Latest / Code & Cure / #60 - We Can Teach Medical AI To Look Like Experts

Transcript

#60 - We Can Teach Medical AI To Look Like Experts · Code & Cure

• 0:00
  SPEAKER_01: What if we could teach AI not just what to see, but how experts view it?
• 0:06
  SPEAKER_01: Researchers are using radiologist gaze to build a new kind of chest x-ray model, one designed to be more accurate, explainable, and trustworthy.
• 0:24
  SPEAKER_00: Hello and welcome back to Code and Cure, the podcast where we discuss decoding health in the age of AI.
• 0:30
  SPEAKER_00: My name is Vasant Sarathi.
• 0:32
  SPEAKER_00: I'm an AI researcher and cognitive scientist, and I'm here as always with Laura Hagopian.
• 0:37
  SPEAKER_01: I'm an emergency medicine physician, but I also do my fair share of reading some chest x-rays myself.
• 0:42
  SPEAKER_00: Yeah, this is a different kind of episode in the sense that the type of technology that they're developing is um, you know, we talk about human in the loop and things like that.
• 0:54
  SPEAKER_00: This is a different kind of human in the loop here, where we're trying to trying to track how radiologists think as they read x-rays.
• 1:01
  SPEAKER_01: Right.
• 1:01
  SPEAKER_01: Because it's not like you're viewing it in a vacuum.
• 1:04
  SPEAKER_01: Sometimes you have like clinical information alongside a chest x-ray.
• 1:08
  SPEAKER_01: Uh, you know, other times you may not, you may not.
• 1:12
  SPEAKER_01: But the way that you, the way that I view a chest x-ray is systematic.
• 1:17
  SPEAKER_01: So I know we've we've talked on prior episodes of like, oh, this is a COVID lung versus a not COVID lung, and like a cat got identified as a COVID lung.
• 1:24
  SPEAKER_01: Yeah.
• 1:25
  SPEAKER_01: But if we like understand, hey, how does a radiologist or how does a physician look at chest x-rays systematically, maybe you could train the AI on that.
• 1:34
  SPEAKER_01: And now you have like um clinical reasoning almost integrated in.
• 1:40
  SPEAKER_00: Yeah, yeah, yeah, yeah.
• 1:41
  SPEAKER_00: And there is a systematic process by which they're trained to read them, right?
• 1:44
  SPEAKER_01: Yeah.
• 1:45
  SPEAKER_01: I mean, I I use a very systematic process whenever I go through a chest x-ray.
• 1:49
  SPEAKER_01: It's like, you know, a mnemonic A B C D E F, where like the first thing I'm gonna look at A is the airways, like, hey, how does the trachea look?
• 1:57
  SPEAKER_01: How do the bronchi look?
• 1:58
  SPEAKER_01: Um, then I move on to breathing and I look at all the lung fields.
• 2:02
  SPEAKER_01: I make sure that there's lung markings going out to the periphery.
• 2:06
  SPEAKER_01: Um, circulation, I'm looking at the heart and also the aorta and how that comes around, what's the size of it?
• 2:13
  SPEAKER_01: Does it look appropriate?
• 2:14
  SPEAKER_01: I look at the diaphragms on both sides.
• 2:16
  SPEAKER_01: Um, are they too flattened?
• 2:18
  SPEAKER_01: Do I see effusions along them, et cetera?
• 2:20
  SPEAKER_01: And then external and foreign material, you know, is there um an endotracheal tube?
• 2:25
  SPEAKER_01: Is there a pick line, et cetera?
• 2:27
  SPEAKER_01: So you don't need to necessarily understand every single word I just threw out there, but the whole concept is like, hey, and and also I look at the bones too.
• 2:34
  SPEAKER_01: Like I'm gonna look at the clavicles, I'm gonna look at the ribs, et cetera.
• 2:37
  SPEAKER_01: But the whole idea is that like we systematically go through these things so that we don't miss stuff.
• 2:43
  SPEAKER_01: Right.
• 2:43
  SPEAKER_01: Right?
• 2:43
  SPEAKER_00: It's like a checklist.
• 2:45
  SPEAKER_01: Exactly.
• 2:45
  SPEAKER_01: And you go through it kind of similarly every time.
• 2:48
  SPEAKER_01: Now you might spend more time looking at certain things based on the indication of the chest x-ray.
• 2:53
  SPEAKER_01: Like if someone's short of breath and you see um, you know, fluid collections at the bottom of the lung fields and along the diaphragm, you might spend more time looking there.
• 3:02
  SPEAKER_01: If I got a chest x ray that's looking at, hey, was the line that I put in, was the central line that I put in, good.
• 3:09
  SPEAKER_01: Then I'm gonna spend more time looking at that foreign material, right?
• 3:11
  SPEAKER_01: So there's like an indication for a chest x-ray that may dictate how you where you spend more of your time, but you're still gonna systematically go through and look at everything there because what if something else is abnormal?
• 3:22
  SPEAKER_00: Right, right, right, right.
• 3:25
  SPEAKER_01: And so in prior episodes when we've talked about, you know, like chest x-rays being interpreted by by AI, it's like, well, we don't even know, we don't even know how they're doing it.
• 3:34
  SPEAKER_01: Like they don't, it's not that they do it back.
• 3:36
  SPEAKER_00: This is interesting, right?
• 3:37
  SPEAKER_00: Yeah.
• 3:37
  SPEAKER_00: Yeah.
• 3:37
  SPEAKER_00: I mean, I think what I find most interesting about this whole space is that this radiology was one of the early field in AI where they said, you know what, radiologists are not gonna be needed anymore.
• 3:49
  SPEAKER_00: We're just gonna have it all be automated.
• 3:51
  SPEAKER_00: Because I think they completely underplayed how difficult a task this is.
• 3:55
  SPEAKER_00: And, you know, when um in the mid-2000s, in the 2010s, was when really vision language models and vision, I shouldn't say vision language models, but vision models started getting really good.
• 4:06
  SPEAKER_00: They started looking at images and trying to figure out if there are, you know, certain characteristics, certain objects in the images.
• 4:12
  SPEAKER_00: They were able to classify images.
• 4:14
  SPEAKER_00: And that became a really powerful piece of technology back then.
• 4:19
  SPEAKER_00: And that's when, you know, kind of radiologist was like the first target, because in the eyes of a computer, it's similar to looking at an image and you're figuring out something about the image.
• 4:29
  SPEAKER_00: And so they thought that with the advanced AI technology back then, that they could basically do the job, what that the radiologists would be able to do and put them out of business.
• 4:39
  SPEAKER_00: That didn't happen.
• 4:40
  SPEAKER_00: No, that has not panned out.
• 4:42
  SPEAKER_00: They don't even close.
• 4:43
  SPEAKER_00: And part of the challenge is because um the way our vision system works and the way our reasoning works and the way our thinking and training works is that it's much more complex than just like looking at a bunch of pixels and then deciding based just on that, right?
• 4:55
  SPEAKER_00: There's like like we just talked about, there's a reasoning process by which one looks at these things.
• 5:00
  SPEAKER_00: It's not just random pixels there.
• 5:01
  SPEAKER_00: It's not the it's not the same as a picture of a cat, right?
• 5:04
  SPEAKER_00: It's totally different.
• 5:05
  SPEAKER_00: It's a different set of skills.
• 5:07
  SPEAKER_00: And I think people started to realize that and started to appreciate the value that, you know, training and all of that comes in and the human look comes in.
• 5:14
  SPEAKER_00: But also there's another piece to this, which is if you just threw an image to a system that detects whether or not something is present, you don't get any explainability.
• 5:23
  SPEAKER_00: You don't have any sense for how it arrived at that conclusion.
• 5:25
  SPEAKER_00: And oftentimes the way it arrived at that conclusion is because it detected similar patterns in its in its dating, uh, it trained on similar patterns in its training data.
• 5:33
  SPEAKER_00: So it concludes that, you know, um, that that particular X-ray has these particular features.
• 5:39
  SPEAKER_00: I mean, this is goes back to the the joke that you made earlier and the comment that we've made before about the uh famous case of the X-ray versus cat, where they trained a model on a bunch of COVID lungs, X-rays, and it was able to tell if a lung you know was infected with COVID or not, and with a certain high degree of accuracy, and then they gave it a picture of a cat, and it said with 100% confidence that this was a COVID lung.
• 6:03
  SPEAKER_00: Now, it's kind of amusing.
• 6:05
  SPEAKER_00: Um, it's it's a little unfair to the AI researchers in that setting, being an AI researcher myself.
• 6:11
  SPEAKER_00: I would say it's a little unfair because that's not the training data, right?
• 6:15
  SPEAKER_00: You can't just give it a picture of a cat and say, is it anything?
• 6:17
  SPEAKER_00: Because it doesn't know what a cat is.
• 6:18
  SPEAKER_00: It's trained on just lungs.
• 6:20
  SPEAKER_00: So as far as it's concerned, it's only looking at the pixels and saying, okay, these pixels have certain properties.
• 6:26
  SPEAKER_00: It looks like there's a pattern here in the picture of the cat that for some reason aligns with the numbers in the pixels of the x-rays.
• 6:34
  SPEAKER_00: And so, you know, it concluded that.
• 6:36
  SPEAKER_00: But the point is that that was an what they call out of distribution.
• 6:39
  SPEAKER_00: That is a picture that was not anywhere there.
• 6:42
  SPEAKER_00: But a human looking at it, even a human child looking at it knows that it's different, right?
• 6:45
  SPEAKER_00: They're not the same domain at all.
• 6:48
  SPEAKER_00: And therefore, they shouldn't be confused at all.
• 6:50
  SPEAKER_00: And so, anyways, this is kind of the historical backing here that kind of prompted us to say, you know what, we need to do better with um images and we need to do better, especially in these high stakes and highly specialized settings where background knowledge comes into play.
• 7:04
  SPEAKER_01: And that's it, right?
• 7:05
  SPEAKER_01: Your eye gaze trajectory when a radiologist is looking at an image is following like a clinical reasoning framework.
• 7:14
  SPEAKER_01: They're not just like looking at it in a vacuum, right?
• 7:16
  SPEAKER_01: They're actually reasoning and they're thinking through things clinically as they do it.
• 7:20
  SPEAKER_01: And so, can we capture some of that reasoning with their gaze?
• 7:26
  SPEAKER_01: And it's not just where they look, which is important, right?
• 7:30
  SPEAKER_01: Um, but also like the temporal sequence, like how long are they spending in each spot and why?
• 7:36
  SPEAKER_01: And that may change based on the abnormalities in the x-ray.
• 7:40
  SPEAKER_01: That's right.
• 7:41
  SPEAKER_01: And that's interesting too, right?
• 7:43
  SPEAKER_01: Because it's not like it's binary.
• 7:46
  SPEAKER_01: It's not like it's normal x-ray versus abnormal x-ray.
• 7:48
  SPEAKER_01: That's not super helpful.
• 7:49
  SPEAKER_01: It's like what's abnormal in the x-ray?
• 7:51
  SPEAKER_01: We want to know that.
• 7:52
  SPEAKER_01: And a radiologist may spend more time looking at the heart in someone who has an enlarged heart.
• 7:58
  SPEAKER_01: They may spend uh more time looking at the lungs in someone who has a pneumonia.
• 8:03
  SPEAKER_01: They may spend more time looking at the periphery of the image in someone who has a support device.
• 8:09
  SPEAKER_01: And the time sequencing of that may vary also.
• 8:12
  SPEAKER_00: Right, right, exactly.
• 8:14
  SPEAKER_00: And that's kind of where the technology comes in, uh, which I think is fascinating.
• 8:19
  SPEAKER_00: So, you know, we can talk about that if um, you know, if that's of interest to people.
• 8:23
  SPEAKER_01: I'm definitely interested in this because I mean, what I saw in this paper is they broke the chest x-ray up into like a grid, um, like of like a four by four grid.
• 8:33
  SPEAKER_01: And they like basically tried to figure out how long are radiologists looking at different spots on the grid.
• 8:39
  SPEAKER_01: But I don't understand how that gets translated into um into like the end output or the end product.
• 8:46
  SPEAKER_01: So please, please explain to me.
• 8:48
  SPEAKER_00: Yeah, so let's um start at the beginning.
• 8:50
  SPEAKER_00: So people know about language, large language models.
• 8:52
  SPEAKER_00: You send in a bunch of text and it predicts the next piece of text, right?
• 8:56
  SPEAKER_00: Um, that those are transformer models and they're designed for predicting sequences.
• 9:01
  SPEAKER_00: Um, you also had what's called convolutional neural networks, which take images and predict objects in them, um, tell you like a classification if this is a cat, dog, whatever.
• 9:11
  SPEAKER_00: Um, and those are not sequence predictors.
• 9:14
  SPEAKER_00: So, what they used here, I mean, typically for a lot of these image processing applications, they use uh like those kinds of convolutional neural networks that don't predict sequences.
• 9:22
  SPEAKER_00: But here, they, because it's a sequence of actions that you would do for reasoning, yeah, they took into they took that into account and they chose a what's called a vision transformer or a vision language model or a VLM.
• 9:35
  SPEAKER_00: A VLM is kind of like an LLM, except that you can provide it not just text, but you can provide it pixels or image data.
• 9:43
  SPEAKER_00: So in sequence.
• 9:44
  SPEAKER_00: And sometimes the sequence is provided literally like a like a like you know from top, top left to bottom right, kind of line by line, pixels, uh rows of pixels, um, are provided sometimes that way.
• 9:54
  SPEAKER_00: There are other ways to provide the information, but the idea is that you get these vision language models in which you can provide um images and text, and the output is text.
• 10:05
  SPEAKER_00: So it again, it predicts some set of sequence.
• 10:08
  SPEAKER_00: So the idea being, hey, if I can give it the x-ray, the output could be the report.
• 10:13
  SPEAKER_01: Yeah, like, oh, this shows, you know, an enlarged heart or this shows um, you know, fluid collection at the bases of the lungs, whatever it is, it's gonna give you that report, which then the next clinician like me could go on and use and say, hey, yes, you have an enlarged heart.
• 10:28
  SPEAKER_01: This is what I want to do with you next.
• 10:29
  SPEAKER_01: This isn't normal.
• 10:30
  SPEAKER_01: Okay.
• 10:31
  SPEAKER_00: And the idea is that these language, these transformers, um, because they're trained on trillions and trillions of images and um, you know, deck text from the internet, they have some degree of reasoning skills to be able to reason through that.
• 10:43
  SPEAKER_00: So that's text from the internet.
• 10:45
  SPEAKER_01: Well, it must be text from like some of this must be text from radiologists reading x-rays, though, right?
• 10:49
  SPEAKER_00: Yes, we'll get there in a second.
• 10:50
  SPEAKER_01: So I was like, you're making me nervous.
• 10:52
  SPEAKER_00: No, no, no.
• 10:52
  SPEAKER_00: So I'm just sort of setting the stage.
• 10:54
  SPEAKER_00: This is not yet what they have done.
• 10:55
  SPEAKER_00: This is just the world of vision language models.
• 10:58
  SPEAKER_00: Okay, jumped ahead.
• 10:59
  SPEAKER_00: Yes.
• 10:59
  SPEAKER_00: So, like, you know, the first simple step would be just, hey, let me just give it a bunch of x-rays and have it generate reports.
• 11:05
  SPEAKER_00: Obviously, that is not satisfactory.
• 11:06
  SPEAKER_00: It has not taken into account any kind of reasoning process um intentionally.
• 11:10
  SPEAKER_00: Uh, and so the what the what these um authors did in this paper, uh, which is wonderful, is actually start to record eye gaze and um sort of um yeah, eye gaze off the radiologists and track what they were talking about as they read the x-rays.
• 11:27
  SPEAKER_00: So they created a brand new data set in which they had radiologists, trained radiologists, looking at x-rays and then talking about them.
• 11:36
  SPEAKER_00: And while they did that, they tracked their eye gaze.
• 11:38
  SPEAKER_00: And eye gaze is just literally, if you think about it, is just where your eyes are looking on the picture.
• 11:42
  SPEAKER_00: Um in that grid, I was saying.
• 11:44
  SPEAKER_01: In that like four by four grid.
• 11:45
  SPEAKER_00: Yeah, I wouldn't even think about the four by four grid for a second.
• 11:48
  SPEAKER_00: Um, I would just look at the whole whole image, the whole image that the radiologist looks at, right?
• 11:53
  SPEAKER_00: Okay.
• 11:53
  SPEAKER_00: They're looking at the whole X-ray, and um, and they're uh we're tracking their eye gaze.
• 11:58
  SPEAKER_00: And every time they say, and maybe we're interested in certain phenomena, right?
• 12:03
  SPEAKER_00: Just for for argument's sake, let's say that certain types of things is what they're looking at.
• 12:06
  SPEAKER_01: Yeah, let's say this this specific x-ray we're looking at has an enlarged heart on it.
• 12:10
  SPEAKER_01: Okay, the radiologist might focus on.
• 12:11
  SPEAKER_00: At some point, might say the words enlarged heart, right?
• 12:14
  SPEAKER_01: Cardiomagalia, actually.
• 12:15
  SPEAKER_01: It's the medical term for it.
• 12:18
  SPEAKER_00: Yeah.
• 12:18
  SPEAKER_00: So the radiologist would say that.
• 12:20
  SPEAKER_00: And so what they also know uh from past cognitive science studies and other studies, that um typically your eye gaze um precedes that by about two and a half seconds.
• 12:30
  SPEAKER_00: So what they did was they look timed when the radiologist said that.
• 12:35
  SPEAKER_00: And then about two and a half seconds before, they kind of isolated where the radiologist was looking, with the point where uh they were looking at on the X-ray.
• 12:44
  SPEAKER_00: Now, it's not really a point, it's kind of like a fuzzy cloud around that point because you have a field of view in your eyesight, so there's like a region that they look at.
• 12:53
  SPEAKER_00: So by doing this, they were able to track in sequence a bunch of regions.
• 12:57
  SPEAKER_00: Now, what's really cool also is what you mentioned earlier, which is that radiologists might stare at a particular thing for longer periods of time.
• 13:05
  SPEAKER_00: And that means that something has either caught their attention or something that needs further thinking, further analysis.
• 13:10
  SPEAKER_00: And so we're able to track that too, because the eye gaze is kind of like throwing darts kind of at the at the at the um X-ray.
• 13:18
  SPEAKER_00: So if you stare at an point for longer, it just means that there's more points over there.
• 13:21
  SPEAKER_00: And the intensity of the eye gaze values goes up there.
• 13:25
  SPEAKER_00: So they're able to track not just the location and the timing of that, but also how intensely the person looked at that place.
• 13:32
  SPEAKER_00: So now imagine you have the radiologist is finished, they're looking through the whole X-ray and has and identified, I want to say, 10 things, right?
• 13:40
  SPEAKER_00: And those 10 things have 10 clouds of points that have been tracked in sequence throughout the X-ray, regions that we care about in the X-ray.
• 13:49
  SPEAKER_00: So then what they did was um they put the X-ray in.
• 13:53
  SPEAKER_00: So they trained the model in multiple different ways.
• 13:55
  SPEAKER_00: And this is called multitask training.
• 13:57
  SPEAKER_00: So you already have that vision language model to start with, the one that's been trained in all the internet data, right?
• 14:02
  SPEAKER_00: But it doesn't have all the specialist information.
• 14:04
  SPEAKER_00: So what you do is you fine-tune it.
• 14:07
  SPEAKER_00: You say, I care about this application, and I'm going to fine-tune it.
• 14:10
  SPEAKER_00: So what they did was they took this radiologist measures and all of this stuff, and they created lots and lots of data points, and then they ran it through and trained the transformer to do different things.
• 14:21
  SPEAKER_00: Like, for example, one thing they had it do was if I gave it a shuffle sequence of um things that they looked at, so let's say the radiologist looked at region one, then region four, and then region three.
• 14:32
  SPEAKER_00: Um, I would train it if I gave it four one three instead, that it would give me the right sequence output.
• 14:38
  SPEAKER_00: So then first thing I trained was to say, hey, if I give you a shuffle sequence, give me the right order.
• 14:44
  SPEAKER_01: That a radiologist would be looking at it.
• 14:45
  SPEAKER_00: Would be looking at, not the sequence.
• 14:46
  SPEAKER_00: So first thing they trained it on it's not when I say first thing, they kind of trained all of these together at the same time.
• 14:51
  SPEAKER_00: But just think of them as separate um tasks that this thing was fine-tuned on.
• 14:56
  SPEAKER_00: So that was one of them, was shuffling it.
• 14:58
  SPEAKER_00: Another one was given the original x-ray, and given the uh all the intensities in sequence, again, the the original x-ray and intensities in sequence and the words, generate the report.
• 15:10
  SPEAKER_00: So that's that's another uh task that was simultaneously trained on.
• 15:14
  SPEAKER_00: Um it was also trained on other tasks that um have to do with uh um when the next item would be, where the next kind of uh step would be in the reasoning process.
• 15:26
  SPEAKER_00: So by doing so, what what they did was essentially have uh take the original language model and then make it do a bunch of specific things, teach it specific skills that then they would do simultaneously.
• 15:37
  SPEAKER_00: And now you have a new language model that they believe incorporates that reasoning process as well.
• 15:43
  SPEAKER_00: And so now you don't just have the original one that's trained in the internet text, but now you have stuff that's actually uh modified or fine-tuned uh to incorporate this all of this process.
• 15:53
  SPEAKER_01: And the idea is that it would be smarter, right, than having something that wasn't fine-tuned on how a radiologist looks at things.
• 16:00
  SPEAKER_01: I mean, I think some of the original articles around chess X-ray interpretation, uh, some of the models did a good job, but there wasn't like a reasoning, but we couldn't understand like exactly what it was looking at and why.
• 16:12
  SPEAKER_01: And the whole idea here is to get that interpretability and understanding.
• 16:16
  SPEAKER_00: I do want to caution though, because at the ultimately the language model is a bunch of knobs, right?
• 16:21
  SPEAKER_00: Billions and billions of knobs that were originally set to some setting from the internet uh training, but then now, based on these additional pieces of training, they're being reset slightly, modified, fine-tuned slightly, right?
• 16:33
  SPEAKER_00: But at the end of it, it's still a a machine that takes in a bunch of information, uh pattern matches, and then outputs something, right?
• 16:42
  SPEAKER_00: There is no guarantee that it's actually using that specific reasoning process that you had it use.
• 16:48
  SPEAKER_00: Right.
• 16:49
  SPEAKER_00: You hope that it's learned enough that the patterns it understands is going to have it apply that, but we're not forcing it to say, hey, do this first, then do this first, and then do this first.
• 17:01
  SPEAKER_00: We're just saying here is a bunch of hundreds of people doing it in that order.
• 17:06
  SPEAKER_00: Please just kind of like try to get the pattern and remember this, please, uh, pretty please, and then run it through the system.
• 17:13
  SPEAKER_00: So, yes, it's very exciting that they have that data from the radiologists.
• 17:18
  SPEAKER_00: Um, and yes, that it probably is much better than just the original language vision language model, but I want to caution that it's still, at the end of the day, a neural network that's kind of just like in isolation.
• 17:28
  SPEAKER_00: So it's not may not have um that reasoning process uh forced into its system.
• 17:35
  SPEAKER_01: It's interesting because I have some drawbacks from like the clinical angle too, um, or some hesitations, I guess I would say.
• 17:41
  SPEAKER_01: Well, chest x ray is like one of the easiest things, so that's what they use.
• 17:45
  SPEAKER_01: But think about going into um things that have multiple images, CAT scans, MRIs, like lots and lots of images.
• 17:52
  SPEAKER_01: It makes it all all the much harder.
• 17:54
  SPEAKER_01: But even for a chest x-ray, like what I routinely do is if, especially if there's an abnormality, is I'm looking at someone's old x-ray and comparing it.
• 18:01
  SPEAKER_01: And that wasn't that wasn't done in the study, and I do that pretty much every time.
• 18:05
  SPEAKER_01: And I think most radiologists do too.
• 18:07
  SPEAKER_01: And the other thing is that chest x rays are done for the most part as two views.
• 18:12
  SPEAKER_01: I order two view chest x-ray.
• 18:14
  SPEAKER_01: One goes like back to front and the other goes sideways.
• 18:18
  SPEAKER_01: So there's a reason we do that.
• 18:19
  SPEAKER_01: It's not like we order a chest x ray and then only look at one view of the chest x-ray, right?
• 18:24
  SPEAKER_01: Which they only looked at one view in this study.
• 18:26
  SPEAKER_01: They did not look at the lateral view or the sideways view of people.
• 18:30
  SPEAKER_01: Um, and you can glean a lot from that.
• 18:32
  SPEAKER_01: Like, especially one of the ones that sticks out in my mind is uh a retrocardiac opacity, which is basically like, is there probably a pneumonia behind the heart?
• 18:42
  SPEAKER_01: Yes.
• 18:42
  SPEAKER_01: Because the heart would block it on the view that goes back to front.
• 18:46
  SPEAKER_01: Yes.
• 18:47
  SPEAKER_01: But on the view that goes sideways, you would be able to clearly see it.
• 18:50
  SPEAKER_01: And that was not included here.
• 18:52
  SPEAKER_01: So it's just it, there, there's limitations.
• 18:54
  SPEAKER_01: And I think it's important to call those out.
• 18:57
  SPEAKER_01: Um, and this is a starting point, right?
• 19:00
  SPEAKER_01: So these are things that could be worked on.
• 19:02
  SPEAKER_01: But if I'm thinking through, hey, what do I actually do when I look at a chest x-ray?
• 19:06
  SPEAKER_01: I I do compare it to the old and I do look at both views of the chest x-ray.
• 19:10
  SPEAKER_00: And well, that that's interesting because it's if you think about the reasoning process here, it is not just a matter of a wrote sequential reasoning process.
• 19:18
  SPEAKER_00: There is a reason behind that reasoning process.
• 19:20
  SPEAKER_00: And when you think about the views front and side, you're already thinking about the model of the human body.
• 19:26
  SPEAKER_00: And you know anatomically what needs to be there.
• 19:29
  SPEAKER_00: You know what, I mean, there's so much more medical knowledge that's um abstract, the abstractions of which you have in your head that you are utilizing.
• 19:38
  SPEAKER_00: And, you know, if you think about it this way, like it's not just a picture to you, right?
• 19:43
  SPEAKER_00: It's a person and their bodily functions and their uh, you know, their anatomy that is what's the mental model in your head as you evaluate this.
• 19:55
  SPEAKER_00: That is the reasoning process.
• 19:56
  SPEAKER_00: Right.
• 19:56
  SPEAKER_00: So again, that is not what neural networks do.
• 19:59
  SPEAKER_00: At least we don't think that they are taking into account a world model or a model of this, of this domain.
• 20:05
  SPEAKER_00: In this case, it's the human body, but it could be anything, right?
• 20:07
  SPEAKER_00: You could have any kind of scientific discipline that has, you know, physics, has all these physics models, right?
• 20:13
  SPEAKER_00: Um, and you don't know that it's taking into account all of these aspects that we understand about the world when we're reasoning about this.
• 20:20
  SPEAKER_01: Right.
• 20:21
  SPEAKER_01: And the reasoning involves like more than just looking at that one view.
• 20:24
  SPEAKER_01: It's like if I'm seeing someone at urgent care, for example, and I do a chest x-ray because I'm worried about an ammonia, I'm gonna look at, I'm gonna look at the side view and I'm gonna look at their last x-ray and see if there's a change from the current x-ray.
• 20:37
  SPEAKER_01: Like there's there's multiple steps that go into it.
• 20:39
  SPEAKER_01: Not to mention the fact that I've, you know, talked to the patient.
• 20:42
  SPEAKER_01: I've heard that they have a, you know, a fever, that they've had a cough, they're coughing up green stuff, and I've listened to their lungs and I hear something funky in the right lower lobe, right?
• 20:51
  SPEAKER_01: So there's all this other stuff that goes into here's what the decision I'm gonna make is for this patient.
• 20:57
  SPEAKER_01: And it and is not just a single of you of a chest x-ray.
• 21:01
  SPEAKER_00: Yeah, and you know, I I'm an AI reasoning researcher, and to me, this is the most fascinating thing about human reasoning is the fact that we work with these abstractions in our head about how the world works or how, you know, if we are trained in the medical space, then how the body works.
• 21:14
  SPEAKER_00: And those are the abstractions that allow us to be super flexible in the way we think about the world, right?
• 21:19
  SPEAKER_00: In the way we think about these problems.
• 21:20
  SPEAKER_00: We're not just looking at these in isolation, we're thinking about the whole system and how the whole system needs to work.
• 21:25
  SPEAKER_00: And anything that's weird is going to be weird in the context of our model of how that system works.
• 21:30
  SPEAKER_00: So I it's pretty cool.
• 21:31
  SPEAKER_00: I think that they are headed in the right direction, uh, incorporating human reasoning.
• 21:35
  SPEAKER_00: I think we have a long way to go still.
• 21:37
  SPEAKER_00: I mean, and radiologists are, you know, still that that the that original comment that people may used to make back, you know, 15, 20 years ago is ridiculous if you think about it, right?
• 21:46
  SPEAKER_00: I mean Yeah, their job is vital.
• 21:48
  SPEAKER_00: Yeah, absolutely vital.
• 21:50
  SPEAKER_00: Yeah, yeah.
• 21:50
  SPEAKER_01: And I think this is one step in the direction of figuring out how they reason and sort of trying to train models on that.
• 21:58
  SPEAKER_01: But that, you know, as you've pointed out, doesn't give full explainability.
• 22:02
  SPEAKER_00: That's right.
• 22:03
  SPEAKER_00: That's right.
• 22:04
  SPEAKER_01: All right.
• 22:05
  SPEAKER_01: Well, um, thank you for giving us a little more background on these vision models.
• 22:09
  SPEAKER_01: This is interesting.
• 22:10
  SPEAKER_00: Yeah, yeah, yeah, yeah.
• 22:11
  SPEAKER_00: And this was a fun episode.
• 22:12
  SPEAKER_00: Thank you for joining us.
• 22:12
  SPEAKER_01: We'll see you next time on Code and Cure.