6 ms·

Google’s lung cancer detection AI outperforms 6 human radiologists

27 points by jseip 7y ago

jseip 7y ago
It's only a matter of time before all X-Ray, CTScans, etc. use algorithms for secondary if not primary screens.
avgDev 7y ago
Having seen many specialist/doctors I generally think they suck at diagnosing, unless its something you could literally discover yourself by a google search and access to a specific test. I don't think they are bad doctors, I just learned that sometimes diagnosis can be rather difficult. Then, you have a computer, it can reference recent studies compare symptoms with hundreds of diseases and generate a test plan accordingly based on probability of a disease or urgency. Frankly, I am surprised not more work is being done with algorithms in medicine. Humans are biased, egoistic and most don't like being told they are wrong.
- gremlinsinc 7y ago
  I tend to agree, doctors is going to be the holy shit moment for society when we begin replacing them across the board. I mean we still might need 100 or so (nationwide in the usa), maybe keep the top 1000 doctors/surgeons for times when creativity and unique problem solving is needed and to 'sign off' on computers' work remotely (for awhile, till they're not needed at all anymore)... If the only job left per se for doctors is to 'okay and sign off' on prescriptions that a computer recommends, after a computer does the entire intake/evaluation, then 1 doctor can replace 100 or 1000 for 20 second evals over skype.
  - mfatica 7y ago
    I don't think surgery will be done without human hand for a very long time
    - gremlinsinc 7y ago
      If you consider 2006 a very long time 'ago', then you're absolutely right... https://www.engadget.com/2006/05/19/robot-surgeon-performs-worlds-first-unassisted-operation/ https://www.engadget.com/2006/05/19/robot-surgeon-performs-w... Then there's a more modern article at: https://curiosity.com/topics/a-robot-performed-soft-tissue-surgery-by-itself-curiosity/ https://curiosity.com/topics/a-robot-performed-soft-tissue-s... So, it's definitely possible. Doctors will be replaced probably before software devs, so I think my job is safe(r).
      - jazoom 7y ago
        I'm a doctor and a software developer. I frequently perform surgery. There's no way a computer/robot will be doing what I do as a doctor in the next 10 years. Believe what you will. I don't really have a personal horse in this race, since I wouldn't mind if I was replaced as a doctor/surgeon. I actually took last year off medicine and focused on my tech businesses and it was great.
        gremlinsinc 7y ago
        I have very high hopes for ai/automation. Maybe not 10 years, but can you be sure 20 years is safe? What if your kid started med school in 10 years, and by the time they get through residency they could be replaced by robots? It's a possibility... I think specific 'parts' of medicine are more adaptable to change... PCP's could be 'mostly' automated. I'm not saying it'll be 100%, but say 20% stay around for a decade or two as management overseeing the robotic slaves -- making sure they have > levels of care and lower death rates than a human would by a magnitude big enough to avoid lawsuits. There's not a job in sight that can't someday be done better by a robot. Even software developer. I don't know when, I just know on a long enough timeline it's almost inevitable.
        jazoom 7y ago
        > can you be sure 20 years is safe? I think you're looking at this all wrong. If robots replace me then that means the robots are safer than I am and do a better job. That's awesome. By the time robots have replaced me, they would have long ago replaced accountants and lawyers. We'd be living in a super efficient world. The human element is extremely important in medicine/surgery so we won't be replaced for a little longer. For a long time it'll be there to help us and make our jobs easier. It's like another industrial revolution where we all have more time to... not work. I am not defined by my career. But... I don't see this happening even in the next 20 years. We can't even yet trust cars to drive forward, back, left and right safely with strict rules, markings and signage to guide them.
Daviey 7y ago
What is the Google's long term goal here? Is there a plan to monetise this?
- gremlinsinc 7y ago
  Seriously? Build AI hospital, charge 1/2 the cost across the board, begin replacing doctors/surgeons worldwide w/ google doctor. If google could take home 1/3rd of all GDP spending on medical that would make them the top company in the world.
  - JPLeRouzic 7y ago
    Are not they searching for new area to expand their business? Their turnover is around $150b, if they go in healthcare they could reach $1 trillion. Which other business path can safely offer similar potential?
- NortySpock 7y ago
  Charge $100 for a hospital to send them an image for screening. Google sends back a cancer confidence score.
ashleyn 7y ago
What makes this different from where IBM Watson failed?
- fanf2 7y ago
  It’s going to be an even better fit than Google Flu trends https://www.wired.com/2015/10/can-learn-epic-failure-google-flu-trends/ https://www.wired.com/2015/10/can-learn-epic-failure-google-...
- forgot-my-pw 7y ago
  Watson was mostly a marketing jargon built by IBM. They marketed and sold the product before the technology was ready. Google, Microsoft, Facebook, (and probably soon Amazon) have advanced the AI industry further than IBM at this point. You can predict this by the number of papers released by each company. Stats from 2017: https://medium.com/@chuvpilo/whos-ahead-in-ai-research-insights-from-nips-most-prestigious-ai-conference-df2c361236f6 https://medium.com/@chuvpilo/whos-ahead-in-ai-research-insig...
j7ake 7y ago
Aren't these algorithms trained on expert-annotated data? If so how does it end up performing better than radiologists? Or are these radiologists only mediocre at detecting cancer?
- jazoom 7y ago
  I'm a doctor but not a radiologist. I imagine it would go like this: Computer makes diagnosis on image of lesion that has already been biopsied and has a known correct diagnosis. I don't know how it would work for human false negatives that were never biopsied. Possibly the AI could be trusted enough to make a biopsy on its recommendation.
option_greek 7y ago
One thing for sure is that all these AIs (or their creators) are pretty good at marketing than the actual doctors :).

27 points by jseip 7y ago

It's only a matter of time before all X-Ray, CTScans, etc. use algorithms for secondary if not primary screens.

Having seen many specialist/doctors I generally think they suck at diagnosing, unless its something you could literally discover yourself by a google search and access to a specific test. I don't think they are bad doctors, I just learned that sometimes diagnosis can be rather difficult. Then, you have a computer, it can reference recent studies compare symptoms with hundreds of diseases and generate a test plan accordingly based on probability of a disease or urgency. Frankly, I am surprised not more work is being done with algorithms in medicine. Humans are biased, egoistic and most don't like being told they are wrong.

gremlinsinc 7y ago
I tend to agree, doctors is going to be the holy shit moment for society when we begin replacing them across the board. I mean we still might need 100 or so (nationwide in the usa), maybe keep the top 1000 doctors/surgeons for times when creativity and unique problem solving is needed and to 'sign off' on computers' work remotely (for awhile, till they're not needed at all anymore)... If the only job left per se for doctors is to 'okay and sign off' on prescriptions that a computer recommends, after a computer does the entire intake/evaluation, then 1 doctor can replace 100 or 1000 for 20 second evals over skype.
- mfatica 7y ago
  I don't think surgery will be done without human hand for a very long time
  - gremlinsinc 7y ago
    If you consider 2006 a very long time 'ago', then you're absolutely right... https://www.engadget.com/2006/05/19/robot-surgeon-performs-worlds-first-unassisted-operation/ https://www.engadget.com/2006/05/19/robot-surgeon-performs-w... Then there's a more modern article at: https://curiosity.com/topics/a-robot-performed-soft-tissue-surgery-by-itself-curiosity/ https://curiosity.com/topics/a-robot-performed-soft-tissue-s... So, it's definitely possible. Doctors will be replaced probably before software devs, so I think my job is safe(r).
    - jazoom 7y ago
      I'm a doctor and a software developer. I frequently perform surgery. There's no way a computer/robot will be doing what I do as a doctor in the next 10 years. Believe what you will. I don't really have a personal horse in this race, since I wouldn't mind if I was replaced as a doctor/surgeon. I actually took last year off medicine and focused on my tech businesses and it was great.
      - gremlinsinc 7y ago
        I have very high hopes for ai/automation. Maybe not 10 years, but can you be sure 20 years is safe? What if your kid started med school in 10 years, and by the time they get through residency they could be replaced by robots? It's a possibility... I think specific 'parts' of medicine are more adaptable to change... PCP's could be 'mostly' automated. I'm not saying it'll be 100%, but say 20% stay around for a decade or two as management overseeing the robotic slaves -- making sure they have > levels of care and lower death rates than a human would by a magnitude big enough to avoid lawsuits. There's not a job in sight that can't someday be done better by a robot. Even software developer. I don't know when, I just know on a long enough timeline it's almost inevitable.
        jazoom 7y ago
        > can you be sure 20 years is safe? I think you're looking at this all wrong. If robots replace me then that means the robots are safer than I am and do a better job. That's awesome. By the time robots have replaced me, they would have long ago replaced accountants and lawyers. We'd be living in a super efficient world. The human element is extremely important in medicine/surgery so we won't be replaced for a little longer. For a long time it'll be there to help us and make our jobs easier. It's like another industrial revolution where we all have more time to... not work. I am not defined by my career. But... I don't see this happening even in the next 20 years. We can't even yet trust cars to drive forward, back, left and right safely with strict rules, markings and signage to guide them.

Daviey 7y ago

What is the Google's long term goal here? Is there a plan to monetise this?

gremlinsinc 7y ago
Seriously? Build AI hospital, charge 1/2 the cost across the board, begin replacing doctors/surgeons worldwide w/ google doctor. If google could take home 1/3rd of all GDP spending on medical that would make them the top company in the world.
- JPLeRouzic 7y ago
  Are not they searching for new area to expand their business? Their turnover is around $150b, if they go in healthcare they could reach $1 trillion. Which other business path can safely offer similar potential?
NortySpock 7y ago
Charge $100 for a hospital to send them an image for screening. Google sends back a cancer confidence score.

ashleyn 7y ago

What makes this different from where IBM Watson failed?

fanf2 7y ago
It’s going to be an even better fit than Google Flu trends https://www.wired.com/2015/10/can-learn-epic-failure-google-flu-trends/ https://www.wired.com/2015/10/can-learn-epic-failure-google-...
forgot-my-pw 7y ago
Watson was mostly a marketing jargon built by IBM. They marketed and sold the product before the technology was ready. Google, Microsoft, Facebook, (and probably soon Amazon) have advanced the AI industry further than IBM at this point. You can predict this by the number of papers released by each company. Stats from 2017: https://medium.com/@chuvpilo/whos-ahead-in-ai-research-insights-from-nips-most-prestigious-ai-conference-df2c361236f6 https://medium.com/@chuvpilo/whos-ahead-in-ai-research-insig...

j7ake 7y ago

Aren't these algorithms trained on expert-annotated data? If so how does it end up performing better than radiologists? Or are these radiologists only mediocre at detecting cancer?

jazoom 7y ago
I'm a doctor but not a radiologist. I imagine it would go like this: Computer makes diagnosis on image of lesion that has already been biopsied and has a known correct diagnosis. I don't know how it would work for human false negatives that were never biopsied. Possibly the AI could be trusted enough to make a biopsy on its recommendation.

option_greek 7y ago

One thing for sure is that all these AIs (or their creators) are pretty good at marketing than the actual doctors :).