The problem is that unlike a blood test, neural networks are not transparent at all. So when an imaging test comes back with "possible malignancy" with red circle around the highlighted area, a GP doesn't have the experience to agree or disagree with the result. So every time that happens the next step will be to send the image to a specialist.

And in the case where the machine learning algorithms don't find anything suspicious, the GP again won't have the training or experience to confirm those results. Now if the person was otherwise healthy and this was just a screening that might be enough, but if the GP was suspicious enough to order the test in the first place, it won't be.

What will probably happens is that this kind of technology increases productivity for radiologists, and maybe increases the number of screenings done on healthy people. But it's not going to reduce the demand for radiologists.

Basically the problem is that to be able to interpret the output of a neural network you need to be an expert. What we need is AI that can present a fully formed argument that is easy for a non expert to follow and validate, but we are nowhere near that in most cases.