[bearsnowstorm]

I own the 1st gen of the Butterfly - in my opinion, it wasn't great image-wise compared with the contemporary conventional crystal based probes (thinking of cart-based machines with less flexible, more expensive probes etc so perhaps an unfair comparison). Would be cool if the newest ones mentioned in the article are becoming comparable with the crystal based probes - I can't comment. But I can say image quality is absolutely key. There are lots of cool AI based applications coming out all the time (I know much more about echocardiography AI than the foetal ultrasound AI mentioned in the article, but this is a similar paper where some ultrasound novices had AI guidance and were able to obtain useful echo images https://www.ahajournals.org/doi/10.1161/CIRCIMAGING.123.0155...). I get to use various machine vision based tools on echo images at work to automate various measurements - but at the moment, I find they fail badly if the imaging is anything but great quality, whereas humans can interpret them. Maybe future training sets will include more "technically difficult studies" (code for poor imaging) and AI tools will do better than they do now? Or there will be more augmentation of data sets with realistically degraded versions of images to add robustness? AI that worked on suboptimal images would be awesome, particularly in my setting (ICU).

[dotancohen]

Is there really room for AI in medical imaging?

Medical imaging is not about getting the expected result as is typical in most bodies, rather, it is about getting the actual result as is found in this body.

[mikepurvis]

The applications I've seen are around analysis/alerting. Submit your raw data to a service and it screens for hundreds of rare conditions your doctor's never heard of.

Similar to blood work maybe nothing comes back conclusive but it might be interesting enough to trigger follow on investigation.

But the general theme is around augmenting the human decision maker rather than "preprocessing" the data in a way that might obscure or hallucinate important details.

[faeriechangling]

False positives on tests are a well understood source of illness and iatrogenic harm, as well as being a well known money printer for those administering the tests. Trying to screen for a whack of rare diseases with AI still sounds dubious to me unless you have good reason to suspect one beyond high levels of medical anxiety.

I’m more interested in AIs ability to do things like read x-rays and identify illnesses with higher accuracy than a tired doctor on a 12 hour shift.

[adolph]

> Is there really room for AI in medical imaging?

Sure, there are many applications and more coming. In your example of an actual result as is found in this body, the image segmentation [0] and recognition task is applicable.

One of the bullet points in the above is "Intra-surgery navigation." Let's say there is a known position of a tumor within an organ based on previous imaging. The human body is squishy and interior organs may shift in position or shape when a procedure is performed. Deformable image registration [1] guides the surgeon (human or robot) to the tumor location with greater precision than what was possible before.

0. https://en.wikipedia.org/wiki/Image_segmentation

1. https://paperswithcode.com/paper/voxelmorph-a-learning-frame...

[iamtheworstdev]

your dentist is probably using it to analyze xrays if you get them every couple of years.. it points out all sorts of neat stuff