‘Embarrassingly simple’ probe finds AI in medical image diagnosis ‘worse than random’
‘Embarrassingly simple’ probe finds AI in medical image diagnosis ‘worse than random’
You're viewing a single thread.
We have models that are specifically made to be good at these kinds of tasks. Why would you choose the ones that aren't and then make generalizing claims about how AI sucks in this domain?
Yeah this is probably just straight up misinformation. By no means is a diagnosis going to be made by a generalist multimodal LLM. Diagnosis is a literally a binary classification (although that is an oversimplification) and on medical CV you are optimizing on that directly.
They did not use a LLM.
In a recent experiment, they set out to determine how reliable LMMs are in medical diagnosis — asking both general and more specific diagnostic questions — as well as whether models were even being evaluated correctly for medical purposes.
Curating a new dataset and asking state-of-the-art models questions about X-rays, MRIs and CT scans of human abdomens, brain, spine and chests, they discovered “alarming” drops in performance.
You've quoted them stating they used LLMs while claiming they did not use a LLM? What am I missing here?
What am I missing here?
"L" "M" "M"
Which in this context just means multimodal LLM, correct?
Correct.
large language models (LLM) vs. large multi-modal models (LMM)
Regardless, they both use an LLM as the main driver. Multi modal just means that the LLM is interfaced with generative and/or predictive AIs for other types of content like images, sound, video, etc.
This is using a generalist tool for a specialized job. I'd expect the limit for LMMs is telling you if your picture is a heart or a kidney... Maybe. With low accuracy. Diagnosing? lol, hell no.
models including GPT-4V and Gemini Pro
What a joke, a few generic LLMs making a judgement call about all AI models.
They used one to create the dataset for their experiments:
In their experiments, they introduced a new dataset, Probing Evaluation for Medical Diagnosis (ProbMed), for which they curated 6,303 images from two widely-used biomedical datasets. These featured X-ray, MRI and CT scans of multiple organs and areas including the abdomen, brain, chest and spine.
GPT-4 was then used to pull out metadata about existing abnormalities, the names of those conditions and their corresponding locations. This resulted in 57,132 question-answer pairs covering areas such as organ identification, abnormalities, clinical findings and reasoning around position.
The seven models tested included GPT-4V, Gemini Pro and the open-source, 7B parameter versions of LLaVAv1, LLaVA-v1.6, MiniGPT-v2, as well as specialized models LLaVA-Med and CheXagent. These were chosen because their computational costs, efficiencies and inference speeds make them practical in medical settings, researchers explain.
It seems like this is a case of "they just aren't using AI right, if they used it right it works" when it sure looks like they are using the models intended for these specific medical tasks.
Those are not the sort of model anybody in the field would use (medical CV with deep learning based analysis is a vibrant field with many breakthroughs in recent years). These are the sort of models tech bros are trying to sell to the public as general AI. There is a world of difference.
Not defending this article, but companies & big tech are generalizing the crap out of AI right now, and forcing it into everything.
They could have (and definitely should've) promoted the strengths and weaknesses of their models, specifically regarding what it can and can't do. But they don't. They get more money when their shareholders & customers think it's the next best thing for everything.