Mentorship in Radiation Oncology
ACR helps empower radiation oncology trainees through mentorship, leadership and advocacy, showing residents can help shape the specialty’s future.
Read more
Main Content
Can and Cannot of AI
Having initially contemplated a career in computer science before pursuing medicine I have always had a fond view of technological advancement and the capabilities for good that it could bring. While I had heard about artificial intelligence for many years, it had never been at the forefront of society until the advent of ChatGPT. With AI being able to create videos, explain complex topics, and draft documents, it appears that future professions could become disrupted by the powerful abilities of AI. Because AI can process information at magnitudes that a human mind could not, professions at risk are those of a cerebral nature that rely on cognitive ability rather than manual labor. Radiology seems especially vulnerable to change and thus, I sought to better understand the impact that AI could have on it. Throughout my conversations with both physicians and non-medical professionals, a dichotomous view of the future of radiology started to emerge.
Some view AI as a productivity booster for radiology that could bring about enhancements in workflows without reducing the need for radiologists. Those that focus on the benefits of AI cite consistent volume growth of imaging, novel modalities, the overall positive job market, and past technology shifts such as the move from film to PACS, which improved efficiency without harming job prospects. Other viewpoints noted that the liability burden of medicine as a field would preclude AI from taking over a physician’s job. Would a software company really want to be liable for a person’s health outcomes? Finally, many highlighted the radiologist’s roles in procedures and clinical consultations—areas AI cannot replace.
However, some respondents expressed a more pessimistic view, arguing that radiology may be uniquely vulnerable to disruptive effects from AI because much of its work involves interpreting a finite dataset with objective key findings—tasks well suited to machine learning. Although current image-analysis tools remain limited (as evidenced by ChatGPT’s inability to reliably interpret a chest X-ray), the rapid pace of AI advancement and targeted algorithm training could overcome these limitations sooner than anticipated. After all, multiple studies have shown AI’s utility in breast and chest imaging especially for subtle nodules and calcifications1,2. Therefore, it is just a matter of time before AI will become better at interpreting images and once that moment has come, the number of radiologists needed will drop precipitously as direct replacement of human radiologists with AI algorithms will commence. AI liability need not be borne by software companies as one radiologist would be able to oversee several AI algorithms and be the one responsible for their accuracy. This is the view espoused by those who are negative on the outlook of radiology as a field. They believe that the total number of radiologists needed will decrease and each individual radiologist will be able to supervise multiple AI algorithms displacing the job of several current radiologists.
These opposing views raise a central question: Will AI displace the radiologist, or will it simply augment their clinical repertoire? There is merit to both viewpoints, and my goal has been to investigate the current state of the AI debate in radiology and evaluate AI’s strengths and limitations.
Throughout my research, I have concluded that there are two main domains that AI can impact: image analysis and workflow optimization. It is already impacting many radiological disciplines including breast imaging, neurological imaging, and lung nodule screening3,4. Some studies have shown superior performance of AI algorithms in lung nodule detection utilizing LDCT compared to six trained radiologists3. Screening exams like low-dose CT and mammograms suit AI well, given the vast datasets used to train algorithms. However, AI struggles with unfamiliar data and rare pathology giving human radiologists a key area where they can provide their judgment. Even if AI algorithms take a foothold in screening modalities, it is unclear if they can displace radiologists who currently occupy that same role because AI outputs must be “checked,” often taking as much time as reading the study independently. Furthermore, truly autonomous AI would require new legislation. Are we, as a society, ready for autonomous algorithms to play a significant role in something so vital as a person’s health? That is an open question, but as of right now, there are no autonomously functioning AI algorithms within radiology.
In addition, AI may have a promising role triaging time-sensitive studies, such as brain bleeds or acute stroke rule outs. These advances seem most promising as the radiologist can triage studies more efficiently with AI’s help, improving outcomes for both patients, hospital workflows, and emergency personnel. However, over-reliance on AI triage can result in a delay of care for pathology which goes falsely undetected. Finally, AI can play a vital role in non-interpretative tasks such as report layout, grammar checking, and helping dictation errors. This seems like another promising avenue for AI in radiology as it can reduce misunderstandings and can enhance the quality of radiology reports.
Notwithstanding these advancements, AI’s lack ability in several key areas. Most importantly, AI is only as strong as its training dataset and fragmented health systems and HIPAA restrictions limit dataset quality. Rare or unconventional pathology may go undetected or may be completely misunderstood by AI, providing a key area for human radiologist input. While human radiologists may miss occult fractures or small nodules, AI may commit nonsensical errors such as calling a bone an artery—a mistake a human would most likely not make. Such mistakes would limit trust and reduce AI’s role in independent decision-making. Because AI struggles to explain the why or how behind its reasoning, its ability to operate autonomously without human radiologist input will be hampered.
Additionally, AI cannot train future radiologists or advance the field through research. Training resident physicians is an inherently human domain requiring close mentorship to help young radiologists understand how to interpret and dictate studies with nuance. Furthermore, clinical research demands creative, human-led inquiry—something AI cannot replicate. Finally, having a human radiologist as a colleague to consult with and as an interventionalist for procedures are both areas that AI cannot impact.
Radiology stands at a crossroads. Disruption by AI is a certainty in the future. However, the extent of this disruption is still unclear. I have sought to better understand the AI debate as it currently stands and evaluate the arguments in favor and against AI. I have investigated the key areas of promise for AI and some areas that AI is not as capable in. While I believe that AI’s maximal impact will be as a resource for future radiologists to enhance accuracy and improve study reporting, I acknowledge that AI’s ability to become proficient in the large volume of screening studies may have an impact on job prospects in the future. Much remains uncertain, and AI’s true impact remains an open question that will depend on technological progress, legislation, and society’s own comfort with algorithms playing a part in health care.
Mentorship in Radiation Oncology
ACR helps empower radiation oncology trainees through mentorship, leadership and advocacy, showing residents can help shape the specialty’s future.
Read moreBalancing Risk, Benefit and Trust in Radiology
Radiologist-led radiation optimization and clear communication about imaging risks and benefits matter more than ever.
Read more
Thriving in Radiology Residency
Match Day marks the start of radiology residency: focus on preparation, safety, consultation skills and involvement — no one expects you to know everything.
Read more
