How AI Is Changing Radiology
Radiology is where AI in medicine has gone furthest, and where the hype and the reality are easiest to confuse. Here is what is actually happening, from someone building in the space.
Radiology is a natural fit for AI. It is image-heavy, the images are already digital, and there is a global shortage of radiologists against a rising tide of scans. So this is where much of the real progress, and much of the overselling, has landed. Having co-invented in this area, I want to separate what AI genuinely does well from what people wish it did.
What AI does well today
The honest list is shorter and more useful than the marketing version. AI is good at triage, sorting a worklist so the studies most likely to need urgent attention rise to the top. It does not diagnose anything there; it changes the order a human looks, which in a busy department matters. It is good as a second set of eyes, flagging regions a radiologist may want to examine more closely, which can catch the thing missed at the end of a long shift. And it is good at measurement and tedium, the volumes, distances, and counts that humans find draining and machines find easy.
Notice what these have in common. They make a skilled human faster and less fatigued. They do not replace the human.
What it does not do
AI does not read a scan the way a radiologist does, with context about the patient, the history, and the stakes. It does not take responsibility for a decision. And it does not work reliably outside the conditions it was trained on, which in medicine is a serious limitation, because patients and machines and protocols all vary. The honest framing is assistance, not autonomy. Anyone selling autonomous diagnosis is selling something that does not exist yet.
The real bottleneck is not the model
The thing that surprises people is that the blocker is rarely the algorithm. It is the surrounding problems. Medical data is intensely private and cannot be pooled casually. A 3D scan is a huge volume of data to search. And when a machine influences care, you have to be able to explain afterward what happened. Solve the modeling and you have solved the easy third of the problem.
This is exactly where my own work focused: teaching a system where to look first, keeping the patient data protected, and making the whole thing auditable. There is more on the patents page.
Where it is heading
The near future of AI in radiology is not a robot radiologist. It is a quieter shift, where routine measurement and first-pass triage are increasingly handled by software, freeing radiologists to spend their judgment where it counts. The tools that win will be the ones that make experts better and can prove they are safe, private, and accountable. The flashy demos will keep coming. The ones that reach patients will be the careful ones.
FAQ
Will AI replace radiologists?
No. Today AI assists radiologists by triaging worklists, flagging regions for review, and handling measurement. It makes experts faster and less fatigued, but it does not read scans with full clinical context or take responsibility for decisions.
Part of the Patents series →
Can You Patent an AI Algorithm?
This is one of the most common questions I get now that I have a patent in AI. The short answer is that you usually cannot patent the math itself, but you can patent a specific, novel system that puts it to work. The distinction is the whole game.
AI and HIPAA: Using Patient Data Without Breaking the Rules
You cannot build useful medical AI without medical data, and medical data is some of the most regulated information there is. Here is how those two facts are reconciled, from a security perspective.
Why Medical AI Needs Security, Not Just Accuracy
Every medical AI pitch leads with accuracy. Almost none of them lead with security. That is exactly backwards if the goal is to put the thing near a real patient.