100% agree.
Forget radiology and medical specialties for a second. Look at general AI progress and the pattern is pretty obvious:
-Math, coding, and anything language-related (text-based): Supra-human and broadly general.
-General image recognition tasks (image-based): below humans on benchmarks such as ZeroBench, BabyVision. Not generalized.
-Self-driving (image-based/perception): suprahuman in limited common settings, but fails at edge cases.
-General everyday tasks such as folding laundry, cleaning your room, taking out the trash, etc (image-based/perception): well below humans and not generalized.
No doubt computer vision will get better, and those at frontier labs may devise clever tools to achieve human-level general vision.
But on a relative scale, right now AI is clearly better at text/language than it is at vision.
This has big implications for medical specialties, and well as white/blue collar work.
In this sense, radiology is actually more of a "blue-collar" type of work.
Maybe you’ve missed the last four years: frontier AI has progressed much faster in language than in vision, particularly in complex visual reasoning.
If that trend continues, full automation of radiology could come well after the automation of most language-based cognitive work.