Artificial intelligence will reduce the need for clinical medical physicists

In 2011, IBM’s supercomputer Watson defeated the former human winners and won the first prize on Jeopardy! game. It has created an overly publicized attention on machine learning and Artificial Intelligence (AI). Early this year, Google AlphaGo has marked a major breakthrough in AI by winning the first game against the world’s best champion human player in the world’s most complex game, the ancient Chinese Go game. With no doubt, the interests in AI and its related products had reached a global frenzy. As scientists advance in technology, a concern of job security has risen up: will robots take our jobs? IBM Watson has evolved from a “question answering machine” to a highly intelligent “cognitive diagnostic engine” or a “decision support system” over the past 6 yr. Based on Carl Frey and his collaborators, future family health centers may transition to a team of nurse practitioners with the support of Watson Health and overseen by one single doctor. Will AI technology also marginalize medical physicists in the near future? In this series, we have Dr. Xiaoli Tang arguing for the proposition that “AI will reduce the need for clinical medical physicists” and Dr. Brian Wang arguing against it. Dr. Xiaoli Tang received a Ph.D in Electrical Engineering from the Rensselaer Polytechnic Institute. She then did her postdoctoral training in Medical Physics at the Massachusetts General Hospital and the University of California at San Diego. She previously worked at the University of North Carolina and now is working as an Assistant Attending and chief physicist at the Memorial Sloan Kettering Cancer Center Westchester regional site. She is an expert in motion management, Deep Inspiration Breath Hold (DIBH) for left-sided breast cancer, and machine learning algorithms on medical physic applications. She is interested in developing related clinical trials, and bringing new technology to the clinic. She is a member of the American Association of Physicists in Medicine (AAPM), and the American Society for Radiation Oncology. Dr. Brian Wang received his PhD in nuclear engineering from Rensselaer Polytechnic Institute in Troy, NY in 2005. He currently works at University of Louisville as the chief of physics and medical physics residency director. Dr. Wang is an associate editor for the JACMP. His research interests include motion management, image guidance, and SRS/SBRT. Dr. Wang has been involved with the AAPM Spring Clinical Meeting and its predecessor ACMP annual meeting as a program director or the subcommittee chair for 8 yr. Dr. Wang serves on several committees at ASTRO, RSS, and ABR.


| INTRODUCTION
In 2011, IBM's supercomputer Watson defeated the former human winners and won the first prize on Jeopardy! game. It has created an overly publicized attention on machine learning and Artificial Intelligence (AI). Early this year, Google AlphaGo has marked a major breakthrough in AI by winning the first game against the world's best champion human player in the world's most complex game, the ancient Chinese Go game. With no doubt, the interests in AI and its related products had reached a global frenzy. As scientists advance in technology, a concern of job security has risen up: will robots take our jobs? IBM Watson has evolved from a "question answering machine" to a highly intelligent "cognitive diagnostic engine" or a "decision support system" over the past 6 yr. Based on Carl Frey and his collaborators, future family health centers may transition to a team of nurse practitioners with the support of Watson Health and overseen by one single doctor. 1 Will AI technology also marginalize medical physicists in the near future? In this series, we have Dr. Xiaoli Tang arguing for the proposition that "AI will reduce the need for clinical medical physicists" and Dr. Brian Wang arguing against it.  Physics as well. Out of many duties that medical physicists have taken upon, the clinical aspect mainly includes treatment planning, chart checking, and machine quality assurance (QA). The need of clinical physicists in these areas has already been slowly reduced over the past several years, and this speed is going to be increased as more AI technologies are implemented in the clinic. Research has shown that on average, these automatic planning systems can achieve planning target volume (PTV) coverage that is highly comparable with the original plan. The normal tissue sparing is also within acceptable range. 7,8 The performance of these commercial systems is expected to further improve in the following years, and these systems will gradually replace the manual planning process, at least for those standard plans. It seems inevitably that the need of routine planning by physicists or dosimetrists is diminishing by then. Some clinics have already started planning on letting dosimetrists do the initial chart checksum the automatically generated plans, where the initial chart checks are now routinely performed by physicists.
Initial chart check is a thorough checking process after the plan is approved and finalized. This is the second area that the need of physicists will be reduced. One source of planning problems is contour discrepancy and/or interuser variations. Machine learning is an application of AI. Currently, the machine learning-based autosegmentation systems can reliably contour structures with standard shapes or can be distinguished out from surroundings, such as bladder, rectum, heart, lungs, etc. With the improvement of AI, segmentation applications might be able to contour more challenging structures, such as prostate, spinal canal, etc. There have been many studies on the diagnosis area on automatic tumor segmentation. 9,10 In the therapeutic area, AI-based systems might later be able to contour PTVs as well. These new technologies will reduce potential variations or inaccuracy of contours and lead to less chart checking problems.
Consequently, the need for physicists will be reduced as well.
Machine QA is the third area that will need less physicist efforts.
There has been research on predicting machine output trends using AI-based algorithms. 11 The proposed data visualization can predict the Linac performance over time and prompt physicists to perform output calibration before the output is drifted away from the tolerance. The routine Linac output checks (daily, monthly or annual) are currently recommended as a standard. Yet, based on our experience, modern linacs are getting so stable that we might not need to routinely perform monthly output calibration. In our clinic, output calibration happens on average once per 6 months. This suggests that , with AI-based predictions, we perhaps no longer need to check machine output on a monthly basis. Similarly, data visualization is also an effective tool to perform data comparisons, alert failures, and potential identify causalities. All these may lead to the reduced frequency of medical physicist interaction.
For diagnostic physicists, the need for medical physicists may also be reducing. One of the major responsibilities is the optimization of the clinical imaging procedures. 12 Some vendors have already automated this process through implementing AI-based solutions.
For instance, IBM Watson is able to review a digital chest x ray and suggest that the patient may have small-cell lung cancer and heart surgery. Watson can then go ahead and search PACS, EMR, and departmental reporting system to bring in related files without any physics interaction. 13 Similar to therapeutic physic situation, the need for physicists' machine QAs might be reduced due to the AIbased predictions of the machine performance. As technologies and treatment options get more complex in healthcare, clinical medical physicists will get involved in more nonroutine activities, which cannot be reduced by AI technologies.
A recent publication advocates for more direct patient interactions by clinical physicists in the future and I full-heartedly agree with the authors for their following remarks. 14   that this might due to the increase in complex technologies, that is, multileaf collimator (MLC) and intensity-modulated radiotherapy (IMRT). However, AI is still at its infancy stage. We know that usually it requires more efforts when we adopt new technologies. What clinical medical physicists would do is more QA at the beginning. It is very likely that we see an initial increase in the need of physicists for QA and clinical implementation. However, eventually, as much as we do not want to admit, the need of clinical medical physicists in cancer care will be gradually reducing as the AI technology matures. For example, it is not a trivial job to develop a knowledge-based planning model that is customized to someone's own institution.

3.B |
Second, as also noted by Dr. Tang, the treatment plans generated by AI have to be extensively checked by a human, either a dosimetrist or a physicist. Last but not the least, it creates a new challenge and task to implement routine QA procedures for these AI technologies.
Currently, most of the research and development effort for AI technologies stems from industries and academic institutions. As AI starts migrating into clinical practice, the associated supporting resource will shift to the hands of clinical physicists.
As for diagnostic imaging, the current clinical physicists are already facing a specialty identity issue. While medical physicists continue making important contributions to the field, their clinical roles have not always been viewed as critical. 15 With AI being utilized more in the field, additional efforts are needed by clinical medical physicists to understand an imaging artifact, to optimize a scan protocol or to monitor an equipment performance. Most of these tasks cannot be generalized and have to be customized to each individual institution and patient group.
For both diagnostic imaging and radiation therapy, the values of clinical physicists are demonstrated not only by routine QA activities, but more importantly by nonroutine activities. Commissioning a new technology for radiation therapy is not as simple as measuring LINAC dosimetric performance anymore. It involves clinical physicists to communicate with all team members within the department. Similarly, setting up a new imaging protocol is not just creating a set of parameters; instead, it requires clinical physicists to discuss the process with technologists. Quite frequently, clinical medical physicists need to provide in-service education to other team members annually or before the initial implementation. A couple decades ago, clinical physicists may be seen as the "phantom" of the department: they do all their work such as patient and machine QA at nights and on the weekends. Now, clinical physicists have already become an indispensable team for patient care, especially in radiation therapy. In the future, clinical physicists will get involved in more patient-specific consultations, which is very difficult to be replaced by an AI machine. Now at the organizational level, AAPM has initiated the Medical Physics 3.0 project to redefine and reinvigorate the role of physics in modern medicine. With the growing public and collegial awareness of our role, the demand for clinical medical physicists can only increase as technologies such as AI are becoming more complex in healthcare.

CONFLI CT OF INTEREST
The authors declare no conflicts of interest.