November 28, 2022

Machine Learning in Radiology: Is AI Coming for My (Future) Job?

Christian Henriksen, BS, MS2, Creighton University School of Medicine

Christian Henriksen, BS, MS2Deciding on a specialty forces each medical student to grapple with a question that has been put on hold for a long time: what am I going to do with the rest of my life? In typical medical student fashion, we try to gather as much information as possible before deciding. What is the work-life balance? The bread-and-butter pathologies? The usual patient interaction? And of course, what does the future of the field look like? In the case of radiology, there is an elephant in the room that underscores that last question. Every aspiring radiologist wonders: is “artificial intelligence” (AI) just a flash in the pan, or is it time that we see the writing on the wall?

Recent innovations in the field of automated diagnostic tools have been made possible by advancements in a specific type of AI called “machine learning.” This technique involves sending input data through a neural network — essentially a very complex mathematical function — that “learns” to choose the correct outputs by iteratively modifying itself. Machine learning has shown undeniable potential through detecting fractures and pneumothoraces, classifying thyroid nodules, and diagnosing lung cancer more accurately than a team of radiologists (under certain conditions).1,2 While these are all notable feats, concerns over the limitations of these tools have been raised by the American College of Radiology® and the Radiological Society of North America.3 A neural network is only as good as the data it was trained on, which calls into question the generalizability of these models. Similarly, the way these models are trained means that it can never be said for certain that a neural network is correctly identifying the important features of inputs or merely relying on unseen confounding variables. However, it is the prevailing opinion of experts in the field that it is a matter of when, not if, AI will overcome these pitfalls.2

The development of AI is accelerating at a rapid pace — but how will that impact the clinical practice of radiology? First, it bears mentioning that the most feared outcome — the complete replacement of radiologists by AI — will probably not manifest within our lifetime. The first inklings of AI-assisted diagnostic tools entered medical literature with the term “computer-assisted diagnosis” in 1966. However, it was not until 2018 that the U.S. Food & Drug Administration approved the first fully autonomous AI-based tool, a program for the diagnosis of diabetic retinopathy.4 Yet, despite impressive diagnostic accuracy, the field of ophthalmology hasn’t even blinked: adoption of this technology remains niche due to concerns of exorbitant cost and limited scope.5

With total replacement being unlikely, a more realistic role for AI in the future of radiology is as a tool to complement the radiologist’s workflow. A program recently approved in the European Union for the interpretation of chest X-rays boasts a sensitivity of 99+% and “zero clinically relevant errors” in identifying normal scans, cutting down on radiologist workload.6 In the United Kingdom, a similar tool was approved on the condition that all results be reviewed by a radiologist.7

It is in these recent developments that we can see a more probable, and less alarmist, role for AI in radiology. Curtis Langlotz, MD, PhD, of Stanford likens it to the role of a plane’s autopilot. As radiologists, he says, “We’re going to get better at getting our ‘autopilot’ to augment what we do and to make our lives better and easier.”8 So the next time you see a headline proclaiming a new AI tool that “reads scans better than radiologists do,” just know that it will be a long time before anyone other than a radiologist is doing the final sign-off on those read reports.

To learn more about AI in radiology, explore on-demand webinars from the RFS AI Journal Club.


  1. Choy, G., Khalilzadeh, O., Michalski, M., Do, S., Samir, A. E., Pianykh, O. S., Geis, J. R., Pandharipande, P. V., Brink, J. A., & Dreyer, K. J. (2018). “Current Applications and Future Impact of Machine Learning in Radiology.” Radiology, 288(2), 318–328.
  2. Ardila, D., Kiraly, A. P., Bharadwaj, S., Choi, B., Reicher, J. J., Peng, L., Tse, D., Etemadi, M., Ye, W., Corrado, G., Naidich, D. P., & Shetty, S. (2019). “End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography.” Nature Medicine, 25(6), 954–961.
  3. “ACR and RSNA Comments—Autonomous Imaging AI.” American College of Radiology, 30 June 2020, Press Release.
  4. “FDA permits marketing of artificial intelligence-based device to detect certain diabetes-related eye problems.” U.S. Food & Drug Administration, 11 Apr. 2018, Press Release.
  5. Chen, E. M., Chen, D., Chilakamarri, P., Lopez, R., & Parikh, R. (2021). “Economic Challenges of Artificial Intelligence Adoption for Diabetic Retinopathy.” Ophthalmology, 128(3), 475–477.
  6. “Oxipit Awarded CE Mark for the First Autonomous AI Medical Imaging Application.” Oxipit, 29 Mar. 2022, Press Release.
  7. “ awarded CE Mark approval for its AI-based chest X-ray diagnosis technology.”, 15 Apr. 2020, Press Release.
  8. “RSNA 2017: Rads who use AI will replace rads who don’t.” Stanford Center for Artificial Intelligence in Medicine & Imaging, 21 Dec. 2017,