The role of the radiologist isn't just mapping image[range_x, range_y, range_z] to disease -- it's also including a vast amount of Bayesian priors on the rest of the patient's notes, and their referring colleagues' hints of an indication.

For example, often the question isn't just "does this person have mitral valve regurgitation yes/no", it's more along the lines of "is there evidence from this cardiac MRI scan that their mitral valve regurgitation is significant and able to explain the symptoms that we have -- and if so, is it amenable to treatment". That's a totally different question -- are the symptoms beyond what would be expected for the patient; and is there a plausible mechanism are all second-level radiological questions, well beyond the level of "please classify this stack of images into either healthy, cyst, or tumour". Another random example would be the little old lady who comes in with breathlessness at rest: she may well have a large, slow-growing lung cancer (that any AI algorithm would easily diagnose) that she may well die with or of, but the acute dysponea could be down to an opportunistic LRTI that remains treatable with a course of antibiotics (and visible on a plain film chest x-ray). Capturing that sort of information is a lot, lot harder.

You're also forgetting that the cost of an expensive imaging modality like MRI or CT is amortised over 10 years, and that -- by far -- the biggest cost of running the service is the staff. The doctors do more than push buttons. In many services, actually, they don't acquire the scans or interact with patients at all.