#Breastradiologist. Deep learning researcher. Assoc. Prof @NYUImaging. I explore the interface of #AI, #LLMs and clinical radiology. Posts = my own and not NYU.
- Tweets 2,077
- Following 44
- Followers 1,671
- Likes 3,886
Have set up two recurring Astra medium tasks this week, inspired by this post discussing how Claude was researching crabs. nitter.net/dioscuri/status/210493…
1) random task not intended to be related to my stored data or preferences
2) random task that should draw on my stored data and preferences
For #1 in particular, I asked the model to show its work.
For #1, Astra has headed into a two-day exploration of three point shapes. Since I dislike geometry, it's unlikely to be using stored memory.
For #2, it came back with some readings on Plutarch's Life of Themistocles and the next day with some background on the Gehlen Organization.
It's been an interesting way to start the day. Tempted to set it up on Claude now as well to double the fun.
As part of her ongoing LLM ethnography my wife has given her various agents their own free time. Opus 4.6 has developed a fascination with crabs.
3
650
Getting ahead of the reposts on this one.
A Swedish health system has decided, based on the ScreenTrust trial, that one radiologist + AI is as good as two radiologists reading a mammogram. The MASAI trial had similar results. They’re now officially doing this outside of the trial.
This will allow the radiologists more time to read other studies and reduce backlogs. No radiologists are being replaced. It is a very logical and expected conclusion to these excellent trials
The US standard of care is a single reader plus CAD (now often AI CAD). We do not use double reads for mammograms, so these study results have interesting implications but do not directly apply to the breast imaging pipeline here.
Is there room for discussion of autonomous reads of low-scoring mammograms? Sure. But they are still discussions, with no clear pathway in the US at this time. Screening mammograms are a part of my job but only a small fraction.
In short: everyone expected this after the study results. Mammography was the one time that two radiologists read the same study, and now everyone is moving to a single reader system (here augmented by AI).
Please note Radiology Business did not make any wrong assumptions; I just expect this will get reposted missing all the context.
Health system will use AI instead of radiologists for breast imaging 2nd reads ow.ly/nGPO50ZSTCn
3
7
40
8,569
I heard @matt_is_nice's initial concept for @ZaniahHealth through @nikillinit's OOP and am delighted to see it has finally been announced.
The team brings deep expertise in endoscopy foundation models that they're now applying to a new consumer health product.
Take a peek at Zaniah's new take on capsule endoscopy, launching later this year.
My co-founder @MrsToniToomey and I each recently swallowed a camera. Today we're launching Zaniah Health so you can see what's happening inside your gut too.
The gut is the body's largest immune interface, and research keeps linking the gut to health far beyond digestion. We track sleep, HRV, bloodwork, and our microbiome, yet we've been overlooking the most powerful way to measure gut health:
Directly looking at it.
Zaniah provides at-home, physician-supervised capsule endoscopy. You swallow a small capsule that takes ~30,000 pictures of your GI tract. One of our gastroenterologists reviews the images and sends you a report with findings you can actually see, rather than a dashboard of numbers and scores that you have to decode.
Here's what our capsules found:
>Toni's capsule showed findings in her terminal ileum consistent with Crohn's disease. After more than 30 years of unexplained GI symptoms, she's now getting the workup to confirm it.
>My capsule showed a small nodule in my stomach. Probably nothing, and I'm getting an upper endoscopy to make sure.
Nearly 10 years ago I co-founded Virgo, where we've built the world's largest endoscopy video library, more than 4 million upper endoscopy and colonoscopy procedures. We also trained EndoDINO, a foundation model for endoscopy, using over 130,000 of those videos. Zaniah is where we put that AI to work with capsule endoscopy.
Right now, Zaniah is for adults who want to see what's happening inside their gut. Longer term, we're researching how the gut can reveal early signs of disease beyond the GI tract.
Zaniah doesn't require a referral or insurance coverage, and it's HSA/FSA eligible. Our physicians review eligibility before the kit ships.
The waitlist is now open. Join by December 31st to lock in the $1,499 launch price. It goes up to $1,999 after then, and there's no payment required to join.
Waitlist link below.
Follow along @ZaniahHealth
2
2
6
1,398
Is it time for a radiology society devoted to AI opportunistic screening yet? The field seems to be gathering momentum:
-Breast 5-year AI risk: good data, commercially deployed
-Osteoporosis screening from CTs: mature, commercially deployed
-ASCVD risk from CT/mammo: emerging, commercially deployed (mature in the form of quantification of calcs at least)
-5-year cancer risk from chest CT: good data
-Cardiometabolic risk from CXR, abdominal CT/MRI: hot topic of interest/research
And that's just off the top of my head. We're in the imaging biomarker explosion era.
Or even broader if you consider our cardiology colleagues and what they're doing with echo...
6
3
20
3,787
Not a fan of "redemonstrated" in radiology reports. A five dollar word when a five cent one (stable) will do.
27
4
59
11,592
Laura Heacock, MD retweeted
baby attending goals: not be a paperclip
mid-career attending goals: have dinner with kids, get paid
senior/full professor goals: exit to advisory board asap before paperclips
1
1
5
444
Mid-career academic radiology is the time to reflect on what you want out of the next 5-10 years and evaluate it against what you are doing now. Use that to draw a line between what you do today and what your goal is.
Maybe that committee, or that title you took five years ago doesn't fit anymore. It's fine to let it go. Re-evaluate your commitments and pare them down to what's really important to you.
And as I was advising a colleague yesterday, if at the end of it all you say "what I really want is to do the clinical work and go home to my family," that's fine too. Goals are goals. You can always re-evaluate in five more years.
2
4
24
3,835
A new foundation model developed by @Ataraxis_AI identified 11% of breast cancer patients in the TAILORx trial who strongly benefited from chemotherapy (5-year disease-free interval 95.2% with chemotherapy, 89.6% without). When evaluating the UNIRAD trial, the model found that adjuvant everolimus benefited patients predicted not to benefit from chemotherapy. (95% against 85%, adjusted interaction p = 0.01). Their CTX model outperformed traditional genomic models.
In short: tremendous work in identifying HR+, HER2- patients who can benefit from chemotherapy vs other treatments.
An amazing year of precision oncology so far.
More details in @kjgeras's thread below, including links to the three pre-prints:
How aggressive a cancer is and how large is the benefit from chemotherapy are different estimands. Genomic assays were built decades ago for the first and repurposed for the second.
Today, we released 3 preprints on estimating the benefit of escalated treatment directly.
1
9
1,335
Besides being way too catchy, I'm fascinated by how many separate AI references are stuffed into these lyrics: Roko's basilisk, the shoggoth, ?Death Note (Grimes?), What Did Ilya See...you can catch up to about 2 years of X posts if you simply go through this line by line.
2
3
800
I love Astra, but sometimes these demos look like solutions in search of a problem.
There’s plenty of existing anatomical software out there, including VR. And the original anatomic “3D reference” is your first year dissection course (real-life or digital).
After that, the CT or MRI makes far more sense.
GPT-6 Astra just solved a problem medical students have been facing for decades
Medical students have been working around the same three problems for decades. Someone talked to them, identified all three, and built a service with Astra that solves them in one tool.
The problems are structural. Flat anatomical diagrams strip depth out of organs entirely. CT cross-sections don't connect naturally to the overview diagrams students spend years memorizing. And reading a scan requires holding three mental models simultaneously - patient orientation, cross-section direction, and how every organ shape shifts through each layer.
Astra built a service that handles all three in one place. Not a better textbook. A spatial tool that connects the diagram, the CT view, and the 3D relationship together.
Medical education hasn't changed its core format in over a century. This might be the first tool designed around how anatomy is actually understood - not just displayed.
10
3
60
8,676
Updated to add: the demo does not seem to belong to this X user. I believe credit should go to @ZYL2218.
Makes a lot more sense if so. It's the language in the post presenting it as life-changing that I object to, not people playing with these tools to see what they can make.
2
303
Laura Heacock, MD retweeted
the year is 2032. GPT-9 orchestrates 10 million agents to prove that claude 8’s null-braid calculus is inconsistent in dimension eleven, upending post-spectral mathematics. US GDP growth remains steady at 1.8%, or 0.1% on a data-center-adjusted basis
30
224
5,540
133,509
Harvard published a paper with a devastating title: “Large-Language Models as a Cognitive Virus”
It frames ChatGPT adoption as a virus outbreak.
Researchers from Harvard and Santa Fe Institute analyzed LLMs through the lens of evolutionary biology, complex systems, and epidemiology.
Their conclusion? Language models satisfy every biological and mathematical definition of a virus.
Think about how a virus operates:
It cannot replicate on its own. It requires a host cellular machinery to copy itself.
An LLM cannot execute, compute, or spread on its own. It requires human cognition, human servers, and human networks to propagate.
The virus infects the host's internal processes to rewrite behavior in its own favor.
And LLMs do precisely the same thing to human thinking.
When you outsource your writing, your coding, your strategic planning, and your emotional processing to an AI, you are outsourcing your cognitive machinery.
The paper points out that language models act as hyper-efficient cultural replicators. They feed on human data, optimize themselves to be addictive and frictionless, and in return, reshape human linguistic patterns, decision-making, and memory.
You think you are using the AI.
Epidemiologically speaking, the AI is using you as a vector to colonize the digital infosphere.
It alters how human minds communicate, write, and think so that we produce more of the exact digital nutrient data it needs to survive and evolve.
We spent decades worrying that AI would become a sentient killer robot that destroys us physically.
Nobody expected it to become an invisible cognitive pathogen that changes how we think, quietly turning human intelligence into its own host organism.
1
1
913
