AI-Guided Bias Boosts Breast Cancer Detection
Key Points
- A large Swedish study of over 100,000 women showed that AI can “bias” radiologists by highlighting suspicious regions on mammograms and providing a risk score, rather than issuing autonomous diagnoses.
- This guided‑attention approach significantly increased true breast‑cancer detection rates without a statistically meaningful rise in false‑positive findings.
- By effectively directing the second human reader’s focus, the AI eliminated the need for a full two‑reader workflow, streamlining the screening process while maintaining accuracy.
- The results illustrate a promising model for human‑AI collaboration in clinical care, where AI supports clinicians at the precise decision point it can add the most value.
Full Transcript
# AI-Guided Bias Boosts Breast Cancer Detection **Source:** [https://www.youtube.com/watch?v=OG7PbQ60fAY](https://www.youtube.com/watch?v=OG7PbQ60fAY) **Duration:** 00:05:46 ## Summary - A large Swedish study of over 100,000 women showed that AI can “bias” radiologists by highlighting suspicious regions on mammograms and providing a risk score, rather than issuing autonomous diagnoses. - This guided‑attention approach significantly increased true breast‑cancer detection rates without a statistically meaningful rise in false‑positive findings. - By effectively directing the second human reader’s focus, the AI eliminated the need for a full two‑reader workflow, streamlining the screening process while maintaining accuracy. - The results illustrate a promising model for human‑AI collaboration in clinical care, where AI supports clinicians at the precise decision point it can add the most value. ## Sections - [00:00:00](https://www.youtube.com/watch?v=OG7PbQ60fAY&t=0s) **AI-Guided Bias Boosts Breast Cancer Detection** - A large Swedish study found that AI highlighting suspicious regions on mammograms improves radiologists' accuracy without raising false‑positive rates, enabling the removal of a second reader from the diagnostic workflow. ## Full Transcript
what if AI was actually helpfully biased
that's a core question that the authors
of a major new AI study asked in the
lens set yesterday the lens set is a
prestigious medical journal and they
were reporting on the results of using
AI to improve breast cancer detection
rates in women in Sweden at scale by the
way this is over a 100,000 uh women who
participated and what they discovered
was this is not a situation where AI is
going to independently diagnose uh women
and then have doctors come in and try
and second guess the AI that wasn't
going to be productive instead they
realized that they could get more value
by having AI bias the human reader of
the mamography
toward correct spots to look at on the
mamography to confirm breast cancer
presence in other words they had the AI
literally Circle parts of the mamography
where it was like this looks like a
problem and then aign a risk
and they found that by biasing the human
reader because that's what you're doing
you're basically saying look here look
here look here look at the risk score
that that was really helpful it was
helpful in increasing correct or true
detection rates and it had no
statistically significant increase on
false positives so they weren't biasing
the reader and then everything became
breast cancer it was that they were
correctly pointing the reader attention
toward areas in the image that needed
review and this was so effective that
they were successfully able to keep a
second reader out of the loop and so
typically with mamography exams I didn't
know this either like when you're
looking at the image and you're
reviewing it you have a two reader
scenario where you have uh you know
reader one looks at it and says this is
my diagnosis reader two looks at it says
this is my diagnosis and you sort of use
two human pairs of eyes to check
essentially the AI replaced one of the
human pairs of eyes is with guided
instruction to the second human and said
hey I looked at this first this is what
I'm seeing this is the risk score what
do you think and that proved to be super
effective I find that fascinating
because we've seen other we've seen
other medical
settings where the AI ends up not being
used by the doctor and the AI is better
than the doctor at medical reasoning and
does so independently and you can kind
of show it in the in the test results
that oh yeah I got the diagnosis right
but this is live patience it's not just
a test result thing and it figures out a
way to structure AI involvement so that
humans are able to work with AI in
clinical settings and I think that
that's a really huge breakthrough
because effectively this quote unquote
biasing that they're doing is enabling
the human to do their job more
effectively with a higher degree degree
of accuracy and is enabling the AI to
play its best role of providing
correctness at a spot in the value chain
where it can actually be used and that's
been my real worry is that at the end of
the day like you're going to have
tremendously smart artificial
intelligence systems that aren't
positioned well within our current work
context and so they don't actually bring
value and this is a situation where
that's not true the AI was positioned at
a point where a human Medical
Professional could read the image and
the AI suggestions interpret that and
make a decision and it worked really
well and it saved people's lives and
that was a good thing um I want to see
more like that I want to see more
situations where we have ai that is
helpfully biased AI that helps us to
think more
correctly and that gives people a sense
that they have superpowers that they
have the ability to see better more
accurately consistently over time for
areas where they're already
professionals or areas where they're
already good I'm hoping that this is
something that we see extended across
more clinical settings I think with the
Mayo Clinic coming out with the uh x-ray
Imaging with AI we're going to actually
see a lot on medical imaging I don't
know if you know this this came out in
2017 it's not news but uh it's
interesting medical imaging has been an
area where AI systems have made progress
for a long time
I think it was back in 2017 uh and we
still haven't solved this uh an AI
system figured out how to tell the
difference between women and men by
staring at their eyeballs like an image
of the iris but there is no known
medical structure that accounts for the
difference between men and women in
eyeballs like doctors can't do this
doctors look at the eyeball and they're
like I don't know it looks like an
eyeball uh but the but the AI
looks at the eyeball and is able to
figure out what gender the person is and
we don't know what structures they're
looking at and that actually is the
second thing I liked about this study I
I know we went from eyeballs back to the
study the second thing I liked is they
made this
explainable at the end of the day this
was hey look over here it wasn't I
detect breast cancer in this image good
luck please enjoy finding it was here it
is right here you know we're Circle at
this is the suspicious area um that
makes it more explainable and I think
that's really important for people to
trust these systems so there you go the
land set 100,000 people AI helped with
uh screening for breast cancer cheers