Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

AI Recognises Race in Medical Images

explainthispaper.com

Post reply on HN
Print view View original

341–350 of 357 posts

Re: AI Recognises Race in Medical Images

#341

Post by goto11 »

JamesBarney wrote: post #247
awild wrote: post #233

Earlier quoted context omitted.

> Maybe these things are social constructs, but if they are then we surely must come to the conclusion that almost everything we care about in the world is a social construct. That's the point really. And it's not inherently a value judgement to say something is a social construct. Using it as a value judgment usually is meant to convey that some secondary attributes are not inherent, or even more so, just historical…

I guess another way to put this when someone says "race is a social construct". What are they trying to communicate about race?

That racial categories and the criteria for deciding which category an individual belongs to depends on history, traditions and politics. There is no objective criteria to decide what constitutes a "race".

Re: AI Recognises Race in Medical Images

#342

Post by oh_sigh »

romwell wrote: post #338
oh_sigh wrote: post #317

Earlier quoted context omitted.

Do you believe that life begins at conception? Or at birth? Or is it somewhere in between? Do you feel comfortable marking the line where life possibly begins? If you mark it before birth, aren't you just giving ammunition to anti-abortion advocates to take away freedom from women? Should we just say that life begins at birth, and shut down anyone who asks otherwise, because anything else is dangerous and possibly ar…

You seem to be confused. What's the continuous variable here that you are measuring to assign the discrete label (alive/not alive) to? If you want to do that based solely on one variable, time from conception, then you get bad science. People can become dead at any point in time. Yes, it's not possible to assign the "alive" label based on time from conception alone . We do have plenty of discrete characteristics base…

Right. A bunch of variables all exist whose values let us ultimately apply the label of alive or not alive. So how can one pick the right set of variables and weight them appropriately to determine if a fetus is alive? Isn't it kind of arbitrary at a certain point? As in, we know when something is definitely not alive (when it is gametes in separate humans), and we know when something is definitely alive (when a baby is born and it is crying), but anywhere between those two is just an arbitrary choice.

You on your own definition of life would apparently want to restrict the rights of women relative to what they have currently. This is a dangerous idea and should not be allowed.

Re: AI Recognises Race in Medical Images

#343

Post by romwell »

oh_sigh wrote: post #342
romwell wrote: post #338

Earlier quoted context omitted.

You seem to be confused. What's the continuous variable here that you are measuring to assign the discrete label (alive/not alive) to? If you want to do that based solely on one variable, time from conception, then you get bad science. People can become dead at any point in time. Yes, it's not possible to assign the "alive" label based on time from conception alone . We do have plenty of discrete characteristics base…

Right. A bunch of variables all exist whose values let us ultimately apply the label of alive or not alive. So how can one pick the right set of variables and weight them appropriately to determine if a fetus is alive? Isn't it kind of arbitrary at a certain point? As in, we know when something is definitely not alive (when it is gametes in separate humans), and we know when something is definitely alive (when a baby…

Yes, it is kind of arbitrary at certain point, that's why we have the debate about it in this country. I am not sure where you are going with this given the subject of the comment I was responding to.

I chose my definition from the rights-of-women perspective. My understanding is that a fetus is generally non-viable outside of a host body, and thus not "alive" by my definition.

On the other hand, we havd C-sections and incubators. If a baby can be safely extracted with a C-section, placed in an incubator, and survive, my understanding is that the choice to abort is no longer available.

Of course, I can be wrong here.

My point was that at least I can make a definition here that does not depend on a choice of an arbitrary value of a continuous variable. My definition of "alive" depends on the choice of which variables to look at, not on arbitrary thresholds. And I don't insist on it being The Truth.

Another example:

"Heart rate" is a continuous vafiable, but the distribution of its values has a large gap between 0 (no heart rate) and nonzero values (the lowest observed was 27bpm). So you don't need to make a choice for a clustering algorithm to work. You can run K-means on "heart rate" and get these 2 clusters: 0 and everything else.

This allows one to make a definition of "alive" based on heart rate. That's not my definition, but it's a usable one.

This is not feasible with race if we use the variables commonly understood to be associated with race: skin color, height, nose shape, etc. There are no gaps in those variables.

Even eye color varies continuously [2], and it's not clear how to assign labels [1].

Color in general is a good analogy for race. You see the colors vary in a rainbow. You can tell the difference between red, green, and blue.

But you can't run K-means on a rainbow to get 7 colors out. Or any number but 1, for that matter.

You need to make a call yourself where to draw those lines.

This reflects in languages. Russian has distinct words for what we'd call "blue" in English; But also English has words like cyan, turquoise, navy, etc., which other languages may not have.

Color, in end, is a social construct.

[1] https://www.edow.com/general-eye-care/eyecolor/

[2]https://udel.edu/~mcdonald/mytheyecolor.html

Re: AI Recognises Race in Medical Images

#344

Post by tpm »

zepto wrote: post #340
tpm wrote: post #322

Earlier quoted context omitted.

My claim is that race is not a meaningful concept.

Do you think that racism is real?

Yes, racism is sadly very real, it is based on perceived ingroup/outgroup differences and as such is only one of such ideologies (I come from a territory where people used to kill each other if they belonged to different branches of christianity, or later if they belonged to different nationality). They are all based on false beliefs and propagated by people who are trying to profit from hate and conflict.

Re: AI Recognises Race in Medical Images

#345

Post by mlnewb2 »

shadowgovt wrote: post #133
hgial wrote: post #95

It might be helpful for folks to look at the blog post written by one of the authors: https://lukeoakdenrayner.wordpress.com/2021/08/02/ai-has-the... or the paper itself https://arxiv.org/pdf/2107.10356.pdf I see a lot of "oh it's probably just picking up on x y z" when x, y, and z are things they explicitly checked for: 1) "It's probably just the names or other metadata" – they only gave it pixel data to train on. T…

One major risk source I see is that the size of the training data for the races isn't the same. For white vs. black patient data, there's between a 2:1 and 3:1 ratio bias in both the training and test data (and a much higher ratio bias for Asian... as high as 20:1 in some of these categories). This gives the CNN more information on one race than another, which can create a classifier that performs very well on the tr…

They tested on tons of different external datasets, and at least one of the training datasets was balanced. Same results were obtained.

Re: AI Recognises Race in Medical Images

#346

Post by mlnewb2 »

abrichr wrote: post #114
hgial wrote: post #112

Earlier quoted context omitted.

It's not "conspicuously absent from the paper". They have a whole group of experiments on this: "Experiments on anatomic and phenotype confounders" and conclude "Race detection is not due to obvious anatomic and phenotype confounder variables."

They show correlation with individual features, not all together.

They show a bunch of non correlations. Have you ever seen a bunch of nothing results add up to an AUC of 95 plus?

Re: AI Recognises Race in Medical Images

#347

Post by mlnewb2 »

shadowgovt wrote: post #136
andi999 wrote: post #92

AIs do not have magical abilities, I do not trust this result. AI can pick up though easily on technical artifacts. Something like a cofactor: Since they used different databases, maybe one dataset had a high number of people of one self declared race and the other the other self declared race; and each using a different intensity maximum or so.

They accounted for specifically intensity maximum, but your overall concern is solid; I don't see anything in the paper suggesting they accounted for a full spectrum of risk factors (broad-image noise, rotation, individual "stuck pixels" that could create a hard-to-spot thumbprint in the image, for example).

Simply testing in multiple external populations already rejects this hypothesis, unless you think they all had the same scanners with the same stuck pixels.

They also tested several variants of noise.

Re: AI Recognises Race in Medical Images

#348

Post by mlnewb2 »

poulpy123 wrote: post #337

AFAIK there is no scientific definition of race so I don't see what could be recognised by an algorithm

In medicine there is an accepted definition which is used by, for example, the national institutes for health in the US. They use that to define "health disparities" between different racial groups.

That definition is what the researchers used here. Race as a social construct, self-reported by the patient.

Re: AI Recognises Race in Medical Images

#349

Post by mlnewb2 »

soundnote wrote: post #331
lostlogin wrote: post #255

Earlier quoted context omitted.

What reality is that? I haven’t seen the thread sorry, have you a link?

That race is a physically real thing that goes beyond skin color and means pervasive differences in all parts of human biology, some meaningful, some not so much. The AI community is very invested in making a colorblind AI that can't or won't use racial characteristics in doing its job, which at least in medicine seems completely idiotic to me. We know that even at the rough proxy level of race that populations somet…

The only Twitter meltdowns were from your ilk. Racist Twitter exploded for some reason. It didn't seem like they (or you) even read the paper. At no point do the authors suggest making AI colorblind.

Re: AI Recognises Race in Medical Images

#350

Post by exporectomy »

vintermann wrote: post #336
exporectomy wrote: post #251

Earlier quoted context omitted.

Skin color isn't race. I would expect this AI to detect your white-skinned friend as black because all his other traits would be black-like.

Well, you would be wrong. Because this AI predicts self-reported race, not ancestry. Think about what this means People who have, say, three black great-grandparents and five white ones, will vary widely in how black they look, and this likely affects how they identify. But there are also invisible markers of ancestry - of the sort you would expect maybe were visible only on an X-ray if you looked carefully. Those co…

You're really missing something obvious here. I don't know if it's the statistical nature of ML or the statistical nature of correlation or what. But the fact that everyone else can see how it probably works while you can't should be a red flag that there's a hole in your understanding somewhere.
Post reply on HN