An apology first, on behalf of everybody who repeated it. It became a fashionable thing to say at conferences, years ago, that we should stop training you, and the people who said it most confidently had never set foot in a reporting room. Meanwhile the lists got longer, the studies got bigger, and there were never more of you. You have been more polite about this than we deserved.
So let me concede the part your trade is tired of arguing about, and then say the thing I actually came to say. Some of the detection is good. On narrow, well-defined tasks a model finds what it was trained to find, reliably, and it doesn’t get worse at the end of a long list. Anyone who has watched a flagged bleed jump the queue and get a patient into a scanner sooner is not going to be talked out of it, and shouldn’t be.
That is a claim about detection. It is not a claim about reading, and the difference is the whole post.
Your day is protocolling, priors, the phone and the dictation. Deciding what the scan should even be, from a request that says abdominal pain, query cause, and nothing else. Pulling the old study and working out whether the thing is bigger. Being interrupted by a surgeon who wants a look now. Telling whoever is at the scanner whether to add the extra sequence. Then the meeting where the oncologists and surgeons sit down with the images and somebody has to say, out loud, what they think that is. Then a list that is never empty, sent overnight to a service in another time zone and coming back for you to live with in the morning.
A great deal of that is being automated and some of it should be. Measurement, comparison with the prior, the segmenting, the reconstruction that lets a scan be quicker and cleaner. The draft on the normal chest film. The chasing of follow-ups that everybody hates and nobody has time for. Good. None of that is the part I came to defend.
The report doesn’t go, because a report is not a description of an image. It’s a signed instruction about what happens next to a person, addressed to a particular reader who will act on it. Change the reader and the same pixels produce different sentences. Change what the referrer can do about it and they change again. That isn’t sloppiness, it’s the job, and it’s why two competent radiologists can write two different true reports on the same study.
Which brings me to the small thing you saw that had nothing to do with why the scan was ordered.
You know the one. It sits at the edge of the field, it is almost certainly nothing, and you have to decide whether to put it in writing. Mention it and you start a chain of appointments and repeat scans and a frightened person who now has a thing, and sometimes a needle goes into something that was never going to harm them. Leave it out, and if it turns out to matter, the name on the report is yours. Nothing in the image tells you which way to go. You settle it with the patient’s age, the referrer, what the follow-up will do to this particular person, and what you can defend later.
Then there’s the phrase. Clinical correlation is advised. Cannot be excluded. Everyone in medicine makes jokes about that language, and most of the jokes miss what it’s doing. Sometimes it is genuine uncertainty written down honestly. Sometimes it is a person moving a decision to somebody with information they don’t have. And sometimes, let’s be adults, it is a person covering themselves. All three are the same act, which is placing the weight of what comes next somewhere specific.
A machine will flag the incidental. It will flag it every time, because flagging costs it nothing. It has no clinic to fill, no patient who can’t sleep, no colleague who has to do the biopsy. The cost of a flag lands entirely on people who are not the flagger, and somebody in the building has to be willing to take a flag and do nothing about it, in writing, with their name on the line.
That is what a tool of this kind actually does, and it is worth saying plainly. A tool that points at things does not settle anything. It moves the question one step along, from whether there is something there to what happens now that somebody has been told there might be. The second question has never been an imaging question. It gets answered in clinics, in waiting lists, and in a conversation with a frightened person who will remember it for years.
The argument that should worry you isn’t about machines at all. If the read comes with the scan, why does the answerable person have to be a radiologist? Let the clinician who ordered it take the report and own the consequence. They know the patient, which you usually don’t. Nothing about the position requires your training, only somebody’s signature.
I think that’s the actual threat to your trade and I don’t think it’s silly. My answer is only that the position doesn’t disappear, it moves, and it lands on somebody with less imaging training and less time, who will then need somewhere to send the hard ones. Where those go is not a technical question. It’s a question about who a health system is willing to have blamed.
And I’m not going to end this kindly, because the honest reading isn’t kind. Every tool your trade has been given has raised what a day is expected to contain. The flags arrive faster than the reporting time you save, and the being-answerable is not a moment at the end of the study. It’s the posture you hold for the whole list, and it’s what makes the ordinary parts tiring. That doesn’t get lighter when the ordinary parts go. It gets more concentrated.
Pick the last incidental you decided not to chase, one of the small ones you settled without slowing down. Work out what you actually used to settle it, and how much of that came from the images. Then answer something I can’t. Who else in that building could have made the same call, and still be willing to explain it a year from now?
