Learning Library

← Back to Library

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
0:00what if AI was actually helpfully biased 0:02that's a core question that the authors 0:04of a major new AI study asked in the 0:07lens set yesterday the lens set is a 0:09prestigious medical journal and they 0:11were reporting on the results of using 0:13AI to improve breast cancer detection 0:16rates in women in Sweden at scale by the 0:18way this is over a 100,000 uh women who 0:21participated and what they discovered 0:23was this is not a situation where AI is 0:28going to independently diagnose uh women 0:30and then have doctors come in and try 0:32and second guess the AI that wasn't 0:34going to be productive instead they 0:37realized that they could get more value 0:39by having AI bias the human reader of 0:44the mamography 0:46toward correct spots to look at on the 0:49mamography to confirm breast cancer 0:52presence in other words they had the AI 0:54literally Circle parts of the mamography 0:56where it was like this looks like a 0:57problem and then aign a risk 1:01and they found that by biasing the human 1:05reader because that's what you're doing 1:06you're basically saying look here look 1:08here look here look at the risk score 1:11that that was really helpful it was 1:12helpful in increasing correct or true 1:15detection rates and it had no 1:17statistically significant increase on 1:20false positives so they weren't biasing 1:23the reader and then everything became 1:24breast cancer it was that they were 1:26correctly pointing the reader attention 1:28toward areas in the image that needed 1:31review and this was so effective that 1:33they were successfully able to keep a 1:37second reader out of the loop and so 1:40typically with mamography exams I didn't 1:42know this either like when you're 1:43looking at the image and you're 1:44reviewing it you have a two reader 1:46scenario where you have uh you know 1:47reader one looks at it and says this is 1:49my diagnosis reader two looks at it says 1:51this is my diagnosis and you sort of use 1:54two human pairs of eyes to check 1:56essentially the AI replaced one of the 1:58human pairs of eyes is with guided 2:02instruction to the second human and said 2:04hey I looked at this first this is what 2:06I'm seeing this is the risk score what 2:08do you think and that proved to be super 2:10effective I find that fascinating 2:12because we've seen other we've seen 2:15other medical 2:17settings where the AI ends up not being 2:21used by the doctor and the AI is better 2:24than the doctor at medical reasoning and 2:27does so independently and you can kind 2:29of show it in the in the test results 2:31that oh yeah I got the diagnosis right 2:33but this is live patience it's not just 2:36a test result thing and it figures out a 2:40way to structure AI involvement so that 2:42humans are able to work with AI in 2:44clinical settings and I think that 2:47that's a really huge breakthrough 2:48because effectively this quote unquote 2:50biasing that they're doing is enabling 2:55the human to do their job more 2:58effectively with a higher degree degree 3:00of accuracy and is enabling the AI to 3:05play its best role of providing 3:08correctness at a spot in the value chain 3:11where it can actually be used and that's 3:13been my real worry is that at the end of 3:15the day like you're going to have 3:17tremendously smart artificial 3:19intelligence systems that aren't 3:22positioned well within our current work 3:25context and so they don't actually bring 3:28value and this is a situation where 3:29that's not true the AI was positioned at 3:32a point where a human Medical 3:34Professional could read the image and 3:36the AI suggestions interpret that and 3:38make a decision and it worked really 3:40well and it saved people's lives and 3:42that was a good thing um I want to see 3:46more like that I want to see more 3:48situations where we have ai that is 3:51helpfully biased AI that helps us to 3:53think more 3:55correctly and that gives people a sense 3:59that they have superpowers that they 4:01have the ability to see better more 4:03accurately consistently over time for 4:05areas where they're already 4:06professionals or areas where they're 4:07already good I'm hoping that this is 4:09something that we see extended across 4:11more clinical settings I think with the 4:13Mayo Clinic coming out with the uh x-ray 4:15Imaging with AI we're going to actually 4:17see a lot on medical imaging I don't 4:20know if you know this this came out in 4:212017 it's not news but uh it's 4:24interesting medical imaging has been an 4:26area where AI systems have made progress 4:28for a long time 4:30I think it was back in 2017 uh and we 4:32still haven't solved this uh an AI 4:36system figured out how to tell the 4:38difference between women and men by 4:41staring at their eyeballs like an image 4:44of the iris but there is no known 4:47medical structure that accounts for the 4:49difference between men and women in 4:50eyeballs like doctors can't do this 4:53doctors look at the eyeball and they're 4:54like I don't know it looks like an 4:56eyeball uh but the but the AI 5:00looks at the eyeball and is able to 5:02figure out what gender the person is and 5:05we don't know what structures they're 5:06looking at and that actually is the 5:08second thing I liked about this study I 5:10I know we went from eyeballs back to the 5:11study the second thing I liked is they 5:15made this 5:16explainable at the end of the day this 5:18was hey look over here it wasn't I 5:21detect breast cancer in this image good 5:24luck please enjoy finding it was here it 5:28is right here you know we're Circle at 5:29this is the suspicious area um that 5:32makes it more explainable and I think 5:34that's really important for people to 5:35trust these systems so there you go the 5:38land set 100,000 people AI helped with 5:41uh screening for breast cancer cheers