1 00:00:15,356 --> 00:00:15,796 Speaker 1: Pushkin. 2 00:00:20,276 --> 00:00:23,916 Speaker 2: Ten years ago, Jeffrey Hinton, the computer scientist who's one 3 00:00:23,956 --> 00:00:26,676 Speaker 2: of the founders of modern AI, came out with a 4 00:00:26,796 --> 00:00:29,716 Speaker 2: very strong take. He said, and this is a quote, 5 00:00:30,196 --> 00:00:36,196 Speaker 2: people should stop training radiologists. Now. Radiologists, as you probably know, 6 00:00:36,556 --> 00:00:40,756 Speaker 2: are doctors who, among other things, read medical scans, X rays, 7 00:00:40,836 --> 00:00:45,116 Speaker 2: CT scans, MRIs and Hinton said it was just completely 8 00:00:45,156 --> 00:00:49,596 Speaker 2: obvious again his words, completely obvious that within five years 9 00:00:49,636 --> 00:00:52,676 Speaker 2: AI would do a better job than radiologists at doing that, 10 00:00:52,716 --> 00:00:57,276 Speaker 2: at reading scans. And this makes sense. Radiology is largely 11 00:00:57,356 --> 00:01:00,436 Speaker 2: pattern matching. Does this scan match the pattern of scans 12 00:01:00,436 --> 00:01:02,916 Speaker 2: that show disease or the pattern of scans that show 13 00:01:02,956 --> 00:01:07,116 Speaker 2: a healthy patient? And pattern matching is basically what AI does. 14 00:01:07,676 --> 00:01:10,956 Speaker 2: And yet here we are ten years later with amazing AI, 15 00:01:11,716 --> 00:01:16,236 Speaker 2: and we still need radiologists. In fact, there's a radiologist shortage. 16 00:01:16,636 --> 00:01:20,156 Speaker 2: So the question is what happened? Why do we still 17 00:01:20,196 --> 00:01:23,796 Speaker 2: need so many radiologists, and more broadly, what is the 18 00:01:23,836 --> 00:01:27,196 Speaker 2: story of AI and radiology tell us about the future 19 00:01:27,236 --> 00:01:29,916 Speaker 2: of AI and work for the rest of us, those 20 00:01:29,916 --> 00:01:44,356 Speaker 2: of us who are not radiologists. I'm Jacob Goldstein, and 21 00:01:44,356 --> 00:01:47,276 Speaker 2: this is what's your problem, My guest, today is Soreb Jaw. 22 00:01:47,556 --> 00:01:50,796 Speaker 2: He's a radiologist at the University of Pennsylvania. He's written 23 00:01:50,796 --> 00:01:53,156 Speaker 2: about his own use of AI, and he's also traveled 24 00:01:53,156 --> 00:01:56,396 Speaker 2: to India and Nepal and has seen the very different 25 00:01:56,476 --> 00:01:59,436 Speaker 2: ways that AI is being used for radiology in the 26 00:01:59,476 --> 00:02:02,796 Speaker 2: developing world. To start, Soreb and I talked about why 27 00:02:02,916 --> 00:02:07,436 Speaker 2: radiologists have adopted AI more slowly than some people expected 28 00:02:07,476 --> 00:02:10,676 Speaker 2: they would. What was it like for you using AI 29 00:02:11,116 --> 00:02:13,316 Speaker 2: five years ago? Like when you tried to use AI, 30 00:02:13,836 --> 00:02:15,596 Speaker 2: you know, specifically sort of what happened? 31 00:02:16,516 --> 00:02:19,156 Speaker 1: So to give you an example, when I was using 32 00:02:19,196 --> 00:02:23,956 Speaker 1: AI for a chest sextuory, I found that it slowed 33 00:02:23,996 --> 00:02:28,876 Speaker 1: me down quite honestly, because what it would do is 34 00:02:28,996 --> 00:02:33,916 Speaker 1: it would pick up stuff and highlight it. And now 35 00:02:33,996 --> 00:02:38,076 Speaker 1: I had to do two things. Firstly make my own 36 00:02:38,116 --> 00:02:41,756 Speaker 1: assessment and then make the assessment with it. AI was 37 00:02:41,876 --> 00:02:45,636 Speaker 1: right wrong, and whereas before I didn't have to do that. 38 00:02:46,076 --> 00:02:49,596 Speaker 1: So AI was kind of like that irritating medical student. 39 00:02:53,036 --> 00:02:55,276 Speaker 1: I didn't mean to say that. They're all very nice. 40 00:02:56,636 --> 00:02:59,476 Speaker 2: That go on, like what was it like? We'd be like, hey, 41 00:02:59,476 --> 00:03:01,276 Speaker 2: what about this? Should we worry about this? 42 00:03:01,516 --> 00:03:04,916 Speaker 1: Yeah? And for most parts it wasn't really finding things 43 00:03:04,956 --> 00:03:07,836 Speaker 1: that I wasn't finding, and it was finding things that 44 00:03:07,916 --> 00:03:10,996 Speaker 1: weren't real. It increased my cognitive burden. 45 00:03:12,116 --> 00:03:15,596 Speaker 2: There is this interesting story from a kind of before 46 00:03:15,716 --> 00:03:22,036 Speaker 2: the modern era of AI. There was computer assisted mimmography, right, 47 00:03:22,076 --> 00:03:24,996 Speaker 2: there was There was an earlier version of this, and 48 00:03:25,036 --> 00:03:28,636 Speaker 2: as I understand it, in that earlier version, in fact, 49 00:03:28,756 --> 00:03:32,236 Speaker 2: it was unhelpful. Right. It led to more false positives. 50 00:03:32,276 --> 00:03:36,236 Speaker 2: Women were getting more biopsies, more follow up tests because 51 00:03:36,236 --> 00:03:38,196 Speaker 2: they maybe had cancer and they didn't, and in the 52 00:03:38,316 --> 00:03:42,876 Speaker 2: end it was not improving outcomes. Right. So I understand 53 00:03:42,876 --> 00:03:44,876 Speaker 2: that that's an old story, but like that points to 54 00:03:44,956 --> 00:03:47,796 Speaker 2: the risk of AI. It's not always good to have 55 00:03:47,916 --> 00:03:50,116 Speaker 2: more flags raised. 56 00:03:50,796 --> 00:03:53,556 Speaker 1: Yeah, I'm glad you raised that because that does address 57 00:03:53,676 --> 00:03:57,316 Speaker 1: two very important things. You know, we think about physicians 58 00:03:57,356 --> 00:04:03,076 Speaker 1: as sort of people that spot diseases and as somebody 59 00:04:03,156 --> 00:04:07,276 Speaker 1: that I trained a long time ago, but not that 60 00:04:07,356 --> 00:04:14,116 Speaker 1: long time ago. The easy bit was finding disease. The 61 00:04:14,196 --> 00:04:16,996 Speaker 1: easy bit was actually finding people that were very ill. 62 00:04:17,676 --> 00:04:21,476 Speaker 1: The difficult bit was telling somebody that you're not ill, 63 00:04:22,596 --> 00:04:27,956 Speaker 1: that you're actually normal. So appreciating what normal is appreciating 64 00:04:27,996 --> 00:04:30,676 Speaker 1: that you don't need to come into the hospital. You 65 00:04:30,756 --> 00:04:35,516 Speaker 1: don't need to be investigated left, right, and center. You 66 00:04:35,516 --> 00:04:39,196 Speaker 1: don't need to be on a bunch of medication. Is 67 00:04:39,236 --> 00:04:41,156 Speaker 1: probably the most challenging task. 68 00:04:41,436 --> 00:04:43,796 Speaker 2: I say more about that. Why do you say that. 69 00:04:44,516 --> 00:04:47,636 Speaker 1: Because when disease looks at you, it stares at you 70 00:04:47,676 --> 00:04:52,636 Speaker 1: straight in the face. Yea, But for many reasons, when 71 00:04:52,676 --> 00:04:56,316 Speaker 1: you're at that border zone between disease and not disease, 72 00:04:57,396 --> 00:05:01,756 Speaker 1: it's very fuzzy. It gets very foggy, and at that 73 00:05:01,876 --> 00:05:05,316 Speaker 1: point that you have to really instill all your expertise 74 00:05:05,476 --> 00:05:08,956 Speaker 1: to make sure that it's not that you're not overcalling. 75 00:05:09,676 --> 00:05:12,196 Speaker 1: So the false positive part is a very big part. 76 00:05:12,956 --> 00:05:18,396 Speaker 1: And what happened in the past was that the mammography, 77 00:05:19,236 --> 00:05:24,996 Speaker 1: the computer aided detection of findings was fine tune. You noticed, 78 00:05:25,036 --> 00:05:28,996 Speaker 1: like a thermostat, you can find tunit to some accent 79 00:05:29,036 --> 00:05:31,396 Speaker 1: the way you want. And they said, we can't miss 80 00:05:31,436 --> 00:05:35,396 Speaker 1: a single case of cancer. Can't make that mistake. It's 81 00:05:35,396 --> 00:05:37,836 Speaker 1: okay to overcall cancer, but we can't have a false 82 00:05:37,876 --> 00:05:42,156 Speaker 1: negative because that would be disastrous. And so that the 83 00:05:42,396 --> 00:05:47,716 Speaker 1: radiologist who had the computer aided detection, now it became 84 00:05:47,796 --> 00:05:50,956 Speaker 1: like a monkey on the shoulder. It kept saying positive 85 00:05:50,996 --> 00:05:53,636 Speaker 1: positive positive you know, you know the story of the 86 00:05:53,636 --> 00:05:56,956 Speaker 1: boy who cried Wolf. Right, So when there truly was 87 00:05:56,996 --> 00:06:00,756 Speaker 1: a cancer, people were like, yeah, are you are you 88 00:06:00,956 --> 00:06:04,396 Speaker 1: just being AI? Or is this truly cancer? So it's 89 00:06:04,436 --> 00:06:07,356 Speaker 1: not very helpful if it has a lot of false positives. 90 00:06:07,916 --> 00:06:10,756 Speaker 2: Right, And in fact, there was a study published in 91 00:06:10,796 --> 00:06:12,556 Speaker 2: the New England Turnal of Medicine. Right, this is not 92 00:06:12,636 --> 00:06:15,876 Speaker 2: just anecdotal, like that early system did not help, Like 93 00:06:15,956 --> 00:06:20,196 Speaker 2: women had more biopsies, more fear of having cancer, but 94 00:06:20,276 --> 00:06:22,956 Speaker 2: they didn't. They were not healthier, Like the computer was 95 00:06:22,996 --> 00:06:25,676 Speaker 2: doing more harm than good at that point, whatever it was. 96 00:06:25,716 --> 00:06:27,276 Speaker 2: And fifteen years ago. 97 00:06:27,356 --> 00:06:31,676 Speaker 1: I would say that things have improved, okay, And trust 98 00:06:31,756 --> 00:06:33,716 Speaker 1: is a very important thing. Trust is a major thing. 99 00:06:34,436 --> 00:06:39,676 Speaker 1: So as AI improves, my trust of using eye increases, 100 00:06:40,276 --> 00:06:43,196 Speaker 1: and then at some point I just kind of relax 101 00:06:44,156 --> 00:06:47,276 Speaker 1: and I'm not so uptight with AI and more like yeah, 102 00:06:47,516 --> 00:06:51,996 Speaker 1: relinquishing some stuff. So we're getting to that point where 103 00:06:52,036 --> 00:06:56,716 Speaker 1: we can basically start trusting, not fully trust, but start 104 00:06:56,756 --> 00:07:02,036 Speaker 1: trusting AI and start asking ourselves, starting to yourselves, maybe 105 00:07:02,196 --> 00:07:02,876 Speaker 1: AI has a. 106 00:07:02,836 --> 00:07:07,276 Speaker 2: Point, And so specifically, how does that work in your 107 00:07:07,676 --> 00:07:10,516 Speaker 2: in your practice of medicine? Like what specific instances where 108 00:07:10,516 --> 00:07:12,156 Speaker 2: AI is helpful. 109 00:07:12,196 --> 00:07:14,916 Speaker 1: Now, so let me give you one very big example. 110 00:07:16,476 --> 00:07:20,116 Speaker 1: So a lot of people have lung nodules, which is 111 00:07:20,196 --> 00:07:25,236 Speaker 1: little dots in their lungs or varying sizes, most of 112 00:07:25,276 --> 00:07:29,036 Speaker 1: which the vast majority of which are not a problem. 113 00:07:29,156 --> 00:07:31,276 Speaker 1: So if somebody says to you you have a lung nodule, 114 00:07:31,276 --> 00:07:34,316 Speaker 1: I wouldn't lose sleep at all. But some of which 115 00:07:34,396 --> 00:07:38,196 Speaker 1: can become conser rather, some of which are precursor to consent. 116 00:07:39,276 --> 00:07:44,716 Speaker 1: So a big example is AI flagging them lung nodules. Right, 117 00:07:45,116 --> 00:07:48,676 Speaker 1: So what that does is it reduces the work on 118 00:07:48,796 --> 00:07:52,356 Speaker 1: my visual field, like I don't need to It's like 119 00:07:52,956 --> 00:07:55,876 Speaker 1: trying to find a needle in a haystack. If it's 120 00:07:55,876 --> 00:07:58,556 Speaker 1: finding those needles in a haystack, I don't need to 121 00:07:58,756 --> 00:08:01,156 Speaker 1: spend my time looking for those needles. 122 00:08:01,676 --> 00:08:04,756 Speaker 2: And it needs to find substantially all of them, and 123 00:08:04,836 --> 00:08:07,196 Speaker 2: not find substantially more than there really are. 124 00:08:07,836 --> 00:08:11,956 Speaker 1: Yeah, and also also needs to tell me based on 125 00:08:12,076 --> 00:08:17,596 Speaker 1: the shape, the size and other aspects which can become 126 00:08:17,676 --> 00:08:19,076 Speaker 1: cancer which can't. 127 00:08:19,676 --> 00:08:25,316 Speaker 2: This presumably is the harder diagnostic problem to solve, so. 128 00:08:25,396 --> 00:08:27,916 Speaker 1: Willing to accept a fewers, it's not that difficult. A 129 00:08:27,956 --> 00:08:32,476 Speaker 1: problem to solve because we can we can ratchet the 130 00:08:32,476 --> 00:08:36,516 Speaker 1: thermostat to tell us that you know, the vast majority 131 00:08:36,556 --> 00:08:41,796 Speaker 1: nodules which are not malignant. You say, they're not malignant. 132 00:08:42,156 --> 00:08:43,196 Speaker 1: We're in good shape. 133 00:08:44,476 --> 00:08:46,756 Speaker 2: Yes, Well, I mean there's always going to be some 134 00:08:46,876 --> 00:08:49,716 Speaker 2: number of errors, and the nature of eliminating errors is 135 00:08:50,356 --> 00:08:52,716 Speaker 2: the more you eliminate, the harder it is to eliminate 136 00:08:52,756 --> 00:08:57,556 Speaker 2: the next one the more. And so presumably the trade 137 00:08:57,596 --> 00:09:02,956 Speaker 2: off with AI is it needs to on net make 138 00:09:03,076 --> 00:09:06,836 Speaker 2: you catch more without taking so much extra time and 139 00:09:06,876 --> 00:09:11,636 Speaker 2: have so much extra intervention so that the benefit exceeds 140 00:09:11,636 --> 00:09:12,036 Speaker 2: the cost. 141 00:09:12,116 --> 00:09:14,596 Speaker 1: Right, Yeah, that is that is pretty much it. 142 00:09:15,076 --> 00:09:18,596 Speaker 2: And are we there now with with with lung nodules. 143 00:09:20,436 --> 00:09:28,596 Speaker 1: I think we candy. I think that we out the 144 00:09:28,636 --> 00:09:33,556 Speaker 1: algorithms that could tell us when there's a lum module 145 00:09:33,836 --> 00:09:38,836 Speaker 1: and when it's a problem without really compromising on the 146 00:09:38,876 --> 00:09:40,196 Speaker 1: false positives too much. 147 00:09:41,036 --> 00:09:42,836 Speaker 2: When you say we can be does that mean it's 148 00:09:42,876 --> 00:09:43,876 Speaker 2: not quite there yet? 149 00:09:43,916 --> 00:09:43,996 Speaker 1: Like? 150 00:09:44,076 --> 00:09:46,396 Speaker 2: Where is this now in your practice? Where is it? 151 00:09:46,436 --> 00:09:50,916 Speaker 2: More generally? And then presumably there's this sort of sociology 152 00:09:51,316 --> 00:09:53,036 Speaker 2: piece of it, for lack of a better word, Right, 153 00:09:53,076 --> 00:09:56,196 Speaker 2: there's does the machine actually work? And then there's do 154 00:09:56,196 --> 00:09:59,156 Speaker 2: doctors actually use it? And the first is a precondition. 155 00:09:59,236 --> 00:10:02,076 Speaker 2: The first is necessary but plainly not sufficient for the second. 156 00:10:02,156 --> 00:10:06,396 Speaker 1: Right, right, And so let's answer your first question about 157 00:10:06,476 --> 00:10:13,076 Speaker 1: the where are we now? We are still in the 158 00:10:13,116 --> 00:10:16,676 Speaker 1: relics of the past, which is that we are still 159 00:10:16,756 --> 00:10:21,356 Speaker 1: obsessed with making sure AI detects correctly and not missus. 160 00:10:21,956 --> 00:10:25,116 Speaker 1: So we're still in where AI is pointing stuff out 161 00:10:25,236 --> 00:10:28,236 Speaker 1: by circles and things of that, as opposed to not 162 00:10:28,396 --> 00:10:30,516 Speaker 1: just pointing it out but putting it in the report. 163 00:10:31,196 --> 00:10:35,476 Speaker 2: I don't really, I mean, I superficially understand the difference 164 00:10:35,596 --> 00:10:37,836 Speaker 2: between those two, But why is that an important distinction? 165 00:10:38,556 --> 00:10:42,076 Speaker 1: Because if you have ten nodules that you flag and 166 00:10:42,116 --> 00:10:45,876 Speaker 1: you don't put it in the report, then radiologists will 167 00:10:45,916 --> 00:10:47,196 Speaker 1: have to put in the report. 168 00:10:47,676 --> 00:10:50,076 Speaker 2: Uh huh, and the ten nodules that it flags. It's 169 00:10:50,076 --> 00:10:52,916 Speaker 2: saying like you're nodules but they're fine, yeah, and then 170 00:10:53,236 --> 00:10:57,436 Speaker 2: making the radio just do that just drudgery of typing 171 00:10:57,556 --> 00:11:01,716 Speaker 2: basically exactly that one. I mean, maybe it's not easy 172 00:11:01,756 --> 00:11:05,676 Speaker 2: to solve for legal or bureaucratic reasons. Sounds easy to 173 00:11:05,716 --> 00:11:06,916 Speaker 2: solve that part of it. 174 00:11:06,916 --> 00:11:09,636 Speaker 1: It shouldn't be difficult it shouldn't be diffult. But you know, 175 00:11:09,676 --> 00:11:12,476 Speaker 1: you got to see your second question about trust. So 176 00:11:12,596 --> 00:11:17,356 Speaker 1: trust is something that obviously takes time. But you know, 177 00:11:18,676 --> 00:11:24,716 Speaker 1: radiologists aren't sort of people that have shied away from technology. Yeah, 178 00:11:24,956 --> 00:11:29,156 Speaker 1: we embraced technology. That's how we've grown over the years. 179 00:11:29,316 --> 00:11:31,956 Speaker 2: I made it only exists because of technology, Like presumably 180 00:11:31,996 --> 00:11:34,916 Speaker 2: you couldn't have had radiologists before you had X rays, 181 00:11:34,956 --> 00:11:38,396 Speaker 2: which are what one hundred and not two hundred years old? 182 00:11:38,436 --> 00:11:39,996 Speaker 2: One hundred hundred plus years old? 183 00:11:40,036 --> 00:11:43,876 Speaker 1: Yeah, I know that old a little awe twenty five. Yeah, 184 00:11:43,916 --> 00:11:48,636 Speaker 1: so if we haven't embraced it, it's probably because it 185 00:11:48,676 --> 00:11:53,236 Speaker 1: hasn't moved the needle for us. And if it moves 186 00:11:53,276 --> 00:11:57,236 Speaker 1: the needle, then it would be self evident. So a 187 00:11:57,276 --> 00:12:01,156 Speaker 1: big example we have from our own technological development was 188 00:12:01,836 --> 00:12:06,876 Speaker 1: the time place where you know, medical imaging. They were 189 00:12:06,916 --> 00:12:09,076 Speaker 1: like films that you would put up on the view boy, 190 00:12:09,476 --> 00:12:13,956 Speaker 1: like the kind of thing. And then it became digitized, 191 00:12:13,996 --> 00:12:18,436 Speaker 1: so you no longer did that picture archiving systems sure, 192 00:12:19,316 --> 00:12:21,796 Speaker 1: and what we call packs or when it became that 193 00:12:23,516 --> 00:12:27,676 Speaker 1: it was adopted left right center, and nobody was complaining 194 00:12:27,676 --> 00:12:31,596 Speaker 1: about it and nobody was asking for evidence it was 195 00:12:31,636 --> 00:12:32,676 Speaker 1: so self evident. 196 00:12:32,916 --> 00:12:36,876 Speaker 2: Yes, although I mean if you extrapolate that out, it 197 00:12:36,916 --> 00:12:39,956 Speaker 2: doesn't have the risk of taking your job in the 198 00:12:39,996 --> 00:12:44,636 Speaker 2: way that AI theoretically might for radiologists at some point, right, 199 00:12:44,716 --> 00:12:47,596 Speaker 2: that seems like a potentially meaningful difference. 200 00:12:48,436 --> 00:12:50,756 Speaker 1: It is, But I don't think that anybody is saying 201 00:12:50,836 --> 00:12:55,596 Speaker 1: no to AI because they are afraid of losing their job. 202 00:12:56,516 --> 00:12:57,916 Speaker 2: Uh, why are they saying no? 203 00:12:58,356 --> 00:13:02,076 Speaker 1: They're saying no because it's not currently being helpful. So 204 00:13:02,116 --> 00:13:05,676 Speaker 1: if you're going to be helpful, you have to either 205 00:13:05,756 --> 00:13:09,596 Speaker 1: make radiologists do less or do things faster. 206 00:13:10,356 --> 00:13:13,716 Speaker 2: And is there any domain yet where AI is helping 207 00:13:13,796 --> 00:13:15,996 Speaker 2: radiologists to either do less or do things. 208 00:13:15,756 --> 00:13:22,236 Speaker 1: Faster as it stands are right now? No? Yeah? 209 00:13:22,436 --> 00:13:25,196 Speaker 2: Interesting is there? This is kind of a question for 210 00:13:25,236 --> 00:13:28,836 Speaker 2: the end, but it's I can't wait because I'm so interested, Like, 211 00:13:29,196 --> 00:13:33,316 Speaker 2: do you feel like there is some broader lesson? Like, 212 00:13:33,356 --> 00:13:37,156 Speaker 2: it's extremely interesting to me that this domain that seems 213 00:13:37,196 --> 00:13:40,236 Speaker 2: like it should be so good for AI, it's still 214 00:13:40,276 --> 00:13:44,436 Speaker 2: not helping. Can you make any inference that goes beyond 215 00:13:44,516 --> 00:13:45,956 Speaker 2: medicine based on that fact? 216 00:13:46,956 --> 00:13:51,436 Speaker 1: That's historically any technology that has come through that is 217 00:13:51,556 --> 00:13:58,156 Speaker 1: automated it hasn't required the person overseeing the automation to 218 00:13:58,436 --> 00:14:02,396 Speaker 1: kind of share the process that leads the automation, the 219 00:14:02,436 --> 00:14:06,596 Speaker 1: scientific calculator and the lock tables. There's a point in 220 00:14:06,596 --> 00:14:09,636 Speaker 1: time people using block tables at a point in time stopped. 221 00:14:10,276 --> 00:14:15,596 Speaker 1: But when they stopped, they weren't checking the calculator's output, right, 222 00:14:15,716 --> 00:14:17,796 Speaker 1: They weren't like, you know, okay, oh, let's just make 223 00:14:17,876 --> 00:14:22,436 Speaker 1: sure that that six figure calculation it did was correct. 224 00:14:22,596 --> 00:14:24,916 Speaker 1: And that's what you need to do in radiology, Like 225 00:14:25,076 --> 00:14:28,916 Speaker 1: if it's going to be AI propelled through, but if 226 00:14:28,996 --> 00:14:32,636 Speaker 1: every time AI is saying something, I have to scratch 227 00:14:32,676 --> 00:14:36,916 Speaker 1: my head and say are you right or wrong? Then 228 00:14:37,156 --> 00:14:40,996 Speaker 1: that's not quite automation. And so if AI is not 229 00:14:41,476 --> 00:14:46,596 Speaker 1: nine nine nine point nine nine percent good, it's not useful. 230 00:14:46,996 --> 00:14:50,916 Speaker 1: It cannot be ninety percent because there's ninety percent and 231 00:14:50,956 --> 00:14:53,756 Speaker 1: there's one ten chance that it could be wrong. I 232 00:14:53,796 --> 00:14:55,316 Speaker 1: don't know what that one in ten is going to 233 00:14:55,396 --> 00:14:59,236 Speaker 1: come from. I still have to treat every single of 234 00:14:59,276 --> 00:15:03,556 Speaker 1: its output as potential one in ten, which doesn't help me. 235 00:15:04,676 --> 00:15:07,116 Speaker 2: You could imagine some world that is not our world, 236 00:15:07,156 --> 00:15:11,156 Speaker 2: that is purely governed by reason, where the AI just 237 00:15:11,236 --> 00:15:13,676 Speaker 2: has to be better than the radiologist. It doesn't have 238 00:15:13,716 --> 00:15:16,036 Speaker 2: to be ninety nine point nine to nine percent, right, 239 00:15:17,156 --> 00:15:20,996 Speaker 2: because you know, radiologists are imperfect like all human beings, right, 240 00:15:22,196 --> 00:15:25,236 Speaker 2: and like all machines, like all probabilistic machines. But we 241 00:15:25,316 --> 00:15:29,676 Speaker 2: don't live in that world. Right For radiologists to trust AI, 242 00:15:29,756 --> 00:15:32,156 Speaker 2: the AI has to be way, way better than radiologists. 243 00:15:32,156 --> 00:15:34,996 Speaker 2: To me, the analogy there is driverless cars. Driveralist cars 244 00:15:34,996 --> 00:15:37,876 Speaker 2: are plainly already safer than humans, but they're not safer 245 00:15:38,076 --> 00:15:42,516 Speaker 2: enough for everybody to trust them everywhere. Does that seem 246 00:15:42,516 --> 00:15:43,916 Speaker 2: relevant in radiology to you? 247 00:15:44,996 --> 00:15:49,596 Speaker 1: So you know when you say that AI should be 248 00:15:49,596 --> 00:15:54,716 Speaker 1: better than radiologists, you know, in terms of diagnosis, better 249 00:15:55,316 --> 00:16:00,756 Speaker 1: is not something that can be easily measured like you 250 00:16:00,796 --> 00:16:06,076 Speaker 1: would measure faster. If somebody is faster, they're faster, and 251 00:16:06,116 --> 00:16:10,476 Speaker 1: there's a timing and they're faster. Into radiology, somebody can 252 00:16:10,516 --> 00:16:14,516 Speaker 1: be better, the better at a certain thing and worse 253 00:16:14,596 --> 00:16:15,356 Speaker 1: than another thing. 254 00:16:15,996 --> 00:16:18,436 Speaker 2: I mean, let me, let me give a naive answer 255 00:16:18,476 --> 00:16:21,036 Speaker 2: and you can tell me why it's not, why it's wrong. 256 00:16:21,836 --> 00:16:24,476 Speaker 2: Fewer false positives and fewer false negatives. 257 00:16:25,116 --> 00:16:27,156 Speaker 1: Oh, I mean, if you can do that, great. 258 00:16:29,996 --> 00:16:31,236 Speaker 2: Like that seems better to me. 259 00:16:31,396 --> 00:16:36,316 Speaker 1: But yeah, yeah, that is definitely better. Okay, fewer false positives, 260 00:16:36,316 --> 00:16:38,596 Speaker 1: a few false negatives, but I would also like to 261 00:16:38,636 --> 00:16:42,356 Speaker 1: see if it can do that. The disease is a spectrum. 262 00:16:42,996 --> 00:16:50,196 Speaker 1: You know, Disease is not just a monolith. It's it's 263 00:16:50,276 --> 00:16:53,676 Speaker 1: certain things are quite so obviously diseased and so obviously 264 00:16:53,716 --> 00:16:57,956 Speaker 1: not disease. So where is it performing the entire range 265 00:16:58,476 --> 00:17:01,196 Speaker 1: of the spectrum. And if it is, then there's no 266 00:17:01,236 --> 00:17:02,236 Speaker 1: need for the radiologists. 267 00:17:03,116 --> 00:17:05,276 Speaker 2: As you're talking about that, I get the sense that 268 00:17:05,316 --> 00:17:07,596 Speaker 2: you feel like that is some distant world that may 269 00:17:07,636 --> 00:17:10,556 Speaker 2: never happens, that a correct inference. 270 00:17:10,476 --> 00:17:12,916 Speaker 1: It could happen. But you know, there are a few 271 00:17:12,916 --> 00:17:16,516 Speaker 1: things to bear in mind. First is I think there's 272 00:17:16,556 --> 00:17:24,276 Speaker 1: this idea that AI has these superhuman skills which somehow 273 00:17:24,556 --> 00:17:27,676 Speaker 1: humans can't figure something out and AI will figure it out. 274 00:17:27,916 --> 00:17:30,796 Speaker 2: In some domains, this appears to be the case some domains. 275 00:17:30,876 --> 00:17:32,196 Speaker 1: Yes, you know, this is not. 276 00:17:32,116 --> 00:17:34,116 Speaker 2: The structure of proteins, notably. 277 00:17:34,276 --> 00:17:38,196 Speaker 1: Sure. But the biggest problem, and I think this problem 278 00:17:38,436 --> 00:17:47,556 Speaker 1: recurs quite consistently, is human biology, human pathology. It's not 279 00:17:47,676 --> 00:17:53,476 Speaker 1: very imaginative. Let's just take something very simple, how disease 280 00:17:53,516 --> 00:17:59,156 Speaker 1: presents itself right, how disease announces itself, So so got 281 00:17:59,156 --> 00:18:03,236 Speaker 1: clinical symptoms, it's not very I maginative. Like if you 282 00:18:03,276 --> 00:18:08,436 Speaker 1: have shoulder pain, for instance, shoulder pain, it could be 283 00:18:08,476 --> 00:18:13,356 Speaker 1: from muscle problems in your shoulder, right. It could also 284 00:18:13,396 --> 00:18:17,716 Speaker 1: be because one of your organs in your abdomen has 285 00:18:17,756 --> 00:18:23,356 Speaker 1: this burst, and it could also be a very uncommon 286 00:18:24,116 --> 00:18:27,876 Speaker 1: manifestation of a heart attack. So there's something about the 287 00:18:27,916 --> 00:18:31,436 Speaker 1: body that just doesn't really have much imagination. So you're 288 00:18:31,476 --> 00:18:36,276 Speaker 1: also saying shoulder pain, three very very different things. Now, 289 00:18:36,756 --> 00:18:41,516 Speaker 1: of course that's just a symptom. Imaging supposed to clarify 290 00:18:41,596 --> 00:18:45,116 Speaker 1: things even further, and it does just a large extent, 291 00:18:45,156 --> 00:18:50,556 Speaker 1: but it also fails. So at some point biology has 292 00:18:50,596 --> 00:18:54,596 Speaker 1: its own limitation. Biology in the way it announces itself 293 00:18:55,116 --> 00:19:00,116 Speaker 1: as a certain limitation. And no matter how much you're 294 00:19:00,116 --> 00:19:04,076 Speaker 1: going to try and decipher that things will look similar. 295 00:19:04,716 --> 00:19:10,756 Speaker 1: Diseases will look similar, and disease and not disease will similar. Yes, 296 00:19:10,876 --> 00:19:13,596 Speaker 1: So the reason many reasons why we you know, you 297 00:19:13,636 --> 00:19:16,636 Speaker 1: wouldn't give it all to AI is because you have 298 00:19:16,716 --> 00:19:20,196 Speaker 1: to have the second check. And even if that second 299 00:19:20,276 --> 00:19:23,796 Speaker 1: check isn't going to make much difference, the mere fact 300 00:19:23,836 --> 00:19:26,636 Speaker 1: of having the second check means that you don't have 301 00:19:26,716 --> 00:19:30,556 Speaker 1: autonomy in a domain where things are less than hun 302 00:19:30,596 --> 00:19:35,356 Speaker 1: percent perfect, and the human in the loop is for 303 00:19:35,516 --> 00:19:39,356 Speaker 1: multiple things, not just simply to make the process more accurate, 304 00:19:39,356 --> 00:19:40,956 Speaker 1: but to have somebody accountable. 305 00:19:45,756 --> 00:19:57,916 Speaker 2: We'll be back in just a minute. Tell me about 306 00:19:58,396 --> 00:20:02,716 Speaker 2: doing X rays at Everest base Camp. 307 00:20:04,236 --> 00:20:10,676 Speaker 1: Yeah, so, so that is one of a big, big 308 00:20:10,756 --> 00:20:13,836 Speaker 1: use case for AI. A few years ago we took 309 00:20:13,836 --> 00:20:16,036 Speaker 1: an X ray machine, which is a portant extra machine 310 00:20:16,076 --> 00:20:22,036 Speaker 1: with AI capabilities, to the Everest clinic in base Camp, 311 00:20:22,756 --> 00:20:26,756 Speaker 1: where they didn't have extras before that and then certainly 312 00:20:26,796 --> 00:20:29,636 Speaker 1: don't have radeologists. And one of the big questions there 313 00:20:29,796 --> 00:20:33,996 Speaker 1: is does the person have what's known as high stud 314 00:20:33,996 --> 00:20:37,796 Speaker 1: pro medema. It's it's it's significant because in a high 315 00:20:37,876 --> 00:20:40,756 Speaker 1: STU pro redeemer, you need to get the person down. 316 00:20:40,876 --> 00:20:43,836 Speaker 1: That's the only treatment. It's not just enough giving them medication. 317 00:20:44,876 --> 00:20:47,036 Speaker 1: So what AI has been able to do in that 318 00:20:47,156 --> 00:20:52,996 Speaker 1: realm is to give assist two people who aren't radiologists. 319 00:20:53,556 --> 00:20:56,276 Speaker 1: And that's a big use case for AI in multiple 320 00:20:56,316 --> 00:20:59,316 Speaker 1: domains where you don't have radiologists and sometimes I just 321 00:20:59,356 --> 00:21:00,516 Speaker 1: don't have physicians. 322 00:21:01,356 --> 00:21:05,716 Speaker 2: So I read that you did a sabbatical not too 323 00:21:05,756 --> 00:21:09,036 Speaker 2: long ago at a university in India, and I'm sure 324 00:21:09,356 --> 00:21:13,356 Speaker 2: more generally if whether during that sabbatical or otherwise, you 325 00:21:14,436 --> 00:21:18,596 Speaker 2: you saw firsthand either people using AI and settings where 326 00:21:18,636 --> 00:21:21,876 Speaker 2: there are not many radiologists, or you saw potential for 327 00:21:21,916 --> 00:21:22,396 Speaker 2: that use. 328 00:21:23,276 --> 00:21:28,996 Speaker 1: So India is a very interesting country because normally when 329 00:21:28,996 --> 00:21:34,716 Speaker 1: you talk about technology, there's a drift that happens between 330 00:21:34,796 --> 00:21:37,996 Speaker 1: high income countries to low medical income countries, where the 331 00:21:38,156 --> 00:21:41,756 Speaker 1: I is kind of the opposite, where its first use 332 00:21:41,836 --> 00:21:45,956 Speaker 1: cases really have been in low middle income countries and 333 00:21:46,196 --> 00:21:52,276 Speaker 1: fighting things like tuberculosis. So tuberculosis is still very rampant 334 00:21:52,396 --> 00:21:59,436 Speaker 1: in Africa in South Asia such as India, and there's 335 00:21:59,476 --> 00:22:03,516 Speaker 1: only so far you can go with detecting tuberculosis using 336 00:22:03,676 --> 00:22:08,636 Speaker 1: the setoscope or looking at the Spoutin further, you have 337 00:22:08,636 --> 00:22:10,956 Speaker 1: to chess sexually, but it's possibly enough to do a 338 00:22:11,036 --> 00:22:14,956 Speaker 1: chest sextuary. That chest sextuary needs to be read and 339 00:22:15,036 --> 00:22:18,716 Speaker 1: read therein then. So imagine you're going around a village 340 00:22:18,716 --> 00:22:22,876 Speaker 1: in India with a portable chest sextualy mounted in a 341 00:22:22,956 --> 00:22:29,636 Speaker 1: bus or something, and you take an image and then 342 00:22:29,676 --> 00:22:32,036 Speaker 1: you send it back to the city and by the 343 00:22:32,076 --> 00:22:35,476 Speaker 1: time it's read three weeks later, you're like, oh, you 344 00:22:35,516 --> 00:22:38,436 Speaker 1: know this person is positive, and now you can't find 345 00:22:38,476 --> 00:22:41,916 Speaker 1: this person. And now this person you know, has TV 346 00:22:43,316 --> 00:22:45,956 Speaker 1: and may very well be infecting other people, but you 347 00:22:45,996 --> 00:22:50,116 Speaker 1: can't find this person because it's three weeks later. Imagine 348 00:22:50,196 --> 00:22:54,356 Speaker 1: if you can give the answer there and then, and 349 00:22:54,396 --> 00:22:58,716 Speaker 1: that's what AI is doing. So AI embedded in the 350 00:22:58,796 --> 00:23:02,236 Speaker 1: chest textually is giving the answer there and then, and 351 00:23:02,276 --> 00:23:03,476 Speaker 1: it's making a huge impact. 352 00:23:04,236 --> 00:23:07,156 Speaker 2: And so is that happening right now? 353 00:23:07,916 --> 00:23:08,236 Speaker 1: AI? 354 00:23:08,836 --> 00:23:11,436 Speaker 2: Yeah, places where there are no radiologists, people are using 355 00:23:11,556 --> 00:23:15,356 Speaker 2: AI to diagnose TB or rule out KIV. 356 00:23:15,676 --> 00:23:19,036 Speaker 1: It's it's been happening for the last I would say, 357 00:23:19,036 --> 00:23:20,876 Speaker 1: at least for the last eight years. 358 00:23:21,316 --> 00:23:22,236 Speaker 2: And how good is it? 359 00:23:22,396 --> 00:23:22,476 Speaker 1: How? 360 00:23:22,596 --> 00:23:24,516 Speaker 2: How how good is it? How good is the AI? 361 00:23:24,716 --> 00:23:32,516 Speaker 1: Excellent? Excellent? Answering the question with high negative predictive values. 362 00:23:32,556 --> 00:23:36,356 Speaker 1: So when the AI says no TV, there's no TV. 363 00:23:37,116 --> 00:23:40,036 Speaker 2: So there are no false negatives they've got They've turned 364 00:23:40,036 --> 00:23:42,516 Speaker 2: the dial way in that direction. How how is the 365 00:23:42,516 --> 00:23:43,356 Speaker 2: false positive rate. 366 00:23:43,716 --> 00:23:46,036 Speaker 1: A false positives? But it's not the only thing. The 367 00:23:46,076 --> 00:23:48,076 Speaker 1: next thing is they have to have a blood tiss 368 00:23:48,796 --> 00:23:50,156 Speaker 1: to see if they have a TV or not. 369 00:23:50,356 --> 00:23:53,316 Speaker 2: How does the false positive rate compare to the false 370 00:23:53,316 --> 00:23:55,956 Speaker 2: positive rate of the median radiologist. 371 00:23:56,956 --> 00:24:00,196 Speaker 1: So I would say that if you have a very 372 00:24:00,276 --> 00:24:03,596 Speaker 1: high end radiologist, like a huge. 373 00:24:03,396 --> 00:24:05,676 Speaker 2: Expense, that's why I gave you median median. 374 00:24:06,556 --> 00:24:09,996 Speaker 1: So it's probably no other median. But it's not as 375 00:24:10,076 --> 00:24:12,476 Speaker 1: good as the expert, or it's just good at the expert. 376 00:24:13,116 --> 00:24:15,116 Speaker 2: So it's so it's pretty good. It's as good as 377 00:24:15,156 --> 00:24:19,076 Speaker 2: like an okay doctor. Yeah, it is okay okay, which 378 00:24:19,156 --> 00:24:21,756 Speaker 2: in a place where there are no doctors or certainly 379 00:24:21,796 --> 00:24:23,596 Speaker 2: not enough to read the X rays, it is great. 380 00:24:23,716 --> 00:24:27,116 Speaker 1: Yeah. But you know, if you ask any physician, are 381 00:24:27,156 --> 00:24:29,756 Speaker 1: you above average? Most of them will say they're above average. 382 00:24:30,356 --> 00:24:33,156 Speaker 2: It's a problem at some level. It's a problem, right 383 00:24:33,716 --> 00:24:38,916 Speaker 2: the lake Wobegonen. The fact is medicine. So you're affiliated 384 00:24:38,956 --> 00:24:41,436 Speaker 2: with the University of Pennsylvania. I assume that means you 385 00:24:41,436 --> 00:24:44,476 Speaker 2: you teach, you work with residents and NED students. Is 386 00:24:44,476 --> 00:24:44,836 Speaker 2: that right? 387 00:24:44,956 --> 00:24:45,436 Speaker 1: That's correct. 388 00:24:46,516 --> 00:24:52,556 Speaker 2: So I'm curious how younger physicians, physicians in training, how 389 00:24:52,596 --> 00:24:55,716 Speaker 2: their relationship to AI is different or is it different 390 00:24:56,596 --> 00:24:57,156 Speaker 2: than yours. 391 00:24:57,556 --> 00:25:01,556 Speaker 1: So that's a very good question, and I think that 392 00:25:01,676 --> 00:25:04,236 Speaker 1: gets sounds to the training that is required in the 393 00:25:04,436 --> 00:25:10,596 Speaker 1: era AI. And you hear all sorts of things about 394 00:25:10,756 --> 00:25:13,956 Speaker 1: AI and radiologists, AI and physicians, et cetera, et cetera. 395 00:25:15,356 --> 00:25:18,956 Speaker 1: But the theme that seems to be very consistent is 396 00:25:18,996 --> 00:25:23,636 Speaker 1: that if you're a real expert, like if you know 397 00:25:23,716 --> 00:25:25,996 Speaker 1: your stuff, if you know medicine and side out, if 398 00:25:25,996 --> 00:25:31,436 Speaker 1: you know radiology inside out, AI can definitely make you better, 399 00:25:32,516 --> 00:25:36,476 Speaker 1: uh huh. But if you're not, if you're starting off 400 00:25:36,596 --> 00:25:40,996 Speaker 1: your nervous your I wouldn't use the word mediocre, but 401 00:25:41,076 --> 00:25:44,916 Speaker 1: I'd say inexperienced, there is a big chance AI can 402 00:25:44,956 --> 00:25:49,796 Speaker 1: make you worse. Huh. And that's because of a phenomenon 403 00:25:49,916 --> 00:25:55,476 Speaker 1: called automation bias where if you don't know enough. You know. 404 00:25:56,196 --> 00:26:01,076 Speaker 1: There's the thing about knowledge, which is that there's a 405 00:26:01,076 --> 00:26:04,796 Speaker 1: certain point. I think it's the Dunark Krugig or Donald 406 00:26:04,796 --> 00:26:06,156 Speaker 1: Krug come over with the name. 407 00:26:06,236 --> 00:26:12,596 Speaker 2: Dunning Kruger Rug. People who aren't very good overestimate their abilities. 408 00:26:12,636 --> 00:26:14,556 Speaker 2: That's the basic that is that right. 409 00:26:15,196 --> 00:26:18,156 Speaker 1: But you know, as you go along that curve, there 410 00:26:18,396 --> 00:26:22,116 Speaker 1: comes a point where you know so much that you 411 00:26:22,156 --> 00:26:26,356 Speaker 1: start knowing where you don't know, like you're actually familiar 412 00:26:26,396 --> 00:26:30,956 Speaker 1: with your ignorance, right, And at that point, at that 413 00:26:31,076 --> 00:26:36,196 Speaker 1: level of expertise, you can use AI intelligently because you 414 00:26:36,356 --> 00:26:39,396 Speaker 1: won't use AI to answer questions that you already know, 415 00:26:41,196 --> 00:26:44,836 Speaker 1: noting that AI can be wrong in those. Rather, you 416 00:26:44,916 --> 00:26:47,036 Speaker 1: know so much that you now you know what you 417 00:26:47,076 --> 00:26:49,836 Speaker 1: don't know, and you can use AI to fill in 418 00:26:49,836 --> 00:26:54,556 Speaker 1: that gap. What happens on the other extreme, the experience extreme, 419 00:26:54,716 --> 00:26:57,156 Speaker 1: is that you don't have that knowledge, so you don't 420 00:26:57,156 --> 00:27:01,316 Speaker 1: know when AI is wrong, and you start relinquishing more 421 00:27:01,316 --> 00:27:03,356 Speaker 1: and more things to AI because you never had the 422 00:27:03,396 --> 00:27:06,636 Speaker 1: point where there was no AI, where you were making 423 00:27:06,676 --> 00:27:10,276 Speaker 1: those decisions on yourself, and you could end up nevi 424 00:27:10,356 --> 00:27:12,556 Speaker 1: getting to the point where you could question AI. 425 00:27:13,236 --> 00:27:17,036 Speaker 2: So this is the trap residents, doctors training to be 426 00:27:17,116 --> 00:27:19,396 Speaker 2: radiologists could fall into, could. 427 00:27:19,196 --> 00:27:22,156 Speaker 1: Fall into, could fall into. And I don't think we've 428 00:27:22,276 --> 00:27:24,516 Speaker 1: kind of figured out quite what to do with that. 429 00:27:25,476 --> 00:27:29,636 Speaker 2: Uh huh. That is an interesting problem that seems to generalize, right, 430 00:27:29,676 --> 00:27:32,236 Speaker 2: You could imagine that trap in all sorts of fields 431 00:27:32,796 --> 00:27:37,516 Speaker 2: where I mean even you know, people worry that kind 432 00:27:37,516 --> 00:27:42,396 Speaker 2: of entry level jobs that senior people view as drudgery, 433 00:27:42,476 --> 00:27:44,036 Speaker 2: but that are in fact the way you learn the 434 00:27:44,076 --> 00:27:46,676 Speaker 2: field you kind of get this sort of intuition by 435 00:27:46,676 --> 00:27:50,076 Speaker 2: doing a thing ten thousand times. Those to the extent 436 00:27:50,076 --> 00:27:52,916 Speaker 2: they're being automated, seem to present that kind of risk 437 00:27:52,996 --> 00:27:55,796 Speaker 2: in many fields potentially. But so what do you do 438 00:27:55,876 --> 00:27:58,876 Speaker 2: about it? You were responsible for training radiologists. How do 439 00:27:58,916 --> 00:28:00,836 Speaker 2: you help them not fall into this trap? 440 00:28:02,516 --> 00:28:07,036 Speaker 1: So you do what Both My kids are undergrunds and 441 00:28:07,796 --> 00:28:13,276 Speaker 1: they're toll very specifically not to use chat gut. They said, 442 00:28:13,676 --> 00:28:16,476 Speaker 1: if we catch you, you're not going to get your grades. 443 00:28:17,316 --> 00:28:20,996 Speaker 1: And better than that is just to examine them in 444 00:28:21,076 --> 00:28:24,796 Speaker 1: scenarios where they can't access it. So I think, I 445 00:28:24,836 --> 00:28:28,356 Speaker 1: think you know, we're in a trial and error situation. 446 00:28:28,996 --> 00:28:33,556 Speaker 1: We're in an experimental area. We haven't quite got any 447 00:28:33,596 --> 00:28:38,596 Speaker 1: precedents for this in the exact manner. But in order 448 00:28:38,676 --> 00:28:42,476 Speaker 1: to leverage AI, you've got to be in a scenario 449 00:28:42,636 --> 00:28:47,476 Speaker 1: where you're a complete physician yourself, and you can't be 450 00:28:47,516 --> 00:28:49,796 Speaker 1: a complete physician if you start using AI early or 451 00:28:49,796 --> 00:28:54,876 Speaker 1: too early, And so that's that's going to require in 452 00:28:54,916 --> 00:28:58,716 Speaker 1: the form of training to force people or at least 453 00:28:58,716 --> 00:29:01,236 Speaker 1: say to people, well, if you use AI, that's fine, 454 00:29:01,276 --> 00:29:05,076 Speaker 1: but your final exams will be without AI. So get 455 00:29:05,156 --> 00:29:07,116 Speaker 1: used to doing it without AI. 456 00:29:08,916 --> 00:29:11,676 Speaker 2: Yeah, I mean, I guess it would seem that in 457 00:29:11,756 --> 00:29:14,756 Speaker 2: radiology it's easier to give people an exam that is 458 00:29:14,836 --> 00:29:18,076 Speaker 2: like what you actually do as a radiologist than it 459 00:29:18,116 --> 00:29:19,996 Speaker 2: is in other domains of medicine. 460 00:29:20,036 --> 00:29:24,596 Speaker 1: Absolutely, And in fact, the American Border Radiology the one 461 00:29:24,596 --> 00:29:29,476 Speaker 1: that conducts the exam years ago, when I was doing 462 00:29:29,476 --> 00:29:33,436 Speaker 1: the exam, it was an oral exam, one on one 463 00:29:33,556 --> 00:29:39,076 Speaker 1: with the examiner. No AI know nothing, then just scans. 464 00:29:39,156 --> 00:29:41,676 Speaker 2: They just show your stands and say what is it? 465 00:29:41,756 --> 00:29:43,956 Speaker 1: Just show scans and on the other side of the 466 00:29:43,956 --> 00:29:48,516 Speaker 1: scan is somebody with an expressionless face. You don't know 467 00:29:48,556 --> 00:29:50,356 Speaker 1: when you're right, you don't know when you're wrong. The 468 00:29:50,356 --> 00:29:54,756 Speaker 1: expression doesn't change. And then they went virtual and they 469 00:29:54,796 --> 00:29:59,236 Speaker 1: gave questions and multiple choice questions. You still couldn't use AI, 470 00:29:59,396 --> 00:30:02,156 Speaker 1: but you had no person overlooking that. Now they're going 471 00:30:02,196 --> 00:30:03,556 Speaker 1: back to the old system, and I think that's. 472 00:30:03,476 --> 00:30:06,036 Speaker 2: A good thing, so that people can't cheat with AI, 473 00:30:06,396 --> 00:30:06,676 Speaker 2: so that. 474 00:30:06,756 --> 00:30:11,476 Speaker 1: People just learn to to people, Yes, the impul thing. 475 00:30:11,996 --> 00:30:14,036 Speaker 1: Most important thing is that you're you know, you're able 476 00:30:14,076 --> 00:30:17,876 Speaker 1: to look somebody in the eye and say, hey, you know, 477 00:30:17,876 --> 00:30:19,876 Speaker 1: I'm reading this scan. Here's what I think it is, 478 00:30:20,476 --> 00:30:22,676 Speaker 1: and that's the best way to make sure that people 479 00:30:22,796 --> 00:30:27,196 Speaker 1: are performing like humans. I mean humans have to be humans. 480 00:30:27,476 --> 00:30:28,796 Speaker 1: I mean to be in the loop, you have to 481 00:30:28,796 --> 00:30:31,356 Speaker 1: be an expert. How could you be in the loop 482 00:30:31,396 --> 00:30:32,676 Speaker 1: if you're not an expert. 483 00:30:34,636 --> 00:30:38,356 Speaker 2: So I want to talk about the future for a minute. 484 00:30:38,676 --> 00:30:41,276 Speaker 2: And you know, maybe there's like the future that you're 485 00:30:41,276 --> 00:30:43,556 Speaker 2: worried about in the future that you hope comes to 486 00:30:43,636 --> 00:30:47,796 Speaker 2: pass in radiology with respect to AI, Like, is there 487 00:30:48,596 --> 00:30:51,196 Speaker 2: do you worry? Like? Is there is there a future 488 00:30:51,236 --> 00:30:52,876 Speaker 2: you worry about in radiology? 489 00:30:53,996 --> 00:30:57,636 Speaker 1: So I'm not worried about my job being taken. I 490 00:30:57,676 --> 00:31:00,996 Speaker 1: think I'm doing to loose sleep about that. 491 00:31:01,516 --> 00:31:04,356 Speaker 2: Would you worry if, like you're a kid who's in 492 00:31:04,396 --> 00:31:06,836 Speaker 2: college wanted to be a radiologist. 493 00:31:06,236 --> 00:31:08,076 Speaker 1: I'd say I'd say that, you know, you should look 494 00:31:08,076 --> 00:31:11,916 Speaker 1: at ten years and be willing to pivot. So the 495 00:31:11,996 --> 00:31:16,156 Speaker 1: thing is what I worry about. I worry about humans 496 00:31:16,196 --> 00:31:20,556 Speaker 1: being stupid or more stupid than we actually are at 497 00:31:20,556 --> 00:31:24,276 Speaker 1: the moment, stupid er. And I think that it's important 498 00:31:24,316 --> 00:31:29,876 Speaker 1: that somehow we don't do much manual work and we 499 00:31:29,956 --> 00:31:32,516 Speaker 1: go to the gym every now and then, but with 500 00:31:32,596 --> 00:31:36,556 Speaker 1: the cognitive stuff, you can't just do that do what 501 00:31:36,676 --> 00:31:38,956 Speaker 1: we do with the biceps. With the cognitive work. We 502 00:31:39,036 --> 00:31:42,516 Speaker 1: have to keep at it otherwise it really will kind 503 00:31:42,556 --> 00:31:47,716 Speaker 1: of disappear. So I do worry about that cognitive disappearance. 504 00:31:48,876 --> 00:31:51,556 Speaker 2: People will stop thinking and get dumber because we have 505 00:31:51,596 --> 00:31:52,956 Speaker 2: stopped thinking exactly. 506 00:31:53,516 --> 00:32:00,436 Speaker 1: I really want to use AI to be smarter and 507 00:32:00,796 --> 00:32:06,276 Speaker 1: to be able to extend my cognitive abilities and to 508 00:32:06,316 --> 00:32:09,796 Speaker 1: see things that I don't see. And that's fun to 509 00:32:09,796 --> 00:32:11,436 Speaker 1: see things that I don't see, and to be able 510 00:32:11,476 --> 00:32:15,156 Speaker 1: to process large amounts of information that I can't process 511 00:32:15,316 --> 00:32:18,036 Speaker 1: because I don't have the time, I don't have the bandwidth, 512 00:32:18,076 --> 00:32:21,756 Speaker 1: and I don't have the computational power. And so I'd 513 00:32:21,796 --> 00:32:24,436 Speaker 1: like to see AI doing that, and I'd like to 514 00:32:24,516 --> 00:32:27,876 Speaker 1: be able to do that. You know, our knowledge is vast. 515 00:32:28,116 --> 00:32:30,836 Speaker 1: A lot of things have changed over the years. We 516 00:32:30,956 --> 00:32:34,276 Speaker 1: used to memorize the bones of the hand, right, the 517 00:32:34,316 --> 00:32:36,396 Speaker 1: bones of the hand. There was a nice pneumonic and 518 00:32:36,436 --> 00:32:38,956 Speaker 1: they're always very kind of your cheesy pneumonic. So I 519 00:32:38,996 --> 00:32:40,676 Speaker 1: can't say it. It's not safe for work. 520 00:32:41,036 --> 00:32:43,476 Speaker 2: I want you to say it now. 521 00:32:44,716 --> 00:32:47,836 Speaker 1: It involves a lot of names and the other acts 522 00:32:47,836 --> 00:32:53,556 Speaker 1: that skateboard lunate tricot piece of form. You probably don't 523 00:32:53,556 --> 00:32:55,636 Speaker 1: need to learn that. You probably don't need to memorize 524 00:32:55,636 --> 00:33:00,556 Speaker 1: all of that. So there's definitely some change in how 525 00:33:00,596 --> 00:33:04,596 Speaker 1: we do things in sort of we can't be doing 526 00:33:04,636 --> 00:33:07,316 Speaker 1: what we were doing in the seventies, eighties and nineties. Now. 527 00:33:07,996 --> 00:33:11,516 Speaker 1: Having said that, there's a call, there's a call of 528 00:33:11,556 --> 00:33:13,956 Speaker 1: being what a physician is. There's a core level of 529 00:33:14,596 --> 00:33:18,796 Speaker 1: totality of knowledge that's important to have. 530 00:33:22,236 --> 00:33:24,356 Speaker 2: We'll be back in a minute with the lightning round. 531 00:33:25,916 --> 00:33:36,556 Speaker 1: M Okay, now we're. 532 00:33:36,396 --> 00:33:38,316 Speaker 2: Going to do the lightning round. The questions are going 533 00:33:38,356 --> 00:33:40,716 Speaker 2: to be somewhat different. What was the hardest thing about 534 00:33:40,716 --> 00:33:41,436 Speaker 2: writing a play? 535 00:33:42,396 --> 00:33:47,996 Speaker 1: It was accepting that nobody would like it. Did anybody 536 00:33:48,116 --> 00:33:54,876 Speaker 1: like it? That was pleasantly surprised. And so it was 537 00:33:54,916 --> 00:33:58,076 Speaker 1: about a doctor in a rural setting that I was 538 00:33:58,116 --> 00:34:04,556 Speaker 1: trying to stay away from technology from information and got 539 00:34:04,596 --> 00:34:08,636 Speaker 1: confronted with a snake bite patient who ended up dying. 540 00:34:09,436 --> 00:34:12,236 Speaker 1: And it was in rural India where they still blame 541 00:34:12,316 --> 00:34:15,236 Speaker 1: doctors when things go bad. They treat doctors like God 542 00:34:15,356 --> 00:34:18,276 Speaker 1: when things go well, but they treat doctors like shit 543 00:34:18,396 --> 00:34:19,756 Speaker 1: when things go bad. 544 00:34:20,076 --> 00:34:23,796 Speaker 2: Do you think AI will replace writers or radiologists. 545 00:34:24,156 --> 00:34:29,996 Speaker 1: First, it will replace mediocre writers before it replaces mediocre ideologists. 546 00:34:30,196 --> 00:34:32,836 Speaker 1: You know. The thing is that do you read the 547 00:34:32,836 --> 00:34:37,036 Speaker 1: New York Times or the economists? Yes? Do you know 548 00:34:37,076 --> 00:34:37,836 Speaker 1: who the author is? 549 00:34:37,836 --> 00:34:41,596 Speaker 2: And the economist only for the oh bits because I 550 00:34:41,636 --> 00:34:42,516 Speaker 2: love her so much. 551 00:34:42,956 --> 00:34:46,036 Speaker 1: Yeah, I have trouble because they all sound the same. 552 00:34:46,476 --> 00:34:49,396 Speaker 1: They all sound the same, but by design, by design, 553 00:34:49,436 --> 00:34:52,396 Speaker 1: by design, because there's a method of writing that they 554 00:34:52,436 --> 00:34:58,956 Speaker 1: all encompass. And AI definitely should take over that. I remember, 555 00:34:59,876 --> 00:35:03,796 Speaker 1: you know, I remember reading people like Christinations, when you 556 00:35:03,836 --> 00:35:06,196 Speaker 1: would read something a boy Christinations or salam ma Reshti 557 00:35:07,356 --> 00:35:08,476 Speaker 1: George orwell. 558 00:35:08,716 --> 00:35:11,076 Speaker 2: Great you were naming great writers. 559 00:35:11,236 --> 00:35:14,876 Speaker 1: You know that the pros came from those people. Yeah, 560 00:35:15,036 --> 00:35:18,796 Speaker 1: because it was unique and for better or worse, it 561 00:35:18,836 --> 00:35:22,796 Speaker 1: had a characteristic And maybe I will take over that too. 562 00:35:22,836 --> 00:35:25,436 Speaker 1: Maybe chap Chept will able to take that too. But 563 00:35:25,636 --> 00:35:29,156 Speaker 1: if it comes to uniformity, if it should, definitely chew 564 00:35:29,196 --> 00:35:32,076 Speaker 1: that up and for breakfast, and I hope it does. 565 00:35:33,836 --> 00:35:36,036 Speaker 2: What's the most beautiful scan you've ever seen? 566 00:35:38,156 --> 00:35:40,436 Speaker 1: I'd say that the most beautiful scan, unfortunately is of 567 00:35:40,436 --> 00:35:43,756 Speaker 1: a fashion that had multiple pathologies. So it was beautiful 568 00:35:43,836 --> 00:35:46,156 Speaker 1: on an aesthetic sense, but not so much for the 569 00:35:46,196 --> 00:35:51,116 Speaker 1: patient's sense. It was somebody that had a stroke from 570 00:35:51,196 --> 00:35:56,916 Speaker 1: a palmary emblass from a clot was called a paradoxical 571 00:35:56,956 --> 00:36:00,276 Speaker 1: emblass crossing from the right side of the heart to 572 00:36:00,276 --> 00:36:03,116 Speaker 1: the left side, and as it was crossing over, it 573 00:36:03,156 --> 00:36:05,636 Speaker 1: was caught in the act crossing over. 574 00:36:06,036 --> 00:36:09,596 Speaker 2: You saw not just the effect but the cause exactly happening. Yeah, 575 00:36:09,836 --> 00:36:12,356 Speaker 2: I've seen you mention a book you have called Atlas 576 00:36:12,396 --> 00:36:16,476 Speaker 2: of Normal rent Invariants that May Simulate disease. Tell me 577 00:36:16,516 --> 00:36:17,316 Speaker 2: about that book. 578 00:36:17,796 --> 00:36:20,796 Speaker 1: It's a book that I started reading when I was 579 00:36:20,796 --> 00:36:25,276 Speaker 1: a resident twenty years ago and I still haven't finished. Keats, 580 00:36:26,116 --> 00:36:30,236 Speaker 1: not the poet, Keats, the iradiologist kids compiled a whole 581 00:36:30,996 --> 00:36:37,196 Speaker 1: range of normal on X rays, and normal is really 582 00:36:37,276 --> 00:36:40,996 Speaker 1: the most difficult thing to learn. 583 00:36:41,956 --> 00:36:44,436 Speaker 2: Say a little bit more like why is normal so hard? 584 00:36:44,476 --> 00:36:49,476 Speaker 2: It is interesting and not immediately obvious. Why that is. 585 00:36:49,436 --> 00:36:54,356 Speaker 1: So because there's so many variations of normal. It's almost 586 00:36:54,476 --> 00:37:00,436 Speaker 1: like what Tolstoy said about happy family all being seen 587 00:37:01,236 --> 00:37:04,116 Speaker 1: and every unhappy family being unhappy in their own way. 588 00:37:05,076 --> 00:37:10,156 Speaker 1: You reverse that disease is a clear disease kind of 589 00:37:11,156 --> 00:37:18,996 Speaker 1: is very similar homogenous to some extent, whereas normal has 590 00:37:19,036 --> 00:37:21,356 Speaker 1: a lot more variability. 591 00:37:21,956 --> 00:37:25,556 Speaker 2: What's your view of healthy people getting a full body scan. 592 00:37:26,276 --> 00:37:26,956 Speaker 1: I don't bother. 593 00:37:28,116 --> 00:37:31,116 Speaker 2: Yeah, why because you will end up. 594 00:37:31,036 --> 00:37:36,516 Speaker 1: Calling overcalling normal, over diagnosing normal. I mean, yeah, there'll 595 00:37:36,556 --> 00:37:40,516 Speaker 1: always be the success story of one person that whose 596 00:37:40,556 --> 00:37:43,796 Speaker 1: life was saved, and genuinely so. But for every success story, 597 00:37:43,796 --> 00:37:46,596 Speaker 1: they're about ninety nine where they'll be just taken down, 598 00:37:47,156 --> 00:37:51,116 Speaker 1: you know, a rabbit hole for no reason whatsoever and 599 00:37:51,236 --> 00:37:53,996 Speaker 1: reassured at the end of it and felt like there's 600 00:37:54,036 --> 00:37:58,036 Speaker 1: some sort of cancer survi not they. I wouldn't bother. 601 00:37:58,796 --> 00:38:01,636 Speaker 1: But I also think that it's it's it's a decision 602 00:38:01,876 --> 00:38:05,556 Speaker 1: made by the individual. Some people are risk averse and 603 00:38:05,556 --> 00:38:08,076 Speaker 1: you have to just take what yours restaurerances. 604 00:38:08,116 --> 00:38:13,196 Speaker 2: As well anything else we should talk about, well, I think. 605 00:38:13,076 --> 00:38:16,116 Speaker 1: I think you've covered a lot. I think that people 606 00:38:16,236 --> 00:38:22,556 Speaker 1: view AI with both. On one hand, you've got this 607 00:38:23,876 --> 00:38:26,836 Speaker 1: view largely coming from Silicon Valley, and I understand why 608 00:38:26,836 --> 00:38:29,996 Speaker 1: they need to make money with view of AI has 609 00:38:30,116 --> 00:38:35,276 Speaker 1: just been this sort of transforming technology. And on the 610 00:38:35,276 --> 00:38:38,476 Speaker 1: other hand, there's there's a fear and skepticism of it. 611 00:38:39,196 --> 00:38:42,716 Speaker 1: And I think the truth is somewhere in between. That is, 612 00:38:43,036 --> 00:38:49,516 Speaker 1: it is a technology that needs nurturing, and it's a 613 00:38:49,596 --> 00:38:52,796 Speaker 1: very personal technology. I think it'll end up being very 614 00:38:52,796 --> 00:38:57,276 Speaker 1: personal to the person nurturing it. No two people will 615 00:38:57,356 --> 00:38:58,516 Speaker 1: view AI in the same way. 616 00:38:59,516 --> 00:39:02,356 Speaker 2: I appreciate your time. Thank you for being so generous 617 00:39:02,356 --> 00:39:11,236 Speaker 2: with your time. Thank you very much, sort of. Jaw 618 00:39:11,436 --> 00:39:15,876 Speaker 2: is a radiologist at the University of Pennsylvania. Please email 619 00:39:15,956 --> 00:39:19,036 Speaker 2: us at problem at pushkin dot fm. Tell us what 620 00:39:19,196 --> 00:39:20,876 Speaker 2: kinds of shows we should do more of or less 621 00:39:20,916 --> 00:39:22,756 Speaker 2: of our particular people you think would be good on 622 00:39:22,796 --> 00:39:25,076 Speaker 2: the show. You can also find me on x and 623 00:39:25,156 --> 00:39:28,556 Speaker 2: on LinkedIn. Our show is produced by Gabriel Hunter Chang 624 00:39:28,796 --> 00:39:31,916 Speaker 2: and Trina Menino. Our editor is Lydia Jean Kott, and 625 00:39:31,996 --> 00:39:35,436 Speaker 2: our engineer is Sarah Bruguier. I'm Jacob Goldstein and we'll 626 00:39:35,476 --> 00:39:37,556 Speaker 2: be back next week with another episode of What's Your 627 00:39:37,556 --> 00:39:37,796 Speaker 2: PALP