4 ms·

My SO is a dermatologist. Her father is also a radiologist, which is probably the most obvious target for this sort of thing. When I asked her about it, she sai

My SO is a dermatologist. Her father is also a radiologist, which is probably the most obvious target for this sort of thing. When I asked her about it, she said that she's looking at the depth and height of things a lot, a dimension not really captured in a photograph. This study pit their AI against dermatologists getting the same photo and information, but I'm curious how it'd do against an in-person examination. Google doesn't appear to be claiming that they could replace dermatologists with AI entirely. Lest anyone's mind start leaping in that direction, here are a few more reasons why I think her job is safe for the foreseeable future. - Much of her work is full-body skin-checks. How able would an AI system be to provide instant feedback on everything sketchy it sees? How comfortable will people be with, rather than a doctor inspecting their entire body, having pictures taken of their entire body? - Her cases can get rather complicated. Skin problems can be caused or complicated by internal illnesses that need to be considered. This point, in particular, calls on her specialized training beyond what a GP would have. - There's a lot more to her work than diagnosing. She's also prescribing treatments, monitoring progress, performing surgeries, etc. - Her time is indeed very expensive (much to by benefit), and more so than a GP's. But there are a lot of challenges involved in deploying things in medicine, and our regulations are probably unprepared for diagnosis-by-computer. I expect it'll be quite a while before, say, a GP augmented by an AI is price-competitive with a board-certified dermatologist.

catoc 7y ago
Dermatologists and radiologists will not be replaced by AI any time soon - for precisely these reasons. Geoffrey Hinton personally saying "we should stop training radiologists" just reflects a complete lack of understanding about what a radiologist does.
- ska 7y ago
  Your broad point is true, but radiology is an odd and relatively new specialty. It's not going away entirely any time soon but I think will be the most radically affected over the next 20 odd years.
  - catoc 7y ago
    It will indeed radically change; and radiologists not-using-AI will be replaced by those who do. However the profession itself will only increase in influence as the radiologists become more and more clinically involved. If they take on the responsibility of employing and regulating medical imaging AI, more radiologists rather than fewer will be needed.
    - ska 7y ago
      I think that "if" is fairly large, and hasn't been resolved yet. I can also see a future with more and more imaging work being done by not-radiologists or hyphenated ones. I guess we'll see. Bear in mind that there have been some effective ML approaches in radiology (for certain tasks) for decades now and the radiologists themselves have been among the largest hurdles to getting them into clinical use. How they respond to broader and more powerful incursions into "their" space will probably have a significant effect on what their future role(s) and scope are. I guess what I'm saying is that we can predict a big growth in the consumption of imaging (and other signal) data by clinicians in the next decades, but it is far less clear if the percentage of that mediate through radiology will remain anywhere near as high as it is now.
      - catoc 7y ago
        I certainly agree to your comment on that "if". Up to now there has been a lot of talk and very little actual action. Although I don't think that it's because of conservative protective resistance - at least not in academia. The main issue is not knowing where and how to start. We'll see indeed who will 'win'. Let's hope it's not the Amazon's and Google's of this world
ska 7y ago
I think the best first target is augmenting specialists, not replacing them. But the problems are cultural much more than technical. Recent improvements in algorithmic performance shouldn't blind people to the fact that for decades now ML approaches have been able to beat average specialists for at least some specific tasks, but uptake of CAD systems and the like has been slowed more by culture than V&V of these approaches.
jszymborski 7y ago
I think automating diagnosis is the least effective and highest risk approach when it comes to creating models for the clinic. The real value for ML in the clinic is to create or refine signals for physicians from which they might better make a diagnosis. Instead of a model rolling dice and saying "I think this is a malignant nevi", it'd be far more useful for it to say something like "here are 10 other nevi in your hospital which I think are similar and share so-and-such characteristics that ultimately resulted in negative outcomes" or "here's a distribution of likely outcomes I foresee occurring, here are the aspects of the lesions that lead me to believe so". ML in the clinic (and in most domains imho) ought to be a tool to increase consistency and efficiency of practitioners, not replace them.
doctoring 7y ago
How comfortable will people be with, rather than a doctor inspecting their entire body, having pictures taken of their entire body? Certain centers do use whole body imaging/photography, e.g. this thing: https://www.canfieldsci.com/imaging-systems/vectra-wb360-imaging-system/ https://www.canfieldsci.com/imaging-systems/vectra-wb360-ima... I expect it'll be quite a while before, say, a GP augmented by an AI is price-competitive with a board-certified dermatologist. I'm curious about whether you could just send a (de-identified) photo with maybe some additional basic info to a web service that then spit back a differential. I don't know about costs (sounds relatively low cost to run, maybe a subscription like UpToDate?), reimbursement, or whether this suddenly would be a regulated device, but if I were a GP I would love something like that to help me -- it would make me faster and better at my job. I don't need to compare to board-certified dermatologists.
MauranKilom 7y ago
I have previously worked with a group that used photometric stereo to incorporate 3D texture information into the process. See for example this paper: https://www.researchgate.net/publication/269051304_3D_Skin_Texture_Analysis_A_Neural_Network_and_Photometric_Stereo_Perspective https://www.researchgate.net/publication/269051304_3D_Skin_T... Didn't keep up with the clinical research that followed though.
mantoto 7y ago
You shouldn't see it like this (competition) at all. There are enough people in the world which do not get diagnosed at all (third world, poor Americans, etc.). Your SO makes mistakes as every doctor does. If the time of your SO is expensive then it means there is a high demand or and high skill required. So we already don't have enough where your SO works. Imagine your SO being able to save, help even more people. This type of large scale medical optimization will help so many people around the world, let's hope together Google can make leaps quickly.

catoc 7y ago

Dermatologists and radiologists will not be replaced by AI any time soon - for precisely these reasons. Geoffrey Hinton personally saying "we should stop training radiologists" just reflects a complete lack of understanding about what a radiologist does.

ska 7y ago
Your broad point is true, but radiology is an odd and relatively new specialty. It's not going away entirely any time soon but I think will be the most radically affected over the next 20 odd years.
- catoc 7y ago
  It will indeed radically change; and radiologists not-using-AI will be replaced by those who do. However the profession itself will only increase in influence as the radiologists become more and more clinically involved. If they take on the responsibility of employing and regulating medical imaging AI, more radiologists rather than fewer will be needed.
  - ska 7y ago
    I think that "if" is fairly large, and hasn't been resolved yet. I can also see a future with more and more imaging work being done by not-radiologists or hyphenated ones. I guess we'll see. Bear in mind that there have been some effective ML approaches in radiology (for certain tasks) for decades now and the radiologists themselves have been among the largest hurdles to getting them into clinical use. How they respond to broader and more powerful incursions into "their" space will probably have a significant effect on what their future role(s) and scope are. I guess what I'm saying is that we can predict a big growth in the consumption of imaging (and other signal) data by clinicians in the next decades, but it is far less clear if the percentage of that mediate through radiology will remain anywhere near as high as it is now.
    - catoc 7y ago
      I certainly agree to your comment on that "if". Up to now there has been a lot of talk and very little actual action. Although I don't think that it's because of conservative protective resistance - at least not in academia. The main issue is not knowing where and how to start. We'll see indeed who will 'win'. Let's hope it's not the Amazon's and Google's of this world

ska 7y ago

I think the best first target is augmenting specialists, not replacing them. But the problems are cultural much more than technical. Recent improvements in algorithmic performance shouldn't blind people to the fact that for decades now ML approaches have been able to beat average specialists for at least some specific tasks, but uptake of CAD systems and the like has been slowed more by culture than V&V of these approaches.

jszymborski 7y ago

I think automating diagnosis is the least effective and highest risk approach when it comes to creating models for the clinic. The real value for ML in the clinic is to create or refine signals for physicians from which they might better make a diagnosis. Instead of a model rolling dice and saying "I think this is a malignant nevi", it'd be far more useful for it to say something like "here are 10 other nevi in your hospital which I think are similar and share so-and-such characteristics that ultimately resulted in negative outcomes" or "here's a distribution of likely outcomes I foresee occurring, here are the aspects of the lesions that lead me to believe so". ML in the clinic (and in most domains imho) ought to be a tool to increase consistency and efficiency of practitioners, not replace them.

doctoring 7y ago

How comfortable will people be with, rather than a doctor inspecting their entire body, having pictures taken of their entire body? Certain centers do use whole body imaging/photography, e.g. this thing: https://www.canfieldsci.com/imaging-systems/vectra-wb360-imaging-system/ https://www.canfieldsci.com/imaging-systems/vectra-wb360-ima... I expect it'll be quite a while before, say, a GP augmented by an AI is price-competitive with a board-certified dermatologist. I'm curious about whether you could just send a (de-identified) photo with maybe some additional basic info to a web service that then spit back a differential. I don't know about costs (sounds relatively low cost to run, maybe a subscription like UpToDate?), reimbursement, or whether this suddenly would be a regulated device, but if I were a GP I would love something like that to help me -- it would make me faster and better at my job. I don't need to compare to board-certified dermatologists.

MauranKilom 7y ago

I have previously worked with a group that used photometric stereo to incorporate 3D texture information into the process. See for example this paper: https://www.researchgate.net/publication/269051304_3D_Skin_Texture_Analysis_A_Neural_Network_and_Photometric_Stereo_Perspective https://www.researchgate.net/publication/269051304_3D_Skin_T... Didn't keep up with the clinical research that followed though.

mantoto 7y ago

You shouldn't see it like this (competition) at all. There are enough people in the world which do not get diagnosed at all (third world, poor Americans, etc.). Your SO makes mistakes as every doctor does. If the time of your SO is expensive then it means there is a high demand or and high skill required. So we already don't have enough where your SO works. Imagine your SO being able to save, help even more people. This type of large scale medical optimization will help so many people around the world, let's hope together Google can make leaps quickly.