What Medical Students See in AI
Is artificial intelligence (AI) good, bad, or both to radiologists? As the buzz around AI has reached near-peak hype levels, many medical students are left wondering what radiology will look like as a specialty in 10 years, or if it will even exist. Will radiologists will be summarily replaced and driven to extinction by their algorithmic counterparts?
Just as there is no stable, long-term demand for the faddish Bitcoin — which has yet to establish its value for merchants worldwide — the sheer complexity of medicine and challenges to implementing AI make questions about its future hard to answer. Despite living on the fringe of radiologists’ daily duties, AI is no fad. Ultimately, AI will transform radiology in much the same way CT and MRI technology did.
Med students and AI
It’s easy to assume that medical students have the same apprehension toward new technologies that doctors do, yet in my experience that’s far from the truth. I recently had an opportunity to present on AI and radiology at the American Medical Student Association. I was surprised to see a younger generation ready to embrace the technology. When I asked if they would jump into a driverless car today, student hands shot up throughout the room signaling an overwhelming “yes.”
I am a little older and more skeptical, but maybe the younger generation sees AI more as an opportunity rather than a threat. As I told the students in attendance, for better or worse, AI will be shaped by them — with the goal of transforming any radiologist into a super-radiologist.
So what questions did the medical students have?
Do I need to learn to program to be a radiologist?
Understanding how AI works is important, and particularly how training data shapes an AI algorithm. Coding is a great skill, but radiologists won’t need to learn to code, just as you don’t know how to assemble a car transmission in order to drive. In fact, in many ways, AI will make data analysis easier than ever.
What will happen to physician compensation in the era of AI?
Clearly, AI will cost money, and the question quickly becomes: Who is going to pay for it? If we look at other industries — airlines and banking, for example — automation hasn’t been a primary driver of change to compensation. It is likely to be the same in radiology.
Insurance programs will likely continue to reimburse a portion of costs with specific procedural codes — and radiologists, in combination with patients, will pick up the remainder. As AI boosts radiological productivity, new opportunities will open for radiologists seeking more direct patient consultations. There may also be cost savings from lower malpractice premiums that reflect more accurate scans and patient outcomes. When CT technology became commonplace and simplified interpretation, pundits similarly predicted that radiologists would be superfluous. That didn’t happen.
Who will be responsible for AI interpretations?
The radiologist will be responsible for AI outputs, much like a pilot is responsible for the airplane, or an attending is responsible for a resident. As we have seen with computer-aided diagnosis, radiologists still need to be the penultimate decision makers. AI can and will be wrong on many occasions and for multiple reasons. It will be up to physicians to catch these mistakes. No one is likely to thank us for catching mistakes, but as physicians, it is already included in our job description.
Most radiologists strive to be highly sensitive and specific in diagnosing patients. We want to be in the top left corner of that receiver operating characteristic curve for performance — to be super-radiologists in our own way. I have yet to meet a fellow diagnostician who wants to over- or under-call a diagnosis, saying, "I really want to miss cancer!"
As many of us know, practicing radiology has a way of breaking down hubris: The "this has to be cancer!" or the "looks ok to me" moments can become famous last words. For the next generation of radiologists, AI will offer a helping hand in detecting subtle findings or for classifying not-so-subtle findings. AI will have a role in more quickly concluding, “maybe that isn't cancer after all."
AI career training
The future generation will need to understand how we build AI and how to make AI work in their institutions, but not necessarily how to program it. Knowing and understanding the Python programming language may certainly be useful, but just as radiologists don’t need the know the programming code to adjust MRI pulse sequences, neither we will need to know programming code for AI to tune AI models for our facilities.
Still, many programmers see AI as more data curation and collection than traditional programming, and that’s where partnerships between data scientists and radiologists will become invaluable. AI is only as good as the data it is fed. As the saying goes, garbage in garbage out (GIGO). Poorly tagged or labeled data creates worthless AI, and radiologists will be required to correctly annotate datasets.
Four steps to embracing AI
How can medical students be great consumers of AI?
1. Learn the basics of data science. AI requires some understanding of data science, from which much evidence-based medicine is derived. It is important to understand basic statistics, as these frame the discussion of how well AI works.
2. Know the applications in which AI works best. AI works well with very specific tasks (such as a classifier), but it isn’t able to classify a disease it hasn't been trained to detect. AI is no panacea, and it won't work like that for a long time. Narrow AI, for specific tasks such as classifying lung nodules, will have real benefits in the short term. Broad AI, intended to classify all forms of lung disease, is likely to provide poor results for now and the foreseeable future. AI will still be beneficial, helping unload some of the tedious work, such as finding micronodules. This allows the radiologist to consider the bigger picture — “Is this perhaps a strange fungal infection?”
3. Understand the role of data in developing and training AI. Those using AI need to understand how the dataset impacts the algorithm and how augmenting the dataset will change the results. For example, if a database has only male patients, we can expect to have some serious problems applying the algorithm to a female population. As a result, fair questions to ask a vendor may include how the algorithm was trained, where it was developed, what studies support it, and what equipment was used for image acquisition.
4. Consider how AI will be incorporated into a workflow. If AI adds the burden of more clicking and time, no one will want to use it. AI needs to make radiologists more efficient. It is important to consider how AI will be integrated into the workflow and how it will perform seamlessly in the background to support radiologists.
Future generations of radiologists are sure to advance and shape the field tremendously — creating new opportunities to improve patient care. The medical students I meet seem rightly excited about this new era.
AI will offer us modern wisdom, and convert radiologists into super-radiologists by vastly expanding the data we use to make decisions. In effect, AI is a means of compressing data and experience. An AI algorithm can be trained on millions of examples in a matter of days. It can process far more images than a human reviewer ever could, which absolves us of our natural human deficits and super-sizes our abilities. Today’s younger generation of medical students hopes to achieve better interpretive results and provide more connective patient care by embracing this technology.
Ayis Pyrros, MD, Data Science Evangelist, undergraduate medical education community | Neuroradiologist, DuPage Medical Group, Hinsdale, IL
Interested in setting up an AI presentation before an undergraduate medical student group? Please contact DSI.