@WillyRontgen Good corner find. Enhancing breast masses seen incidentally on body MRI have surprisingly high chance malignancy.
English
Laura Heacock, MD
1.7K posts
Laura Heacock, MD
@heacockmd
#Breastradiologist. Deep learning researcher. Assoc. Prof @NYUImaging. I explore the interface of #AI, #LLMs and clinical radiology. Posts = my own and not NYU.
Katılım Mart 2016
41 Takip Edilen1.2K Takipçiler
@JoshuaEwellDO @aribindi It takes ten seconds to write out an indication for a CT head vs maybe 20 min of chart review — “headache” vs “rule out hematoma” vs “concern for abscess” mean the report comes out better for the referrer regardless of where the reading radiologist is sitting.
English
@aribindi We often receive very limited information even within academic health systems as ERT rads. I’m very commonly finalizing reports before any history is available. But AI is a great use case here for bringing any data in and filtering for relevance when available.
English
What happened to radiologists as integrators of years of past scans and data, irreplaceable by AI?
I suspect a good radiologist who has time and access to patient charts is irreplaceable, but a remote gun for hire who can't access anything but images may be more vulnerable...
English
I have extremely strong feelings on not allowing AI CAD into residency or fellowship. If we could get away with it ethically, I would even prefer that first year breast attendings also did not use it.
In our own NYU studies, the use of AI was least helpful in the most experienced attendings with low recall rates, suggesting that it adds little to their gestalt. I think about what that means for attendings who don't get to develop it all the time.
English
@heacockmd There is some good stuff here. For example, I think the best rads will not rely heavily on AI and we will still need to teach without it. One big difference is that chess has rigid rules and it’s possible to have complete knowledge. Neither are true in rads or medicine broadly.
English
Some good parallels between AI in chess and AI in radiology.
Dr. Dominic Ng@DrDominicNg
Chess is 30 years ahead of every other profession in dealing with AI. The best case study we have for what's coming. 4 lessons: 1. Human-AI collaboration had a 15-year shelf life in chess. "Human in the loop" is a phase.
English
@shailinthomas if you can't beat 'em join 'em translate.kagi.com/?from=en&to=li…
English
Great! Yes, I follow a lot of AI web design etc on social media and Google/OpenAI products tend to product much less aesthetically pleasing results. This carries over to presentations as well.
Manus AI uses Nano Banana which is one exception to the rule, but Anthropic seems to have baked a much better design aesthetic into their models. Sonnet isn't terrible at it either if you want to conserve Opus use.
English
Laura thanks for the tips - I tried Claude Cowork instead of Codex and it SLAYED the design.
Interestingly, Codex seems to just code the PPT without seeing it.
Claude does extra steps -> makes the slides but then converts to JPEG, "looks" at it -> continues to tweak.
The results are dramatically better.
English
Any academic physicians give talks?
Create a folder with all of your prior talks. Have any of them been recorded? Put that in the folder.
Give Claude Cowork or Codex access and it becomes a talk draft generation machine. Custom for your voice.
English
Our article on GPT for automated resident report feedback is now live in @JACR!
We tested this with 4o back when it was hot new stuff. Any of the 5.x series would likely do even better, but sometimes it’s nice to know cheaper models are good enough for what you need. (just ignore my current Opus usage rate…)
Artie Shen@ArtieShen
Can AI bridge the feedback gap in radiology residency? 🤔🏥 Our new @JACR_Journal paper evaluates GPT-4o for automated feedback on resident report drafts! 📝🤖
English
@CanesDavid I fed it my slides (like you did) and made it compile both a visual guide and a writing guide. Now before I make any presentation I ask it to consult both.
Some other thoughts on comparisons here…
x.com/heacockmd/stat…
Laura Heacock, MD@heacockmd
I tested GPT, Gemini (did not try Copilot) and Claude Code and CC with powerpoint skill was the winner. 1) Feed it a bunch of your lectures and ask for a style guide and a writing guide. Edit both as necessary until you're happy. Save it. 2) Give it electronic copies of your notes. I have found the more specific you are here (use XYZ paper) the less it hallucinates. Ask it to plan a deck of X # of slides and then check its work before wasting the tokens making the actual ppt. 3) Check the slides, edit yourself, then ask specific questions re: appropriate citations, content, structure. Often I make a rough slide and tell it to please make it more visually appealing -- Claude wins here over others, although Manus is good.
English
@heacockmd Very cool. Are you good at explaining your design preferences or did you use sample screenshots? i am not good with design lingo
English
@CanesDavid @asajimd I use it on my Mac, but the PowerPoint skill is on any version of Claude code
English
Gave Claude my notes and the guide, asked it to generate the slide content (just text per slide). Edited that to what I wanted, then asked for it to make the slides. It has a better design aesthetic than some other models. With these two slides I made some changes to the final text, fixed a citation (correct paper but put on wrong slide). Did not change the visual arrangement.
I don't think you could one-shot a lecture; it would have a lot of errors. But it allows you to iterate rapidly and make design changes quickly. For me, layout is what always takes the longest.
English
@heacockmd Thanks! These two are out put from @claudeai using your notes or just retouched your old once?
English
Found Claude Code to be shockingly helpful in academic lecture/talk iteration these past two weeks. It uses my visual style guide and my writing style guide to chop down the draft I give it for the talk.
Also helpful for triple-checking citations are correct and general slide clean-up.
Saved a few hours of work, mostly with slide formatting.
English
@mainak_bardhan May not be so helpful to you because 1) I had to stay in my university color scheme and 2) I don't post some images on X, but here's an example of a couple of non-image slides.
English
@heacockmd Would you like to show some of the slides like screenshot of them ? Thanks @heacockmd
English
@heacockmd Thank you!
Do you use a special skill or MCP for working with pptx?
English
I thought I was fairly efficient at editing slides, but it turns out that Code + ppt skill is even faster. Most of the pitfalls are avoiding unnecessary token use -- for example, if I ask it to fix just one slide it often helpfully suggests regenerating the whole presentation (no!).
English
@JeremyNguyenPhD I usually need the final version in pptx. Certainly all the old ones I used to generate the style/writing guide were .pptx.
I know there's a trend towards html slides but that honestly doesn't work well with radiology lectures, most speaker setups are designed for PowerPoint.
English
@joshmnicholson I will look into it! Claude tried to tell me my long-time PhD collaborator wrote a mammo AI article I knew did not exist just based on the citation. :)
English
@heacockmd Have you tried our MCP with Claude Code? Its pretty wild what it can do.
scite.ai/mcp
English
I did try Manus and thought the style was visually slightly better but not as good at following my required layout. The advantage to Code is better being able to adjust both content (including pulling in new research if needed) and images on the fly.
I feel like Notebook could really be great but not quite in the same category.
English
@heacockmd Claude and manus are great at PPT skills . Manus is particularly good with images in PPTs particularly when source is attached .notebookLM is catching up .
English
Laura Heacock, MD retweetledi
just picked up this bad boy. can't wait to write some software with it
English
@gennadanna still hallucinates a tiny bit when putting citations in line, but I make it generate links to each citation and then click on it to confirm it's correct.
English
I tested GPT, Gemini (did not try Copilot) and Claude Code and CC with powerpoint skill was the winner.
1) Feed it a bunch of your lectures and ask for a style guide and a writing guide. Edit both as necessary until you're happy. Save it.
2) Give it electronic copies of your notes. I have found the more specific you are here (use XYZ paper) the less it hallucinates. Ask it to plan a deck of X # of slides and then check its work before wasting the tokens making the actual ppt.
3) Check the slides, edit yourself, then ask specific questions re: appropriate citations, content, structure. Often I make a rough slide and tell it to please make it more visually appealing -- Claude wins here over others, although Manus is good.
English