4 ms·

AI recognition of patient race in medical imaging: a modelling study

13 points by hunglee2 4y ago

JaceLightning 4y ago
What's amazing is they were even able to determine race "even from corrupted, cropped, and noised medical images, often when clinical experts cannot" and after accounting for differences in body-mass index, disease distribution, and breast density. I feel the profoundness of the study was lost on the authors. They said it's a warning not to create racist AI's. But the fact that human physiology varies among races, and that change is undetectable by people, is incredible. That could imply that certain groups of people are more responsive to certain treats--and AI could help detect which treatment works best for you.
Vecr 4y ago
This is very interesting. Could a cluster analysis using this AI combined with genetics be used to create a new science of race, instead of the dubious pseudo-scientific theories of race from before the 1940s? It might help racialists actually get published, instead of laughed at like they are now.
- heyitsguay 4y ago
  Well, it's probably more like "the social concept of race is roughly correlated with certain phenotype clustering patterns, and those phenotypes extend beyond those visible on people's exteriors", which is mildly interesting, not that surprising, but good for stirring outrage if you pitch it to the right crowds. I've never heard "racialists" used by anyone but white supremacists, so presumably this is one more thing they can use in bad faith to promote their BS.
- dekhn 4y ago
  I don't understand where you're going with this. We already have a science of genetic history (which is what a lot of people sort of mean when they say race) that is based on whole genome sequencing. Medical images aren't going to help you much, except as a phenotypic readout. Also Im not sure why you want "racialists" getting published, there is nothing interesting going on scientifically.
Imnimo 4y ago
Am I misunderstanding Figure 1? It looks like even with a very aggressive high-pass filter (see HPF-50 in Figure 2), the model is still extremely accurate. That feels like a red flag to me - it makes me think there is some sort of leakage in the imaging that isn't being accounted for. It seems so implausible that you could tell race from the HPF-50 image that I'm worried I've completely misread the figure.
- heyitsguay 4y ago
  Agreed, though it could be one of those things where Fourier weirdness and neural network weirdness interact in counterintuitive ways. It'd be easy enough to perform the same experiment with some benchmark image classification dataset, do it with a few and you'd have good context for evaluating whether Figure 1 is an indication of leakage or not.

13 points by hunglee2 4y ago

What's amazing is they were even able to determine race "even from corrupted, cropped, and noised medical images, often when clinical experts cannot" and after accounting for differences in body-mass index, disease distribution, and breast density. I feel the profoundness of the study was lost on the authors. They said it's a warning not to create racist AI's. But the fact that human physiology varies among races, and that change is undetectable by people, is incredible. That could imply that certain groups of people are more responsive to certain treats--and AI could help detect which treatment works best for you.

Vecr 4y ago

This is very interesting. Could a cluster analysis using this AI combined with genetics be used to create a new science of race, instead of the dubious pseudo-scientific theories of race from before the 1940s? It might help racialists actually get published, instead of laughed at like they are now.

heyitsguay 4y ago
Well, it's probably more like "the social concept of race is roughly correlated with certain phenotype clustering patterns, and those phenotypes extend beyond those visible on people's exteriors", which is mildly interesting, not that surprising, but good for stirring outrage if you pitch it to the right crowds. I've never heard "racialists" used by anyone but white supremacists, so presumably this is one more thing they can use in bad faith to promote their BS.
dekhn 4y ago
I don't understand where you're going with this. We already have a science of genetic history (which is what a lot of people sort of mean when they say race) that is based on whole genome sequencing. Medical images aren't going to help you much, except as a phenotypic readout. Also Im not sure why you want "racialists" getting published, there is nothing interesting going on scientifically.

Imnimo 4y ago

Am I misunderstanding Figure 1? It looks like even with a very aggressive high-pass filter (see HPF-50 in Figure 2), the model is still extremely accurate. That feels like a red flag to me - it makes me think there is some sort of leakage in the imaging that isn't being accounted for. It seems so implausible that you could tell race from the HPF-50 image that I'm worried I've completely misread the figure.

heyitsguay 4y ago
Agreed, though it could be one of those things where Fourier weirdness and neural network weirdness interact in counterintuitive ways. It'd be easy enough to perform the same experiment with some benchmark image classification dataset, do it with a few and you'd have good context for evaluating whether Figure 1 is an indication of leakage or not.