It’s definitely an imperfect signal, but we are capable of making decisions under conditions of noise and uncertainty. In other fields we would quantify the uncertainty, the consequences of making the error in either direction, and then act accordingly.

As a thought experiment, if MRI was as cheap and fast as testing blood pressure, do you think they’d still be given as rarely?

I think if we had a perfect tool where you can get MRI (and CT) data in the time it takes to get a blood pressure reading, it would only be used to detect slam-dunk diagnoses for further screening which is how we use blood pressure.

You’re assuming that we would skip this step, implicitly.

Now when would we do further screening? In a Bayesian framework, when we have a certain pre-test probability that we think we can improve to our desired level of post-test probability in order to take an action.

The fact that this is all automated doesn’t change our uncertainty about whether borderline calls should be targeted. Indeed, we’ll likely have more stringent criteria about things to ignore than are used today, and that may be worse for the small number of times we currently catch an actionable incidental finding.

And we’re not even beginning to consider cost, resource misuse, and the risks we’re incurring (an MRI machine isn’t safe to be around and costs a lot to maintain)

Edit: the things that kill people are smoking, heart disease, lack of exercise and poor diet. Targeting these is so much more useful than detecting a larger number of incidental findings