The Last Read

It is 03:14 in a reading room that no longer needs the lights on. A chest CT lands in the worklist. Before the on-call registrar has finished their coffee, the study has been segmented, every nodule measured, prior comparisons pulled, a structured draft written, and the three findings most likely to change management pushed to the top. The machine has, in effect, already read the scan. The question that defines the next decade of medicine is not whether it can. It is what the human read is for once it has.

I have spent my career in and around radiology, through the arrival of PACS, of computer aided detection, of teleradiology and the great unbundling of the night shift. Each wave was supposed to change everything. Each one changed something, and the profession absorbed it and moved on. I am writing this from inside that profession, not above it, because I think this wave is genuinely different, and I think we are mostly having the wrong argument about it.

The argument we keep having

The argument we keep having is about replacement. Will AI replace the radiologist? It is the wrong question, asked with the wrong verb. AI is not a smaller, cheaper radiologist. It is a new kind of instrument, and like every instrument before it, it does not remove the clinician so much as move them, up the stack, toward the parts of the work that are irreducibly human.

Computer aided detection promised to find the cancers we missed and mostly gave us more boxes to dismiss. What is arriving now is categorically different in four ways, and it helps to name them, because they are often blurred into one word.

Four kinds of intelligence

The first is perception: detection, segmentation and quantification. Models that see the nodule, measure it, and track it across years of priors more consistently than any tired human at 3am. This is the layer that already works, quietly, in production.

The second is language: the move from images to reports. Generative models that turn a set of findings into a clean, structured, contextually aware draft, and ambient systems that let a clinician simply speak and have the note assemble itself. This is where most of the time actually goes, and most of the burnout lives.

The third is reasoning: synthesis across modalities. Not just the scan, but the labs, the history, the genomics, the prior episodes, reasoned over together to suggest what the images alone cannot. This is the layer that turns a finding into a differential, and a differential into a recommendation.

The fourth is agency: systems that do not wait to be asked. They triage the worklist, flag the stroke before anyone has opened it, chase the missing prior, and prepare the tumour board. The radiologist stops being a queue of one and becomes the conductor of many.

The read is no longer the act of finding. It is the act of deciding.

What only medicine can still do

Here is the part the replacement argument misses. When the machine can find everything, finding stops being the scarce skill. Judgement becomes the scarce skill: knowing which of the forty true findings actually matters for this patient, on this day, given everything else that is true about them. Knowing when the confident model is confidently wrong. Knowing how to tell a frightened person what the scan means without flattening them into a probability.

Medicine has always been more than pattern recognition. It is pattern recognition in service of a human being, under uncertainty, with consequences. AI makes the pattern recognition abundant. It makes the human service, the judgement and the care, more valuable, not less. The radiologist of 2035 will read fewer images and decide more cases. That is not a demotion. It is the job finally rising to the level it always aspired to.

The futures this opens

Look further out and the shape of it gets exciting. Imaging stops being episodic and becomes continuous: a baseline that quietly updates, so disease is caught as a change in a trajectory rather than a shadow on a single film. Screening inverts from reactive to predictive, modelling who will need the scan before the symptom. The patient gains a kind of longitudinal digital twin, assembled from every modality, against which any new study is read in context.

Documentation dissolves into the background, captured by voice as a natural part of the consultation rather than a tax paid after it. Inference moves to where the patient is, running at the edge in the clinic and the scanner, so the intelligence reaches the bedside without the wait. And population health becomes tractable, because the same models that read one scan can, carefully, learn across millions.

The condition for all of it

There is one condition on every future in that list, and it is the reason I work where I do. None of it is acceptable if it requires the most sensitive data on earth to leave the patient, the hospital, or the nation it belongs to. An AI future for medicine that is brilliant and not sovereign is not a future any health system should accept. The intelligence has to come to the data, run inside the boundary, governed by the right law, and provable to the right regulator. Anything less trades the patient’s trust for the model’s convenience.

That is the quiet thesis underneath all the noise about AI in healthcare. The breakthroughs will not be limited by the models. They will be limited by whether we can run them somewhere worthy of what they are reading.

An AI future for medicine that is brilliant and not sovereign is not a future worth having.

The last read

So, the last read. Not the last time a radiologist reads a scan, but the last time the read means only finding. After it, the read becomes something larger: a human judgement, made over a machine’s perception, about a single person’s life. The machine will see everything. Deciding what it means, and what to do, and how to say it, will remain the most human thing in medicine.

We are building the ground that future runs on: sovereign, intelligent, and answerable to the people it serves. The scans will keep arriving at 03:14. The lights can stay off. The judgement, and the care, will still be ours.