Luke Oakden-Rayner

@DrLukeOR

PhD student, radiologist, blogger, deep learning in medicine researcher

Adelaide, South Australia
lukeoakdenrayner.wordpress.com
Joined November 2016
68 Photos and videos Photos and videos

Tweets

You blocked @DrLukeOR

Are you sure you want to view these Tweets? Viewing Tweets won't unblock @DrLukeOR

  1. Pinned Tweet
    Jun 5

    To open the black box of decisions, we decided to copy humans and use language. We train a system to mimic the important parts of radiology reports and show that doctors prefer this text over widely used visualisation methods. Blog post:

    7 replies 53 retweets 143 likes
    Undo
  2. Retweeted
    5 hours ago

    For those interested, here is my earlier peer review of the chatbot study undertaken by

    3 replies 35 retweets 39 likes
    Undo
  3. Retweeted
    Oct 31

    Wow, can’t believe my wife pulled this pumpkin carving challenge off!

    7 replies 11 retweets 90 likes
    Undo
  4. Retweeted
    Oct 28

    Sneak preview - How AI can detect breast cancer - Diagnosis on Demand? The Computer Will See You Now - Horizon, BBC Two Thurs 1st Nov, 9pm. Featuring and . Also available on iPlayer.

    2 replies 33 retweets 60 likes
    Undo
  5. Retweeted
    Oct 21

    This 👇

    0:13
    manuscript_submission.mp4
    Show this thread
    1 reply 6 retweets 29 likes
    Undo
  6. Retweeted
    Oct 19

    Summit has been a great venue to understand how RSNA can catalyze the use of AI in for multiple stakeholders. Look for the white paper on this topic coming out in the new Radiology:AI journal in a few months! …

    1 reply 4 retweets 8 likes
    Show this thread
    Show this thread
    Undo
  7. Retweeted
    Oct 19

    , , and 7 others
    4 retweets 5 likes
    Show this thread
    Show this thread
    Undo
  8. Oct 19

    Time to kick off day 2 of 's strategic summit. What better way than with from :)

    1 reply 2 retweets 8 likes
    Undo
  9. Retweeted
    Oct 18

    Kicking off the 2-day AI summit

    3 retweets 17 likes
    Undo
  10. Oct 17

    So far I'm meeting for dinner with at 6. Is anyone else interested? (missed last tweet, not sure who else is on Twitter)

    2 replies 3 likes
    Show this thread
    Show this thread
    Undo
  11. Oct 17

    Phew, after 24 hrs on planes I've arrived in Chicago for the RSNA AI summit! Is anyone catching up tonight? or any others?

    3 replies 17 likes
    Show this thread
    Show this thread
    Undo
  12. Retweeted
    Oct 3

    I'm constantly in awe of our machine learning team at - Here is our recent work on generating synthetic medical imaging data at the highest resolution published to date.

    3 replies 14 retweets 38 likes
    Undo
  13. Retweeted
    Sep 30

    If you’re interested in AI research, this is a must read. It’s a thoughtful commentary explaining recent progress. “Recent developments tell us more about the problems tackled & the structure of our world than about neural networks per se”

    1 reply 58 retweets 158 likes
    Undo
  14. Sep 30

    The biggest statistical problem lies with assuming the human is perfect. Human-model PPV diff is 0.88 in the paper. If the human overcalled one case it would be 0.44, with 2 overcalled it is zero. Assuming a human spec of 90%, at least one overcall would occur ~ 20% of the time.

    8 likes
    Show this thread
    Show this thread
    Undo
  15. Sep 30

    To be clear, since it always comes up, I actually like the study. The effort required to apply a model like this in a clinical environment is no joke. But I'm skeptical of the numbers, and more skeptical of the discussion. Not convinced. Like most topics, more research needed.

    1 reply 9 likes
    Show this thread
    Show this thread
    Undo
  16. Sep 29

    But et al (JAMA, 2016) found individual ophthalmologists to have a specificity around 92% (compared against a panel of experts), exactly the same as the model in this study. So humans probably have a PPV of 12% as well!

    2 replies 1 retweet 10 likes
    Show this thread
    Show this thread
    Undo
  17. Sep 29

    The baseline in this case was a single expert interpretation, which precludes any assessment of how humans might perform at the task. Since humans are the baseline, they are assumed to be perfect.

    1 reply 8 likes
    Show this thread
    Show this thread
    Undo
  18. Sep 29

    In this paper specificity was 92% (pretty good), and PPV was 12% (sounds bad, right?). But even if the specificity was 99%, the PPV would still have been only 50%! That is the problem with low prevalence disease, and screening in general.

    1 reply 1 retweet 6 likes
    Show this thread
    Show this thread
    Undo
  19. Sep 29

    Interesting facts about this study: Sensitivity could have been one of three values. 0%, 50%, or 100%. They found it was 100%.

    1 reply 5 likes
    Show this thread
    Show this thread
    Undo
  20. Sep 29

    Is anyone else a bit flabbergasted that we have been heavily discussing a study with results based on a cohort which only contains 2 patients with referable retinopathy?

    Primary care doctors/nurses used a deep learning algorithm prospectively to help diagnose diabetic retinopathy. It didn't work very well. by
    2 replies 19 retweets 61 likes
    Show this thread
    Show this thread
    Undo
  21. Sep 22

    About to kick off the course on in Paris. I'm up first, with a light-hearted introduction to modern AI.

    2 replies 7 retweets 40 likes
    Undo

@DrLukeOR hasn't Tweeted yet.

Loading seems to be taking a while.

Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.

    You may also like

    ·