[jepper]

These articles are always way overhyped. A lot more development is needed before it can even be safely used as a tool alongside radiologists. The dermatology example is an excellent one. It would be great to have a simple image classifier that can help the diagnosis before the clinical appointment.

Software on its own can't even tackle a 'simple' task as reliable 6 lead EKG (electric heart film) analysis yet. A chest x-ray has a lot more variables. Plus clinical variables as patient history etc can make a big difference in diagnosis on a similar image.

[ska]

But there have been successful commercial applications of ML/AI working alongside radiologists safely as a tool for at least a couple decades.

With all the current interest in "AI" it's easy to forget that this is an old problem area, and current techniques don't fundamentally change anything. In most applications the biggest issues remain access to and quality of the data. For the right application though, you can do useful things.

[jrowley]

Additionally the anatomy between people can very greatly. It's pretty hard for computers to determine what is reasonable and unreasonable.