The perceptions of medical physicists towards relevance and impact of artificial intelligence

Artificial intelligence (AI) is an innovative tool with the potential to impact medical physicists’ clinical practices, research, and the profession. The relevance of AI and its impact on the clinical practice and routine of professionals in medical physics were evaluated by medical physicists and researchers in this field. An online survey questionnaire was designed for distribution to professionals and students in medical physics around the world. In addition to demographics questions, we surveyed opinions on the role of AI in medical physicists’ practices, the possibility of AI threatening/disrupting the medical physicists’ practices and career, the need for medical physicists to acquire knowledge on AI, and the need for teaching AI in postgraduate medical physics programmes. The level of knowledge of medical physicists on AI was also consulted. A total of 1019 respondents from 94 countries participated. More than 85% of the respondents agreed that AI would play an essential role in medical physicists’ practices. AI should be taught in the postgraduate medical physics programmes, and that more applications such as quality control (QC), treatment planning would be performed by AI. Half of the respondents thought AI would not threaten/disrupt the medical physicists’ practices. AI knowledge was mainly acquired through self-taught and work-related activities. Nonetheless, many (40%) reported that they have no skill in AI. The general perception of medical physicists was that AI is here to stay, influencing our practices. Medical physicists should be prepared with education and training for this new reality.


Introduction
Over the last decade, artificial intelligence (AI) research in healthcare and medicine had started to show promising results and practical applications, from facial recognition to fully automatic detection and even finding new biomarkers. AI has been recognised as both a productive and disruptive force in healthcare [1]. In particular, radiology, radiotherapy and pathology are the three medical specialities that saw the more prominent AI role [2][3][4][5]. Some clinicians may also 1 3 view AI as a threat to the future of their medical practices [6]. This results in some ambivalence in the attitude towards the acceptance of AI.
The involvement of medical physicists with AI has the potential to extend the expertise of medical physicists and boost patient care [7]. For example, Cui et al. have recommended that medical physicists should be leaders in implementing AI for routine clinical practices [8]. Despite that, there were currently few medical physicists involved directly in AI projects [9].
The number of research and possibilities of applications AI in medical physics is increasing rapidly. The European Federation of Organisations For Medical Physics (EFOMP) white paper proposed that training and education programmes are needed to prepare the medical physicist for this revolutionary transformation in healthcare and medicine [10]. Based on this white paper, the curriculum has been discussed, aiming its expansion to include AI for the education and training of European medical physicists, based on their needs [11,12].
The practices and daily routines of medical physicists may also be affected by AI [2]. It has recently been suggested that big data and deep learning would profoundly change the profession, in general and future research [10]. To prepare for this future, it may be advisable to incorporate AI into the medical physics graduate programme curriculum to better adapt the professionals to a possible new reality where medical physicists collaborate with computer scientists to manage AI applications [13]. This proposition, however, still divides opinions.
Several studies have focused on healthcare professionals' perception and attitude towards AI [14][15][16][17][18][19][20]. In medical physics, the only study on the perceptions, practices, and education needs about AI was Diaz et al.'s Europe-based study of medical physics experts and their needs [9]. Diaz et al. reported on the results of a web-based survey applied to a total of 219 respondents from 31 countries (93%, n = 203 of European countries and 7% of non-European countries) [9]. Therefore, the current study aims to determine the awareness of medical physicists regarding AI and their perception of the relevance and impact of AI in the practices of medical physics, including opinions of medical physicists experts, researchers, trainees, and postgraduate students in medical physics around the world.

Method
An online survey questionnaire (see Appendix) was designed and distributed to professionals, trainees and students in the medical physics field using Google Form (Google LLC, Mountain View, CA, USA) and Smart Survey (SmartSurvey Ltd, Tewkesbury, Gloucestershire, UK), specifically to respondents in China. There were 13 multiple-choice questions (MCQ) about the participants' perception of AI, and six questions about their profile. The MCQs were designed using a 5-point Likert scale with the descriptor of agreement. Participants were also free to provide general comments. After a pilot test, the questionnaire was launched publicly and disseminated via regional and international professional organisations such as the AAPM, ACPSEM, AFOMP, ALFIM, EFOMP, FAMPO, IOMP, IPEM, MEFOMP, and SEAFOMP. The survey was held from February until May 2020, reaching 1019 global respondents. This research was approved by the University of Malaya Research Ethics Committee (UM.TNC 2/UMREC).
The survey included questions about the role of AI in medical physicists' practices and the increasing potential for AI applications in areas such as in quality control (QC) and treatment planning. The participants' opinions were also asked concerning the possibility of AI threatening/disrupting the medical physicists' practices and career, the need for medical physicists to acquire at least some basic knowledge of AI and the need to teach AI in postgraduate medical physics programmes.
Some questions about the participants' knowledge of AI (the origin of that knowledge, basic understanding, working knowledge, relevant skills, proficiency, limitations) were also included. The survey was anonymous, although questions covering demographic factors (e.g., nationality, gender, age, academic qualifications, occupation) were also included.
We used Mann Whitney U, Spearman correlation and Kruskal Wallis + post hoc pairwise comparisons as statistical tests for analyses purposes. The demographics of the study were presented using percentages. We explored the association of gender, age, country economic status, academic qualifications and medical physics practices on the responses ( Table 1). The data were re-categorised based on age groups (> 56-year-old, 40-56-year-old, 24-39-year-old, < 24-yearold), country economic status, academic qualifications and medical physics practices.

Demographics
A total of 1019 respondents from 94 countries participated in this survey. The highest number of respondents was from the UK (10%), followed by Malaysia (8%), the USA (8%), Australia (6%) and Japan (5%). Figure 1 shows the geographical distribution of survey respondents. Thirty-five percent (35%) of the respondents were female and 65% of the respondents were male. The highest proportion of respondents came from the 24-39-year-old age range, contributing to 59% of the total respondents. The second largest group were from the 40-56-year-old age range (28%), followed by the > 56-yearold age range (11%).
In terms of academic qualifications, 91% of the respondents held one or more postgraduate degrees, including 34% who held a PhD degree. However, only 5% of the respondents reported being board-certified. Seventy-nine percent (79%) of the respondents were currently working as academic and/or clinical physicists, while 10% carried multiple roles. Eleven percent (11%) of the respondents were currently (at the time of the survey) undertaking a postgraduate course. Sixty-seven percent (67%) were affiliated with hospitals, and 34% were working in universities. The rest of the respondents were affiliated with research institutes, consultancy, government agencies, regulatory bodies, etc. There were also 20% of respondents who had multiple affiliations. Figure 2 shows the survey summary for questions Q1 to Q10, and Q13. An exceedingly high number of survey respondents (91%) agreed that AI would play an essential role in medical physicists' practices. In comparison, 85% agreed that more and more applications such as QC, treatment planning would be performed by AI.

Survey results
When asked about their opinions if AI would threaten/ disrupt the medical physicists' practices and career, the polls were more spread out. Half of the respondents did not agree with this notion, while 25% of the respondents thought otherwise. Albeit the apprehension, 92% of the respondents acknowledged that all medical physicists should acquire some basic knowledge of AI. Towards that end, 87% of the respondents expressed support that AI should be taught in postgraduate medical physics programmes.
More than half (62%) of the respondents reported that they had a basic understanding of AI relevant to their field, while 21% of the respondents did not agree. However, only 34% of the respondents could confidently said that they had a working knowledge of AI. In comparison, 24% of the respondents neither agreed nor disagreed, and 41% of the respondents did not have a working knowledge of AI. In terms of skills, only 22% of the respondents reported that they had the relevant expertise in AI, while 25% of the respondents were undecided, and 53% of the respondents reported that they did not have the appropriate skills. Upon further probing, only 14% of the respondents could confidently reported that they were proficient in AI, were able to design, code, and implement an AI program; 15% of the respondents were undecided, while 71% of the respondents did not agree. Figure 3 shows that AI knowledge was mainly gained through self-taught (41%) and work-related activities (25%). Attending courses and postgraduate training only contributed to 12% and 5% of the respondents, respectively. We also noticed similar approaches towards developing AI skills, with 39% of the respondents reported being self-taught. Attending courses and postgraduate training only contributed to 10% and 7% of the respondents, respectively. Interestingly, while only 16% of the respondents claimed that they did not know AI, a more significant percentage (40%) reported that they had no AI skill. Contemporary AI systems were not without limitations. Just over half of the respondents (59%) reported that they understood the limitations of AI, while 16% of the respondents did not agree. In terms of the acceptance of AI, 67% of the respondents reported they were ready to learn and apply AI in their practices, while 15% of the respondents did not agree. Table 2 shows the results of statistical tests on the different factors affecting the levels of agreement and method of knowledge and skill acquisitions, comparing across gender, age groups, country economy status, academic qualifications and practices.
Males, more often, strongly agreed that AI would play an essential role in the medical physics practices, and they possess the working knowledge and skills in AI (Fig. 3). Gender association was found in the method of knowledge acquisition. More male respondents reported that they were self-taught, while more females learned from work-related activities.
There were weak correlations between the age groups with Q1, 7 and 13. All respondents from the < 24-year-old age group were convinced that AI would play an essential role in medical physics practices, while 2%, 4% and 3% of the respondents from the 24-39-year-old, 40-56-year-old, and > 56-year-old age groups, disagreed with this statement, respectively. Older medical physicists tended to report that they had a working knowledge of AI relevant to their field with the highest percentage of respondents in the > 56-yearold age group (40%), compared to only 14% respondents in the youngest age group. However, 29% of the respondents from the < 24-year-old age group strongly agreed that they were ready to learn and apply AI to their practices, compared to 17% respondents from the > 56-year-old age group.
The country economic status affected the perception of medical physicists. Higher income countries correlated with higher academic qualifications (r = − 0.210, p < 0.001). Forty percent (40%) of the respondents from high-income countries held a PhD degree, followed by upper-middle income (34%), low-income (20%) and lowermiddle income countries (18%). Medical physicists from higher income countries were more likely to agree that QC and treatment planning would be taken over by AI, as shown as by 34% respondents from high-income and 37% from upper-middle-income countries compared to 18% in low-income countries. Sixty-three percent (63%) of respondents from high-income countries agreed that they had a basic understanding of AI relevant to their field. In contrast, only 36% of respondents from the lowincome countries agreed to this statement. However, 75% of respondents from high-income countries disagreed that they could design, code and implement AI compared to 64% respondents from low-income countries. All respondents from the low-income countries believed that all medical physicists should acquire some basic knowledge of AI and that AI courses should be taught in the postgraduate programme while only 92% of the respondents from highincome countries agreed that medical physicists should acquire some basic knowledge of AI and even less (86%) thought that AI education should be implemented in postgraduate programmes. Respondents from low-and lowmiddle-income countries (73% and 81%, respectively) were also keen to learn and apply AI in their practices.

3
The country economic status association was found in the method of skill acquisition. While more than 40% of the respondents self-taught their AI skills, 17% of respondents from the upper-middle-income countries claimed that they learned by attending courses while 18% of respondents from the low-income countries reported learning from postgraduate training.
Significant correlations were found between academic qualifications with Q2, 6-8, and Q10. Medical physicists who held a PhD degree were more likely to think that AI would have more and more application in the medical physics profession (86%). They were also more likely to say that they had a basic understanding (74%), working knowledge (43%), and skills of AI (28%), and they understood the limitations of AI (70%). An almost similar proportion of respondents who held only an undergraduate degree (at the time of this study) agreed with the Q2 statement (83%) and understood the limitation of AI (63%). Nevertheless, a lesser proportion of the respondents reported that they had a basic understanding (51%), working knowledge (29%), and skills of AI (22%). Analysis of questions Q6 to Q10 showed a significant difference in academic, clinical physicists' opinions and those involved in both areas, henceforth called mixed-field physicists. Pairwise comparisons showed that academic (70%) and medical physicists who were involved in both academic and clinical work (75%) were more likely to agree that they have a basic understanding compared to clinical physicists (58%). An about equal percentage of the academic (44%) and mixed-field physicists (45%) reported to having working knowledge relevant to their fields, compared to 31% of the clinical physicists. Similarly, 35% of the mixed-field physicists and 31% of the academic physicists reported having relevant skills in AI, compared to 18% of clinical physicists. This follows that only 11% of the clinical physicists were proficient in designing, coding, and implementing AI, compared to 22% of the academics and 23% of those involved in both academics and clinical work.
Significant difference was found in the preferred methods of knowledge acquisition (χ 2 = 36.4, p < 0.001) and skill development (χ 2 = 28.1, p < 0.001). While the academic (59%) and clinical physicists (37%) were mostly self-taught, the mixed-field physicists do not have any particular preferences. Eighteen percent (18%) of the clinical physicists claimed they have no knowledge or skill in AI.

Discussion
Our studies found that almost all respondents agreed that AI is essential and would only continue to play increasing roles in medical physics. More and more tasks such as QC and treatment planning can be performed by AI systems, supporting medical physicists' work. Some of the examples quoted by the survey respondents included that AI can help automate and speed up processes, thereby allowing medical physicists to focus on areas that required improvement. In countries suffering from a shortage of qualified medical physicists, AI could help fill the human resource deficiency.
A pertinent concern raised by respondents of this survey was that dependencies on AI in handling crucial tasks in clinical medical physics might result in the decline in the competencies of the physicists and thereby leading to patient safety issues.
Many respondents of this survey reported believing that AI would change medical physics practices; if not disruptive, it would most certainly change the current state of practice. However, only half of the respondents agreed that AI would threaten/disrupt the medical physicists' practices and career. AI can be viewed optimistically as another tool that can help enhance the role of the physicist, although the results of this survey indicated that there was a certain amount of distrust of AI, which may stem from the lack of knowledge around it.
Given the inevitable presence of AI in medical physics, it is important that medical physicists are prepared for this future. The survey showed that about almost one-half of the respondents reported not having a working knowledge of AI and up to three-quarter of them reported not able to design, code and implement AI programmes. To address this gap, education is the key [10,11]. Currently, medical physics is often taught as postgraduate programmes [21]. While students may be exposed to some form of computing or programming courses in the programme, the emphasis on AI may not be extensive [5,10,13].
This survey shows that while most agreed that AI knowledge should be taught to new generations of medical physicists, the approach to do this was still being debated. Many respondents argued that AI and machine learning courses should be taught in an undergraduate course instead of a postgraduate course. One of the reasons cited was that postgraduate courses are usually one to two years, covering all conventional aspects of medical physics, and would be unable to provide extensive education and training on practical AI and machine learning techniques. However, the majority contend that medical physics should be taught in the graduate pogrammes. For example, one of the respondents wrote, "AI, along with many other disciplines such as signal processing, are profoundly related to the practice of medical physics. These areas are something that a simple one-week summer school will not be sufficient to impart requisite knowledge. It is extremely important that medical physics graduate programmes be augmented with a diverse range of subjects that, among other areas, cover AI and its applications. We are a discipline that cannot simply afford and rely on an approach that lacks evolution in time. The best place where we can equip the next generation of medical physicists is graduate programmes that can provide a diverse and solid knowledge base to tackle practical problems of the real-world practices of medical physics. This type of academic preparedness supplemented with clinical training as in resi-dency programmes is THE best way to produce the best clinical physics practitioners." Questions that asked about the respondents understanding and skills of AI showed an insightful trend. While 62% of the respondents claimed to have a basic knowledge of AI, the number quickly reduced by half for those who have a working knowledge of AI. Understanding how AI works do not mean one has the necessary skill to be a computer programmer, which in this case was a mere 14% of the respondents. The personal experience and observations of the authors indicate that many medical physicists have used or at least were acquainted with AI that comes as part of the vendor-supplied packages, rather than coding and designing in-house AI programme.
In terms of training and education in AI, the survey showed that many medical physicists (41%) acquired and developed their knowledge and skills through their own effort, i.e. self-taught. Perhaps this was due to the lack of AI education and training in the previous and current medical physics education programmes. Notably, there was an imbalance in gender concerning the reported level of knowledge and skills in the area of AI. This may be related to the mode of knowledge and skill acquisition as more male medical physicists reported having been self-taught. In contrast, female medical physicists reported having learned from work-related activities. These results suggested that if AI knowledge and skills were delivered more systematically, AI knowledge could be higher in female medical physicists with precise medical physics practice applications. Generally, it may be advisable to deliver AI training, like other physics training, to improve training outcomes for all via multiple pathways.

Conclusion
The general perception of the medical physicists surveyed in this study was that AI is here to stay, and it would definitely influence our practice. Respondents reported believing that AI is likely enhance rather than replace medical physicists. This survey suggests that medical physicists should be prepared with education and training to face a new reality where medical physicists work with and create AI systems.
Global medical physics organisations can play a significant role in providing up-to-date information on AI in medical physics clinical practice guidelines. Leadership in steering the medical physics profession in preparation for the AI-enhanced future is advisable.

Appendix
Please rate your agreement with the following questions: *Answer range from: 1(Strongly agree) to 5 (Strongly disagree) 1 AI will play an important role in the practice of medical physicists.* 2 More and more applications such as quality control, treatment planning will be performed by AI.* 3 In my opinion, AI will threaten/disrupt the medical physicists' practices and career.* 4 All medical physicists should acquire at least some basic knowledge of AI.* 5 AI should be taught in the postgraduate medical physics programme.