-
Interesting New Yorker article about Deep-Learning, Medicine and Radiology
[link=http://www.newyorker.com/magazine/2017/04/03/ai-versus-md]http://www.newyorker.com/…017/04/03/ai-versus-md[/link]
In mammography, too, computer-aided detection is becoming commonplace. Pattern-recognition software highlights suspicious areas, and radiologists review the results. But here again the recognition software typically uses a rule-based system to identify a suspicious lesion. Such programs have no built-in mechanism to learn: a machine that has seen three thousand X-rays is no wiser than one that has seen just four. These limitations became starkly evident in a 2007 study that compared the accuracy of mammography before and after the implementation of computer-aided diagnostic devices. One might have expected the accuracy of diagnosis to have increased dramatically after the devices had been implemented. As it happens, the devices had a complicated effect. The rate of biopsies shot up in the computer-assisted group. Yet the detection of small, invasive breast cancersthe kind that oncologists are most keen to detect[i]decreased[/i]. (Even later studies have shown problems with false positives.)
Thrun was convinced that he could outdo these first-generation diagnostic devices by moving away from rule-based algorithms to learning-based onesfrom rendering a diagnosis by knowing that to doing so by knowing how. Increasingly, learning algorithms of the kind that Thrun works with involve a computing strategy known as a neural network, because its inspired by a model of how the brain functions. In the brain, neural synapses are strengthened and weakened through repeated activation; these digital systems aim to achieve something similar through mathematical means, adjusting the weights of the connections to move toward the desired output. The more powerful ones have something akin to layers of neurons, each processing the input data and sending the results up to the next layer. Hence, deep learning.Thrun began with skin cancer; in particular, keratinocyte carcinoma (the most common class of cancer in the U.S.) and melanoma (the most dangerous kind of skin cancer). Could a machine be taught to distinguish skin cancer from a benign skin conditionacne, a rash, or a moleby scanning a photograph? If a dermatologist can do it, then a machine should be able to do it as well, Thrun reasoned. Perhaps a machine could do it even better.
Traditionally, dermatological teaching about melanoma begins with a rule-based system that, as medical students learn, comes with a convenient mnemonic: ABCD. Melanomas are often asymmetrical (A), their borders (B) are uneven, their color (C) can be patchy and variegated, and their diameter (D) is usually greater than six millimetres. But, when Thrun looked through specimens of melanomas in medical textbooks and on the Web, he found examples where none of these rules applied.
Thrun, who had maintained an adjunct position at Stanford, enlisted two students he worked with there, Andre Esteva and Brett Kuprel. Their first task was to create a so-called teaching set: a vast trove of images that would be used to teach the machine to recognize a malignancy. Searching online, Esteva and Kuprel found eighteen repositories of skin-lesion images that had been classified by dermatologists. This rogues gallery contained nearly a hundred and thirty thousand imagesof acne, rashes, insect bites, allergic reactions, and cancersthat dermatologists had categorized into nearly two thousand diseases. Notably, there was a set of two thousand lesions that had also been biopsied and examined by pathologists, and thereby diagnosed with near-certainty.
Esteva and Kuprel began to train the system. They didnt program it with rules; they didnt teach it about ABCD. Instead, they fed the images, and their diagnostic classifications, to the neural network. I asked Thrun to describe what such a network did.Imagine an old-fashioned program to identify a dog, he said. A software engineer would write a thousand if-then-else statements: [i]if[/i] it has ears, and a snout, [i]and[/i] has hair, and [i]is not[/i] a rat . . . and so forth, ad infinitum. But thats not how a child learns to identify a dog, of course. At first, she learns by seeing dogs and being told that they are dogs. She makes mistakes, and corrects herself. She thinks that a wolf is a dogbut is told that it belongs to an altogether different category. And so she shifts her understanding bit by bit: this is dog, that is wolf. The machine-learning algorithm, like the child, pulls information from a training set that has been classified. Heres a dog, and heres not a dog. It then extracts features from one set versus another. And, by testing itself against hundreds and thousands of classified images, it begins to create its own way to recognize a dogagain, the way a child does. It just knows [i]how[/i] to do it.
In June, 2015, Thruns team began to test what the machine had learned from the master set of images by presenting it with a validation set: some fourteen thousand images that had been diagnosed by dermatologists (although not necessarily by biopsy). Could the system correctly classify the images into three diagnostic categoriesbenign lesions, malignant lesions, and non-cancerous growths? The system got the answer right seventy-two per cent of the time. (The actual output of the algorithm is not yes or no but a probability that a given lesion belongs to a category of interest.) Two board-certified dermatologists who were tested alongside did worse: they got the answer correct sixty-six per cent of the time.
Thrun, Esteva, and Kuprel then widened the study to include twenty-five dermatologists, and this time they used a gold-standard test set of roughly two thousand biopsy-proven images. In almost every test, the machine was more sensitive than doctors: it was less likely to miss a melanoma. It was also more specific: it was less likely to call something a melanoma when it wasnt. In every test, the network outperformed expert dermatologists, the team concluded, in a report published in [i]Nature[/i].
about radiology:
Geoffrey Hinton, a computer scientist at the University of Toronto, speaks less gently about the role that learning machines will play in clinical medicine. Hintonthe great-great-grandson of George Boole, whose Boolean algebra is a keystone of digital computinghas sometimes been called the father of deep learning; its a topic hes worked on since the mid-nineteen-seventies, and many of his students have become principal architects of the field today.
I think that if you work as a radiologist you are like Wile E. Coyote in the cartoon, Hinton told me. Youre already over the edge of the cliff, but you havent yet looked down. Theres no ground underneath. Deep-learning systems for breast and heart imaging have already been developed commercially. Its just completely obvious that in five years deep learning is going to do better than radiologists, he went on. It [i]might[/i] be ten years. I said this at a hospital. It did not go down too well.
-
55 Replies
-
Ok, let’s all check back in 5 years.
-
Well. That’s depressing.
-
-
Quote from qwerty89
[link=http://www.newyorker.com/magazine/2017/04/03/ai-versus-md]http://www.newyorker.com/…017/04/03/ai-versus-md[/link]
In mammography, too, computer-aided detection is becoming commonplace. Pattern-recognition software highlights suspicious areas, and radiologists review the results. But here again the recognition software typically …
You chose to highlight the most “sky is falling” portion of the paper. It’s actually not that bad. People with purely computer science/math backgrounds have a hard time understanding what doctors do and tend to overestimate their creation’s ability to replicate it. In fact it took me several years in medical school before I realized that automating a doctor’s job isn’t so simple despite how algorithmic it seems on cursory glance. On the other hand, as physicians we tend to overestimate how immune we are to the effects of automation.
Other, less alarming parts of the article are below:
“When I brought up the challenge to Dr. Angela Lignelli-Dipple, she pointed out that diagnostic radiologists arent merely engaged in yes-no classification. Theyre not just locating the embolism that brought on a stroke. Theyre noticing the small bleed elsewhere that might make it disastrous to use a clot-busting drug; theyre picking up on an unexpected, maybe still asymptomatic tumor.”
“Hinton now qualifies the provocation. The role of radiologists will evolve from doing perceptual things that could probably be done by a highly trained pigeon to doing far more cognitive things, he told me. His prognosis for the future of automated medicine is based on a simple principle: Take any old classification problem where you have a lot of data, and its going to be solved by deep learning.”
-
Stop posting doom and gloom stuff. There are plenty of quacks out there. He doesn’t justify being printed on it. He says he is in it to help others but he is just another person in it for himself. We really need to stop focusIng so much on all these people yelling use radiology and medicine are going to crap. Aunt minnie please stop posting these doom and gloom articles. It’s stuff like this that drives your traffic away!
-
The article starts by discussing how radiologists learn, then goes into how computers could learn in a similar way, brings up dermatologists, and then ends by saying the fact that dermatologists actually see and talk to patients saves them from automation. Not the same for us, unfortunately! I’ll see you at the unemployment line.
-
RE: AI and other computer detection taking over radiologists’ livelihoods, I’ve never understood the interest in the topic. If it happens, when it happens… it happens. I don’t lose sleep or feel anxious about that prospect. My work currently includes breast. If ever needed, I’ll do IR fellowship and do interventions and biopsies. If that isn’t available, then something else. I’m confident that whatever happens, I’ll figure something out and keep practicing good medicine and keep riding my bicycle 200-300 miles a week. “it’ll be awl-riiiight”
-
“Hintons actual words, in that hospital talk, were blunt: ‘They should stop training radiologists now.’
[i][b]Exactly[/b][/i] the point I’ve made when I’ve written that we need to slash the number of training spots.
-
It’s really not as bad as some of you are making it out to be. Machine learning will give us productivity tools far sooner than it will give us general intelligence. It would take general intelligence to replace everything we do, and at that point, society in general will not look like it does today. Productivity tools will allow us to focus on things that provide more clinical value, such as consulting with clinicians. We will just have to evolve with the times.
-
-
Quote from Flounce
RE: AI and other computer detection taking over radiologists’ livelihoods, I’ve never understood the interest in the topic. If it happens, when it happens… it happens. I don’t lose sleep or feel anxious about that prospect. My work currently includes breast. If ever needed, I’ll do IR fellowship and do interventions and biopsies. If that isn’t available, then something else. I’m confident that whatever happens, I’ll figure something out and keep practicing good medicine and keep riding my bicycle 200-300 miles a week. “it’ll be awl-riiiight”
There might be a pretty long line for that IR fellowship…
-
Yeah there isn’t work for 35000 IRs (or whatever the number of radiologists there are now). But, even if there is a big effect on the ability to work as a rad (not that likely) it’ll happen slowly in terms of employment prospects and hopefully people will have time to adjust.
-
The transition will be over a decade. The problem is, when people finally wake up and understand the implications of AI, everybody will be clamoring for the alternatives.
The effect will be similar to shouting fire in crowded theater.
-
Of course, then there’s this:
[link=http://www.carestream.com/blog/2016/11/01/why-computers-cant-replace-radiologists/]http://www.carestream.com…-replace-radiologists/[/link]
As with hospital admin, pretty big disconnect between what the techies think something does and how it actually performs (much like PACS engineers designing something totally impractical for use by radiologists).-
“The Sky is going to Fall.”
Chicken Little.
I would encourage all to listen to the AI talk given at RSNA last year, and hear all that it would take to implement in terms of support. You will rest a little easier. Listen to someone who really understands BOTH radiology and AI.
Great article in Radiographics this month on AI as well. Worth a read to help us understand. -
Quote from vomer
Of course, then there’s this:
[link=http://www.carestream.com/blog/2016/11/01/why-computers-cant-replace-radiologists/]http://www.carestream.com…-replace-radiologists/[/link]
As with hospital admin, pretty big disconnect between what the techies think something does and how it actually performs (much like PACS engineers designing something totally impractical for use by radiologists).
Siegel seemed like a pretty cool guy I saw him at the ergonomics lecture at RSNA this year. Very open to learn and discuss what works. Also, you can tell he likes being an academic doing essentially non-rad things.-
AI- the least plausible threat to radiology. Let residency applicants believe it. Maybe we can a better job market out of the sky is falling crowd
-
AI won’t completely replace us anytime soon, but it will make us much more productive (much like the transition from film to PACS).
If increased productivity is not accompanied by increased volume, then that means fewer radiologists (if not zero yet).
-
Remind me again, how long as CAD been around for Mammo? How many of you would use it primarily on your family members’ Mammos?
yeah, right. -
Most computer vision techniques in common use before 2012 are completely irrelevant to current progress in image classification. And yeah, that includes the CAD in your mammography workstation.
The sky isn’t falling yet, but it might be wise to get your head out of the sand and look at some recent progress. Most of the exciting stuff has been applied to non-medical images, but with a few caveats the same techniques could be just as easily applied to medical imaging as long as there are good datasets available.
Expect to hear big news about screening mammography sometime this summer.
-
Are you implying mammographers will be out of a job by next summer?
-
thank you very much but I think i’ll take Dr. Siegel’s opinion expressed in the carestream link above over some random AM poster’s fake insight about screening mammograms that will start reading themselves next summer:-).
-
What Siegel says makes you guys feel comfortable, don’t it? Really, it doesn’t matter who says it anymore. You guys will believe anyone who says “everything will be just fine, you have nothing to worry about”.
I don’t know what anon is referring to but things are defintely approaching fast.
BTW, Siegel’s article is from last year. It would be interesting to hear what he says now.
-
I’ll start panic when someone actually shows me a report generated by a machine learning tool in real time on a study I’ve fed it.
-
Siegel has been actively involved in teaching Watson. As I described in a post-RSNA blog post, Siegel doesn’t buy into the doomsday scenario. IBM has no plan to train Watson to take over, even if it could do so.
AI will be another tool in our armamentarium, nothing more, nothing less. It will NOT replace any of us.
I gave a talk to a bunch of medical students interested in radiology yesterday. The first question out of their mouths? Is AI taking over? If the goal of the chicken-littles around here is to scare away the competition, you are marginally succeeding. -
I do not think AI is near taking over our jobs.
But this thread does provide an opportunity for some reflection:
– if your PP group downsized and only retained 20% of its current radiologists, would you be kept on or let go?
– if you were let go and diagnostic radiology jobs were few, and you decided to compete with the many applicants for an IR fellowship position, how competitive would you be as an applicant and how robust is your network, i.e. would you be able to secure a position?
– if you had the opportunity – let’s see it as a forced opportunity – to change fields from radiology to something else, whether medical or non medical, what vocation would you pursue that could pay the bills and that you would enjoy much more than radiology? What would be so fun to do that – despite a huge drop income – you’d say, “screw it, I’m gonna do it” ?
-
Improved efficiency/cost from AI could lead to substantially increased volumes as well, so it could be a wash for the job market. Of course, the competing hardware vendors will continue to find ways to make their imaging more complex and interesting, leading to continual and expensive AI software upgrades (all those curated cases, training, QA, and FDA approval, all of which are unnecessary with human radiologists).
However, having worked with neural networks for the past 15 years including for medical applications, and having worked with software packages (PACS, RIS, VR) from all the major vendors – it does not seem realistic to me that enterprise-grade software with the capability to replace a radiologist is likely to be forthcoming from any of the established players. Maybe new players are interested (Google, IBM, little startups) but there are opportunities for them to learn deep and painful lessons as their 22-year old Stanford CS grads venture forth to conquer the medical device arena. -
Quote from DoctorDalai
Siegel has been actively involved in teaching Watson. As I described in a post-RSNA blog post, Siegel doesn’t buy into the doomsday scenario. IBM has no plan to train Watson to take over, even if it could do so.
AI will be another tool in our [i][b]armamentarium[/b][/i], nothing more, nothing less. It will NOT replace any of us.
I gave a talk to a bunch of medical students interested in radiology yesterday. The first question out of their mouths? Is AI taking over? If the goal of the chicken-littles around here is to scare away the competition, you are marginally succeeding.
nothing to add here, i just want to give kudos to Dr. D for his use of the term “armamentarium”. -
At least no one brings up Terminator or Hal from 2001 references anymore.
No longer a laughing matter, is it?
Stay tuned.
-
Jan the third your link isn’t working.
-
Jimboboy, serious question: what do you think current radiologists should realistically do about (as you see it) the oncoming train of AI that decimates the field?
-
The answer to that depends on your own circumstances.
I am not here to give you answers. Just letting you aware, so you can plan. whatever that means for you.
I could be wrong about the whole AI thing. But I seriously doubt it. The capability is there already. It’s just a matter of applying it specifically to radiology.
From there, you have to postulate what would happen if the machine capability supasses (or even merely approches) an average radiologist, which it undoubtedly will.
All of the ancillary arguments, like who to sue, regulatory hurdles will melt away.
How do you justify using a person interpreter if the machine can do it just as well and much cheaper? What if it can do better?
From healthcare cost perspective, reason to replace exists even when AI is only marginally worse.
-
OK, so you don’t have answers….
I think I’ll just go back to reading this train wreck of a postop CT abd/pelvis. The surgeon is very likely going to ask for CT guided drain placement for abscesses.
-
No, I dont have answer for YOU.
Get back to work, its a start.
-
But soon there will be a Robot to drain the abscess.
-
Actually, I do appreciate this thread. Although I don’t expect to ever need it…. I just looked up the current IR faculty at my former residency program: same chief, and two of the attendings were my co-residents, one of whom I advised. I’ll shoot them each a “what’s up” email, and next time I’m in the area I’ll drop by to visit. If AI suddenly takes over diagnostic radiology, securing a top IR fellowship will hopefully be just an email away.
-
Not everyone is lucky like Flounce!
-
Not sure if that was a cheap shot at Flounce, but it’s called working hard at making connections with colleagues.
Doesn’t just come out of thin air. -
Did not mean for it to be a cheap shot. It takes a lot of work to end up in places like Boston. Hard work comes with priviledges. However, a lot of DR folks don’t have that connection.
-
[image]http://www.valuenews.com/images/news_images/photo-1-roadrunner-and-wile-e-coyote-value-news-story-august-2015_080615161423.jpg[/image]
For added effect (in case you didn’t read/watch it as a kid): [image]https://cdn-images-1.medium.com/max/592/1*3_h4T_3qxebCOJR6dxhZZg.jpeg[/image]
-
-
-
-
I am confident that for the near future the increasing demand for imaging combined with the aging baby boomer generation will keep things balance even if AI starts to eat away bits of our lunch. On long term? Well, 1) “In the long run we are all dead” (Keynes) 2) The same AI experts who think radiologists are going to be displaced by robots also say that the singularity will happen in 20 years – and if it comes to that job security will be the least of our problems.
-
all I know is that every time I’ve freaked out about some “game changer” on the horizon I’ve ended up working harder and making more money. well, marginally in a lot of those cases.
many years ago a bloc of physicians left my main hospital practice and our group quietly freaked about the upcoming loss of income, only for them to be replaced by folks dumber and less experienced, including a bunch of PAs and FNPs, who order waaay more than the old group did, and income went up….
when some women started raising hell about breast density we semi-freaked about all the imaginary pain the state legislation was going to bring forth. fast forward several years out and we’re making sweet little coin from adjunctive ultrasound screens of dense breasts most of which are negative.
mammo CAD has added a tiny fraction to revenue per mammo and has engendered a few work ups in the process of potential lesions that can’t be left alone. LDCT is bringing us a little more coin, and if someone makes CAD for lung nodules that will add a little more money per study, I hope.
unless Americans fundamentally change their DNA and entrust their precious lives to the whims of a machine (eg. EKG reader without cardiologist input) we’ll keep reading films. but folks have warmed to automatic check-out machines at the grocery store so who knows how they might react to a print out from the “Google MRI UltraReader 3000” detailing their knee MRI findings….
selfishly, by that time, with paid for house over my family’s head, college tuition taken care of for the kids, and some dough stashed away, I’d be ready to be a bum. read books I’ve been meaning to read, play the guitar. maybe become volunteer faculty at the local med school. life is good-
Part time fly fishing guide, part time ski instructor. Live “close to the earth”, and live small. But only part time.
Still not ready to dissuade 14 yo from pursuing radiology, though.
-
What is the coyote comparison?
-
[link=http://www.valuenews.com/images/news_images/photo-1-roadrunner-and-wile-e-coyote-value-news-story-august-2015_080615161423]http://www.valuenews.com/…gust-2015_080615161423[/link]
-
Will it really be less inexpensive for AI to interpret exams (assuming it becomes possible at some point)? -
[link=https://www.youtube.com/watch?v=2HMPRXstSvQ&t=2s]https://www.youtube.com/w..?v=2HMPRXstSvQ&t=2s[/link]
You are welcome -
Hinton is an attention seeking stuffed shirt. In 2016 he predicted radiologists would be gone in 5 years. Obama, for one, unquestionably believed him and repeated the prediction. (this is the quote above).
Well, 2021 came and went and we are desperately short of radiologists. He was dead wrong. Should we believe anything he says.I doubt he has spent 5 minutes with a radiologist watching us read. He would not understand that what we do is far more than see spots.
This is not to say he is necessarily wrong. Stopped clocks, and all that. Make a prediction that is broad enough, and you can be right inadvertently.
What I would like here is an illustration I remember. It was from a newspaper article written in about 1964 or so. It made predictions about the future, and targeted, naturally enough 2000.
One of the predictions I remember that seemed so very obvious is that we would all commute in flying cars, or helicopters if you like. CERTAINLY this would happen, it was a natural extension of what existed at the time.(Aside – the neighborhood I live in is about 15 miles from Downtown Columbus. It is configured as a circle, about 0.8 miles in diameter, with houses arranged around the periphery. I learned that the central empty area was designated to be a heliport for the residents commuting downtown. At the time, it was probably a 30-40 minute drive, as there were only local streets on which to commute. So, maybe sensible. BUT, what happened is that we built high capacity freeways instead, and now it is maybe 17 minutes downtown. And offices were built away from downtown, and now, there is telecommuting. All of which made heliports irrelevant.
So the central circle was sold off for more home sites in the 70’s after it became obvious helicopters wouldn’t happen. For dramatic illustrative purposes, I am including an aerial photo of our neighborhood in 1961 and 1980. Just because I could.
It is said “Prediction is hard – especially about the future” and this (and AI) are examples.)
check out this guy – a welcome antidote to all that:
[link=https://www.uab.edu/reporter/people/achievements/item/9925-this-radiologist-is-helping-doctors-see-through-the-hype-to-an-ai-future]https://www.uab.edu/reporter/people/achievements/item/9925-this-radiologist-is-helping-doctors-see-through-the-hype-to-an-ai-future[/link] -
By the way, there are SOOOOOOOO MANY true experts in AI and radiology the author could have talked to. RSNA and ACR have large number of real imaging scientists doing AI work. They know both AI and radiology, But they chose Hinton . Reasons? Quotable? Out there? Dramatic?
-
Read this as well:
[link=https://towardsdatascience.com/why-data-scientists-should-see-like-radiologists-13762a212f9]https://towardsdatascience.com/why-data-scientists-should-see-like-radiologists-13762a212f9[/link]
-
Quote from irinterview2017
What is the coyote comparison?
Radiologists aren’t aware that they will soon fall, i.e; cease to exist at least in the capacity that they do now.
-
-
-
-
-
-
-
-
-
Quote from Teedevil
“Hinton now qualifies the provocation. The role of radiologists will evolve from doing perceptual things that could probably be done by a highly trained pigeon to doing far more cognitive things, he told me. His prognosis for the future of automated medicine is based on a simple principle: Take any old classification problem where you have a lot of data, and its going to be solved by deep learning.”
I love how his opinion has become softer since the infamous coyote comparison went viral – it was even featured in a talk at ECR 2017 in Vienna.
Altogether this article is a good primer, but its added value is negligible. The same people as always had yet another chance to reiterate their opinion, but there’s not much new knowledge to be gained from this. Machine learning and AI or not, radiology will certainly undergo major changes in the coming decades and only one thing is sure, we have to be able to adapt to an increasingly rapidly changing environment.
-
-
Quote from qwerty89
[link=http://www.newyorker.com/magazine/2017/04/03/ai-versus-md]http://www.newyorker.com/…017/04/03/ai-versus-md[/link]
Geoffrey Hinton, a computer scientist at the University of Toronto, speaks less gently about the role that learning machines will play in clinical medicine. Hintonthe great-great-grandson of George Boole, whose Boolean algebra is a keystone of digital computinghas sometimes been called the father of deep learning; its a topic hes worked on since the mid-nineteen-seventies, and many of his students have become principal architects of the field today.
I think that if you work as a radiologist you are like Wile E. Coyote in the cartoon, Hinton told me. Youre already over the edge of the cliff, but you havent yet looked down. Theres no ground underneath. Deep-learning systems for breast and heart imaging have already been developed commercially. Its just completely obvious that in five years deep learning is going to do better than radiologists, he went on. It [i]might[/i] be ten years. I said this at a hospital. It did not go down too well.Quote from sandeep panga
Ok, let’s all check back in 5 years.
[link=https://twitter.com/doRadiology/status/1662951855649849346]https://twitter.com/doRad…us/1662951855649849346[/link]
So [link=https://twitter.com/geoffreyhinton]@geoffreyhinton[/link] finally admits he was wrong… ..when 8 years ago he claimed: “it’s completely obvious we should stop training radiologists because in 5 yrs my AI will replace them”
…wait for it…
Only to now claim he was right all along! It will be in 10 years, not 5.
“AI is now comparable to radiologists, just not *way* better than all of them yet”
…
so I guess we have to check back again in 2 years when we will hit Hinton’s 10 year mark.
-
Apparently all it takes is some capital letters to completely fool ChatGPT.
-
The guy is more of a celebrity than a programmer and certainly not a scientist. I dont even get why he is so popular/listened to. I think its because he makes these absurd, not scientifically backed and crazy statements that the press believes are edgey viewpoints. So he gets more of a following. The way he even talks about computers are about the same as radiologists sounds so dumb. We are talking about a few limited use examples of AI, like ct perfusion, where in a real setting it performs similar to a rad. So were saying because in this one very particular use they are similar, overall a computer is as good as a radiologist. So ask Hinton to provide an ER doc with a computer and ask if hed rather have a computer or radiologist reading studies.
None of them get it. They think its all pattern recognition and auntminnie type diagnostics, not requiring critical thinking or judgement. To them there is no difference between teaching an image analysis NN to identify a malignant lymph node on Ct images compared with identifying a dog in photographs.
-
Thats very well put
-
No chance CT perfusion is even as good as a radiologist. Ill admit that it has picked up some small infarcts my eye didnt catch, but if you allowed it to act solo there would be a giant number of artifactual strokes being treated and posterior fossa strokes going untreated.
-
-
-
-
