Will Artificial Intelligence Replace Radiologists?

Yes, when it replaces engineers and everybody else

Photo by National Cancer Institute on Unsplash

In my experience, most people don't understand what radiologists do. This hasn't stopped otherwise smart people from insisting that radiologists are going to soon be replaced by artificial intelligence. I disagree, and want to outline why.

In particular, I'd like to address this to medical students who are trying to decide on which specialty to enter, and young software developers and engineers working in this space.

I will mention that I do have a little first-hand experience with machine learning. I did all of the coding myself. Fortunately I had a background from my master's program in biomedical engineering. Obviously I do not hold myself out as having any true expertise in AI.

It is a puzzling fact that while there is an enormous amount of talk about this topic, and massive amounts of hype and marketing effort around it (something like 200 companies in the radiology space alone), the use of so-called AI products in actual clinical radiology practice remains quite low. This is a hint that the task is harder than proponents suspected.

There is an analogy to be made between radiology and driving: both tasks involve complex pattern recognition. We've been promised truly autonomous driving for a while but its arrival keeps being postponed. There is a general recognition now that it is much more difficult than first anticipated. It is true that there is a subset of functions that can be effectively handled by automation in both tasks, but also a residuum of mortally important ones that (so far) cannot be.  

Life or death decisions turn out to be tricky.

The whole of diagnosis in medicine requires pattern recognition. Radiology uses images so seems uniquely amenable to current AI approaches, but it is not much of a stretch to envision encoding clinical data such as physical exam findings, laboratory values and so forth onto a graphical, even image-like, display. AI could then operate on these as well. For this reason, if AI replaces radiologists, I believe replacement of internists and the like will not be far behind.

There is a widespread belief that medical images by themselves contain diagnoses. That is, that the images alone encode enough information that a sufficiently skilled and intelligent agent could extract the correct diagnosis. This is sometimes true. A radiograph of a broken femur doesn't leave much doubt. But these cases are not where the radiologist expends most of his or her effort. More common are situations where the images suggest a diagnosis, or a set of diagnoses, but the correct choice depends on incorporating more information.

This is the crucial role of the medical history. Not only does it direct the attention of the radiologist to the clinical problem at hand, it also gives a clue to the pretest probability of various possible diseases. Determining these probabilities is its own complex conceptual task. It requires looking at basic points like age and gender, but also history of other conditions, recent changes in symptoms and so forth. In my experience, the best radiologists become quite skilled at this, though time constraints impact their efforts severely. It takes time to sift through the medical record. Improved retrieval and display of information from the EMR (electronic medical record) could be very impactful for radiologists.

If a small lesion is noted in the liver for example, whether it was present previously and whether it has changed in size are important determinants of its clinical significance. There are numerous benign such processes, but a few malignant ones as well. Features on the original images are not always definitive. Further tests are sometimes needed. The correct imaging technology must be selected (ultrasound? CT? MRI? other?) along with the appropriate technique (with intravenous contrast? without? both? what kind of contrast? how fast to inject it? when to image during the injection? what coils to use? what pulse sequences to use? etc.). These are all decided by the radiologist, and the subject of decades of medical research by radiologists.

The human body has many normal anatomic variations, and many benign processes that must be recognized and discounted. This too is a big part of the radiologist's job.

Also, various pathological processes can result in the same or similar imaging appearance; this is largely because the body has only so many ways in which it can respond to insults. The lung will fill its airspaces with fluid whether the cause is a bacterial pneumonia, pulmonary edema from heart failure, or aspiration of ingested substances. The overall patterns are quite similar, but a skilled radiologist can often use subtle aspects of the pattern along with details from the clinical history to distinguish between them.

AI doesn't have an overall "concept" of human anatomy, physiology or pathology. Most current models don't pay attention to causal links, that is, arrows of causation. There is no understanding that one thing leads to another. This impedes progress, and is clearly not how humans think.

Thus the task of a radiologist is sufficiently complex that replacement by AI will take quite a long time, if it ever happens. Problem solving by radiologists is not enormously different from that by engineers. Both professions focus on practical solutions to real, complex problems. Both professions will be impacted by AI gradually, but relentlessly.

The radiology profession will change over the years, just as it has in the past. The difference between what I actually did day to day in 1982 versus what I actually did day to day 40 years later is stark. Many methods fell out of favor, and many others came into use. AI will play a major role in new approaches, and I applaud this. It's exciting. There is much progress to be made. Who do you think is going to be making that progress? Not engineers by themselves. They will still need doctors as partners, because solving the problems involve more than the images. Those doctors are going to be radiologists.

I haven't even touched upon the human component. Again contrary to many non-radiologist's conceptions, many radiologists do see and talk to patients. I regard this as the best part of my job. I would encourage the radiology profession to expand this function, and not allow the enormous pressures for so-called efficiency to destroy one of the most rewarding dimensions of our profession. But that is a topic for another post.