Thread by @iScienceLuvr

How does GPT-4 do in the medical domain?

I got to play around with its multimodal capabilities on some medical images!

Plus a recent Microsoft paper examined its text understanding and got SOTA results on USMLE medical exams!

A quick thread ↓

As I showed earlier, I had the chance last week to play around with GPT-4's multimodal capabilities:

I also tried some medical images too! Here I started with some histopathology. I passed in an H&E image of prostate cancer and asked GPT-4 to describe it. It knew it was an H&E image of glandular tissue but was unable to identify it as low grade prostate cancer.

Here I passed in an image of invasive lobular carcinoma with characteristic single file lines of tumor nuclei. It fails to notice this unfortunately not matter how hard I try.

Here is an example of a glioblastoma (severe brain tumor). It has a characteristic feature again that suggests the glioblastoma diagnosis (pseudopalisading necrosis) but it fails to notice that. It does realize the presence of what looks like tumor nuclei.

This image shows H&E of basal cell carcinoma (skin cancer). GPT-4 notices that it is of skin but cannot identify the pathology.

Overall though, GPT-4 mostly refuses to provide anything similar to a diagnosis. Here is one such example with and X-ray image.

My conclusion on the multimodal side is that GPT-4 is a impressive first step towards multimodal medical understanding, but its understanding right now is fairly rudimentary, and there is a lot of room to improve here.

On the text side of things, however, the situation is different. In a recent paper from Microsoft Research, "Capabilities of GPT-4 on Medical Challenge Problems", GPT-4 obtains SOTA on USMLEs (medical student exams), significantly outperforming GPT 3.5.

Other benchmark datasets were tested as well, with GPT-4 again reaching SOTA for most of them.

This was all done without any sophisticated prompting techniques, as shown here

One may worry the high performance is due to data contamination. Interestingly this paper performed a memorization analysis, and they didn't find any of the tested USMLE questions with their memorization detection (though it doesn't 100% confirm no memorization).

Plus the USMLE material is behind paywall and probably unlikely to be in the GPT4 training set anyway.

Overall, seems the medical understanding of text-only GPT-4 is significantly improved & multimodal GPT-4 has rudimentary understanding.

Many more experiments should be done to study GPT-4's medical knowledge/reasoning. Some previous studies using GPT-3 concluded domain/task-specific fine-tuned model are better, and I wonder if the conclusion changes now with GPT-4.

#MedTwitter #PathTwitter

If you like this thread, please share!

Consider following me for AI-related content! → @iScienceLuvr