in five years and 10 years, that it would really be used by doctors for people who, let's say, do come into the ER? [POLINA] So it's very interesting. I can tell you anecdotally what's happening. One of the startup companies that I was an advisor for, that I'm really proud of because basically, they were acquired by a very large radiology provider, and now it's deployed on every radiology workstations that that radiology provider is using.

22:32 So it's absolutely there. The kind of algorithms we are working on right now. Basically, I feel like it's our responsibility to work on things that when I describe them to people, they say, whoa, it's out there. And so for those, it will take longer. My colleagues in radiology at Children's Hospital, with whom I collaborate a lot, already have applications that measure, for example, limb length for kids who might need surgery to correct for discrepancy in limb length, running right now in radiology workstation. They did it as part of experimental setup again.

23:12 So it runs at Children's in Boston, but I don't think it's available anywhere else. So I think we also need to start thinking about what does diffusion of algorithms into the common practice looks like. So I can say some of our algorithms or of course products based on the algorithms are already in the hospitals, but it's single hospitals. For it to actually be commonly used, you need real deployment, which academics are not great at. You need a company to do that. [KARA] So it sounds like we're kind of at the beginning of the algorithms that you have created and that others have created in the same vein.

23:53 They're breaking through into actually being used. But we aren't at the maybe huge, widespread deployment. Does that sound right? [POLINA] That's right. We are in this very exciting and unpredictable transition phase. Think about how cars got created and introduced into the society. At the beginning, it was a fancy toy. And then sort of a bunch of rich people could afford them. And now everybody has them. We're nowhere near everybody having them in a health care. [KARA] So let me take you back to 2016. There was is a famous quote by Geoffrey Hinton, who won the Turing Award.

24:34 And this is the quote. We should stop training radiologists now. It's just completely obvious that within five years, deep learning is going to do better than radiologists. So not quite a decade ago, but we're getting close to a decade at this point. We, as you know, still have a lot of human radiologists, and the numbers are growing. What do you think it was that Hinton didn't anticipate? [POLINA] It's a great question. So I think many things. One of them is this mismatch. So we have now neural networks that are exceedingly accurate on very narrow, very specific problems.

25:18 But that doesn't mean that tomorrow, I can translate that performance into any specific clinical domain. If I move to a different clinical question or any other question, I have to start from scratch effectively, meaning collecting training data, annotating it, and retraining the network. So retraining the network in some sense is the easiest part. But getting access to sufficient data and annotations in that very narrow sense of getting the deep learning network to work better than the radiologist is tremendous amount of effort.

25:53 And I think it's hugely underappreciated how much harder it is to get any domain-specific data than it is to get images where the entire world is snapping pictures and uploading them to the internet. The other aspect is that I think, again, he has completely underestimated the challenges of introducing technology, so that it helps the physician rather than slows them down. And so the combination of those, it's not surprising to me that he was off the mark. [KARA] Right. But it isn't to say that radiologists aren't making use of AI now.

26:34 I mean, you mentioned a case in which it's happening at Children's. I know at other big and especially cutting edge hospitals, that is happening. Do you feel like that's widespread, that radiologists just across the board are able to make use of more AI tools, or that's more specialized? [POLINA] No, absolutely. And I think we are in this transition where we are in the exponential growth at the very flat part of the exponential, as in the numbers are very small, but the uptake is extremely fast. [KARA] Interesting.

27:09 [POLINA] So I suspect that in a few years, we will have explosion of tools that are useful to radiologists. The other interesting nuance that people don't appreciate but I've seen over and over again is that actually, in my field, there have been lots and lots of papers showing how the deep network outperformed either radiologist or surgeon, depending on the domain, outperformed the physician. And if you look carefully, what happens is that it outperformed in a very fair trial side by side comparison. The problem is that that setup was so unnatural for the radiologist that their performance plummeted.

27:49 And then the next step after that is when the network is deployed in the clinical setting and does much worse than it demonstrated in the trial. And it is because basically we're comparing it to a radiologist who was put in a very unnatural situation. It's taking excellent hockey players and asking them to play golf. [KARA] And when you talk about the deep network, is that as opposed to the stuff that you're doing? [POLINA] No. Many of these algorithms suffer from the same problem. [KARA] Yeah. [POLINA] So evaluating the algorithms in the very realistic scenarios that match how radiologists operate on the domain is crucial for these applications.

28:34 If you make radiologists being evaluated in a way that is natural for the neural network, the neural network will win. But you won't create a product that is useful for health. I wonder if the fact that there are still radiologists and not significantly fewer of them than there were a decade ago, [KARA] I wonder if that actually says anything to you about not just AI and medicine but AI in the workforce more broadly. And that when people think that AI is going to replace people, there's things they're not taking into account.

29:08 There's pieces of this that it's hard to do a one for one switch out. [POLINA] That's right. So at the danger of making similar mistakes to Geoff Hinton, I looked at other domains. And I think saying we shouldn't train radiologists because that will be replaced is like somebody looking at personal computer invention and spreadsheets and saying, we shouldn't train accountants. They will be useless. And yet we have accountants decades after spreadsheets were invented. Because the profession itself changed. This is the one thing that makes humans unique.

29:47 They will change to the circumstance. So right now, big part of what they counted is doing is not what they were doing 100 years ago when computers didn't exist. But the profession of accountants remains. And so I think similar transformations will happen everywhere. Radiologists will look very different in 50 years than they do now. They won't do mind-numbing measurements, but they will exist as a profession to understand how incoming information can help treat the patient better. So I think the professions will change.

30:21 [KARA] So let's end there, sort of back to what you're working on. If you project forward 10 years, what's the big picture. How do you think-- I mean, you spend so much time thinking about images and how we can make them more useful to doctors and obviously, in some ways, ultimately to patients. How do you think I might change our ability to look at these images in 10 years? [POLINA] In two ways. One is, again, providing lots and lots of background measurements that will be provided to physicians, sort of as the images are acquired, and they won't even notice that this is happening, and it will help them make diagnoses, treat the patients.

31:08 The other way, and I feel like this part is sort of already straightforward. There is a lot of research to be done. There is a lot of translational work to be done. But this will happen. But what I find more interesting is that, and this is what we are starting to work on. That's why nothing is working, and that's what we're working on right now, is to bring in computation into a way we acquire images. So right now, there is a pretty serious wall between acquiring the images and analyzing them. The MRI scanner acquires the image, the radiologist looks at it.

31:44 And so we're saying, oh, well, we'll add the computation to how radiologist analyzes the images. But what we are starting to do is actually intervene in how the imaging equipment, such as the MRI scanner or x-ray machine, acquire the images so that they acquire the most useful information. And I think this is the more exciting part because we don't know how to do that yet. And so for us, the beginning of that research is oftentimes much more exciting than when I can give talks on it, and it's sort of done. [KARA] Well, and it also, a person probably has limitations.

32:22 Like they can't look at the 25 last images that have been taken of this person's lungs. [POLINA] Exactly. [KARA] But a computer doesn't care. It can look back, and it doesn't take up much extra time to throw that in too as additional information. [POLINA] Exactly. So to shift gears. We've been focusing on chest x-ray and these collections of images. But to shift gears, a lot of my research is in MRI. And now, when the patient goes into an MRI, sometimes 1,000 2D images are acquired, together comprise dozens of 3D scans.

32:58 And we ask the radiologist now to look at all of them for one patient. So we're trying to change that by acquiring exactly the right images. So that either then an algorithm or a radiologist, if it's complex case, can look and synthesize based on a small number of images that are acquired to serve the purpose rather than hundreds of images, and this is what we could acquire. And we didn't think about what you're going to measure from these images when we acquired them. [KARA] Is there anything that we didn't talk about that you think would be surprising to listeners in this domain, that's probably obvious to you at this point but may not be to most people?

33:44 [POLINA] I don't know. Because I feel actually both excited and confused in the sense that it seems that introducing these tools, I hear lots and lots of media, conversations about how this will eliminate some problems, eliminate some jobs. Basically, it all sounds like it's closing the doors. Whereas to me, the development of this technology feels like it opened up so many interesting problems we can tackle that we couldn't hope to tackle before that. To me, it's just sort of like, what of the next exciting 10 problems that I couldn't imagine ever attempting we should work on in the next few years?

34:24 [KARA] That's great. That's actually a really interesting and a perspective you don't often hear I think or don't hear enough. [POLINA] I agree. I'm surprised by how everybody thinks that this is limiting. They think of AI developments as sort of again, chopping off some pathways or closing the doors for people. Whereas to me, it feels like it's opened up so many interesting questions that will spend decades working on. [KARA] And in some ways, that's why it's interesting to go back to the Hinton quote, because that prediction, while it was made by a brilliant person, didn't exactly pan out.

35:00 And so the question is like why didn't it pan out? And I think now, we can see that it's more complicated, as you said, than maybe was assumed a decade ago. [POLINA] That's right. [KARA] Polina Goland is a professor of Electrical Engineering and Computer Science at MIT. She's a principal investigator in CSAIL. Polina, thank you so much. Such an interesting conversation. [POLINA] Thank you. I enjoyed it.

35:31 [KARA] And before we go, if you enjoyed the show, we'd love for you to leave us a comment on Apple Podcasts. It helps other listeners discover us. We've got a lot more about CSAIL on our website. It's csail.mit.edu/podcast. You can find out how companies can access research, talent courses, and a whole lot more. Again, the website, csail.mit.edu/podcast. I'm Kara Miller. Our show is produced by Matt Purdy and Andrew Zucosky with help from Audrey Woods. Join us again next time and stay ahead of the curve.

Questions about this episode

Who speaks in "How AI is Reshaping Medical Imagery with MIT CSAIL Professor Polina Golland"?

Kara Miller (Host), Polina Golland (Guest).

What is "How AI is Reshaping Medical Imagery with MIT CSAIL Professor Polina Golland" about?

The discussion centers on how artificial intelligence is transforming the field of radiology and medical imaging, countering the initial fears that AI would replace radiologists. MIT professor Polina Golland highlights the evolving roles within the profession, the potential for AI-driven tools to enhance diagnostics, and parallels with other occupations that adapted rather than disappeared in…

What is mentioned in this MIT CSAIL Alliances episode?

7 people, works and ideas are mentioned, including Artificial Intelligence, Polina Goland, MIT, CSAIL, heart failure, MRI. Each one on this page links to its own page with every episode that mentions it.

How long is "How AI is Reshaping Medical Imagery with MIT CSAIL Professor Polina Golland"?

The episode runs 36 min; the transcript is about 5,647 words in 54 timestamped sections.

More from MIT CSAIL Alliances

• All analyzed episodes
• People mentioned on MIT CSAIL Alliances
• Topics mentioned on MIT CSAIL Alliances