Demo of gemini 2.0 reading a CT scan, opinions?

[Doc mu]

looks pretty promising, pretty terrifying how this will become 5-10 years from now seeing how AI is still in its infancy

---

Members don't see this ad.

 

[BadRadRedemption]

I'm just going to say that it is impressive how AI can see the CT scan. That being said let's get into detail what is wrong here.

For starters CT isn't needed at all to diagnose pancreatitis. Abdominal pain and increased pancreatic enzymes are enough to give it a working diagnosis. CT is used 72h after the onset of symptoms to evaluate the damage done by the pancreatitis as well as to determine any complications. This is how it is done in my academic center with a large ER departmant. Usually the patient gets an ultrasound exam of the abdomen on admission before a CT is even considered. A CT scan is used only if the patient is in a very poor condition and has an ambigous presentation or if the ED/GI/AbdSurgeon is feeling frisky.

This exam also isn't done in proper phase. This looks like a portal venous phase which isn't sensitive enough to evaluate the pancreatic tissue well. It's doable but usually when doing an initial CT exam for pancreatitis, you need non-contrast, arterial and portal venous phase to full evaluate the pancreas. This differs from institution to institution but arterial is the most sensitive.

When he scrolls, the AI takes a while to identify the structures. This is just a minor issue because it will get faster but it still is significatly slower than a competent radiologist. It also missed the fluid near the liver and the paracolic gutters. The initial differential that it gave was very very vague "inflamation or injury" that it could easily be a hallucination. I see no signs of injury on that CT scan and the way fluid is spreading and is located is very clearly a text book pancreatitis. The guy in the video had to point towards the pancreas a lot with his cursor and guided thr AI "is anything wrong with the pancreas?" before it even considered that it may be enlarged.

The way it talked about the complications is very bad as well. It said the complications are necrosis(fine), pseudocysts (why not just say collections? Pseudocysts are a very specific collection) and infection (this is fair but CT cannot fully evaluate infections unless you see air inclusions in the collections or abcesses). It didnt mention many more complications of pancreatitis which made me feel as If I was listentning to a student buying himself time to figure out the answers.

This is a bit nit picky, but it also didnt mention hyperdense content in the stomach which isn't a normal finding even on routine scans unless the patient drank contrast.

This is just my two cents. I'm impressed that we have an AI that can do this and what it can do now definitly surpasses the ability of non radiologists and lay people but it still has lots to improve. I'm also just an R2 resident so my knowledge might not be the best. If I'm worng or have missed something, please let me know.

 

[clutch21]

Here we go again.. like clockwork, another AI post showing a whole lot of nothing. This is a medical student level case of acute pancreatitis performed with good CT technique, ideal body habitus, and no motion.

We’ve had AI tools that can recognize findings for sometime now. Take the PE algorithm for an example. It works pretty well most the time. How about when it confuses a PE with motion or streak artifact?

Radiology is straightforward until it isn’t. That’s where AI still falls miserably short.

 

[Radology]

The first time I saw our AI software correctly detect a subdural hemorrhage I was pretty impressed

Then I saw it call 10 straight subdural hemorrhages that did not exist

AI is the computerized version of the best bull****ter in your med school class. Act super confident to a layman, and you will always sound impressive

 

[Cognovi]

Case report