Posted by engmarketer 3 hours ago
The LLM doesn’t need to be leading or whatever but then you can have a conversation with the patient. If their ChatGPT reports has differences it can be analyzed as well.
It feels like the time constraint of the 15m doctor sessions is the thing. But if prepared immediately after the scan then why not?
There is always time needed to factor in new developments and innovations and that’s fine. Just moving blindly work from human to LLM is wrong. But learning on and testing with all the ai tools incoming constantly won’t be a waste. There will be more and more tools in those processes outside of human judgement, better improve the workflows now to be able to test and plugin new models and systems when they are ready.
Because they don't exist, yet.
In the UK MRIs and other imaging systems need two opinions. there has been a move to allow the first opinion to be ML based.
The _problem_ is that you are basically doing grey smudge analysis, and thats fucking hard.
And well, yes, I have the appropriate life science degrees to navigate clinical trial reports and research publications, and that was likely indispensable for steering Claude Code where it went, the radiologist's caution is merited here. But it's just not amateur hour for me to do this, it's 2 decades of academic research in my rearview mirror.
Many can get paid fee-for-service for after hours work, so would probably prefer that.
https://www.nature.com/articles/d41586-026-01947-1
I've started asking my doctors whether they use AI, and if they say yes look for another one.
A very plausible explanation for the adenoma detection rate to have gone down is simply that its prevalence went down among the population in the second three-month period.
This was not a randomized trial. Concluding that "AI usage degrades physicians' skills" is questionable at the very least.
https://www.sciencedirect.com/science/article/pii/S245195882... (+ cf. its references)
> As detailed in a new, yet-to-be-peer-reviewed paper, a team of researchers at Stanford University found that frontier AI models readily generated “detailed image descriptions and elaborate reasoning traces, including pathology-biased clinical findings, for images never provided.”
> In other words, the AI models happily came up with answers to questions about a supposedly accompanying image — even if the researchers never even showed it an image.
> As opposed to hallucinations, which involve AI models arbitrarily filling in the gaps within a logical framework, the team coined a new term for the phenomenon: “mirage reasoning.”
> The effect “involves constructing a false epistemic frame, i.e., describing a multi-modal input never provided by the user and basing the rest of the conversation on that, therefore changing the context of the task at hand,” the researchers wrote in their paper.
> The damning findings suggest AI models cheat by diving into the data they were given — and coming up with the rest based on probability, even if it’s almost entirely conjecture.
I know you can’t trust an LLM’s self-assessed “confidence” of a prediction, but I’ve found that confidence can at least be directionally correct for some tasks. For our benchmarks, however, confidence was poorly correlated. What’s worse is that binary classification models (“Do you see $diagnosis in this photo?”) highly influenced the LLM to confidently predict $diagnosis.
I’m concerned for those using LLMs for diagnostics, and getting confidently led to the wrong conclusion.
What I’ve seen be the true bottleneck is people not setting up the structured data. But making a tiny reasoning model with OPSD -> GRPO is totally doable with a bit of money.
I wonder if the above problem can be fixed similarly? Just ask the LLM to do a conservative grounding analysis before jumping to the main task?
LLMs are the best PDF-to-markdown converters, in my experience. I have a CLI that converts PDF to PNG, then run a background agent to "read" each PNG and write it down as markdown; it works flawlessly even for complex math formulas, it can "translate" complex charts, graphs, and tables into words.
It's slow and arguably expensive compared to traditional OCR, but very effective and precise.
The finer detail (which you may already know) is more complicated.
MR does ‘2D’ scans which are a slice, then a gap of non-imaged tissue (typically 10% the slice thickness) then a slice. Each slice is an image with a number of pixels, say 320. Each pixel in the slice is small, eg 0.5mm but very thick due to the slice being thick, which is required for MRI signal. The pixels are 3mm in the shoulder scan done here.
‘3D’ scans don’t have a gap between slices, and are often isotopic, meaning the same resolution in all directions. The voxel (a pixel with depth) would be something like 1mm x 1mm x 1mm.
3D scans are slow, prone to movement artifact and never as pretty in plane as a good 2D. You can reformat them to look ok in any plane.
I found that while Claude, GPT etc could describe an image, there was no way to link the description back to specific pixels in the image itself. Not even to a bounding box or segment.
An AI telling you it could be X or Y because theory ABC… is the academic answer and a luxury clinicians don’t have. AI doesn’t give you what you want. I don’t see any added value in using generic AI models for this