Radiological education in the era of artificial intelligence: A review

Abstract

In the era of artificial intelligence (AI), a great deal of attention is being paid to AI in radiological practice. There are a large number of AI products on the radiological market based on X-rays, computed tomography, magnetic resonance imaging, and ultrasound. AI will not only change the way of radiological practice but also the way of radiological education. It is still not clearly defined about the exact role AI will play in radiological practice, but it will certainly be consolidated into radiological education in the foreseeable future. However, there are few literatures that have comprehensively summarized the attitudes, opportunities and challenges that AI can pose in the different training phases of radiologists, from university education to continuing education. Herein, we describe medical students’ attitudes towards AI, summarize the role of AI in radiological education, and analyze the challenges that AI can pose in radiological education.

1. Introduction

The word “artificial intelligence (AI)” was coined more than 60 years ago. AI can be simply defined as “intelligence achieved by machines” and is a branch of computer science. In the past few years, AI has generated great expectations in medicine,^[1] especially radiology.^[2] The emerging field of AI has attracted great attention in radiology.^[3,4] Radiology-based research using AI has developed rapidly, with the rapid growth in the number of publications on “AI” and “radiology” as well as the increase of the discussions about AI topics at various international radiology forums.^[5] Moreover, the number of radiology AI products has rapidly increased in the past few years.^[6] There are currently more than 150 AI products on the radiological market based on X-rays, computed tomography, magnetic resonance imaging, and ultrasound.^[7] The areas of AI applications in radiology can be classified as follows: Image segmentation, lesions detection; Labeling, measurement, and comparison with historical lesions; Generating radiological reports; Semantic error detection in radiological reports; Workflow analysis, outcomes measures and performance assessment; and Data mining in research.^[6,8,9]

The society of radiological sciences has already been aware of the possible impact of AI on our profession, and various media outlets have predicted the gradual disappearance of radiologists with the advent of AI and the wide application of AI in radiology.^[10–12] There are even doubts about the value of continuing to train radiologists.^[13,14] Nevertheless, there seems to be a consensus that radiologists must continue to be trained. However, novel AI knowledge and skills should be incorporated into the training for radiologists, initiated in the university phase, strengthened in the residency stage, and maintained in the continuing education stage after graduation.^[15,16] However, there are few literatures that have comprehensively summarized the attitudes, opportunities and challenges that AI can pose in the different training phases of radiologists, from university education to continuing education. Therefore, in this article, we describe medical students’ attitudes towards AI, summarize the role of AI in radiological education, and analyze the challenges that AI can pose in radiological education.

2. Methods

This is a narrative review where search parameters were used. Searches of PubMed were performed to obtain the data and articles for this review. The search terms used were “radiology,” “education,” “artificial intelligence,” “AI,” “medical students,” “attitude,” “role,” and “challenge.” Abstracts and articles were reviewed and included only if they met our criteria of discussing medical students’ attitudes towards AI, the role of AI in radiological education, and the challenges that AI can pose in radiological education in the last 10 years.

3. Medical students’ attitudes towards artificial intelligence in radiology

In the practice of radiology, AI can be used to solve many practical problems of radiology, including; Reducing missed diagnoses; Improving diagnostic accuracy; Screening out the cases from work lists that need to be prioritized; Improving the quality of reconstructed images, and; Extracting features from images that are not visible to the naked eye.

Overall, the majority of medical students believe that AI will have a significant impact on the practice of radiology.^[17] This impact will be positive for most radiologists, such as reducing errors and time spent on each diagnostic report, and increasing the time dedicated to each patient.^[17,18] In recent years, many surveys have been studied to describe the attitudes of medical students towards AI in radiology (Table 1).^[17,19–22] The viewpoint that radiologists will be replaced by AI is a minority, but at least a significant portion of these groups believe AI will lower their professional expectations.^[19,23] In this sense, several studies showed that the predictable impact of AI discouraged medical students from selecting radiology as a future career, and brought anxiety even among those who wanted to select radiology as their future career.^[20,24] However, there is heterogeneity in the view of AI replacing radiologists among medical students. Gong et al’s study showed that 29.3% of medical students believed radiologists would be replaced by AI in the future, 67.7% believed the demand for radiologists would be reduced by AI, and almost half of medical students (48.6%) agreed AI caused anxiety when considering the radiology as a specialty.^[19] Moreover, Sit et al’s study found that 49% of medical students report they were less likely to consider radiology as a career in due to AI applications.^[21] However, it is worth noting that Sit et al’s study demonstrated that those students who received AI teaching were more likely to consider radiology as a career compared to those who did not receive AI teaching.^[21] In addition, tech-savvy medical students were less concerned about the potential negative effects of AI in radiology compared to those who were not tech-savvy.^[17] Only 15.2% (40/263) of medical students believed that radiologists would be replaced by AI.^[17] Furthermore, a majority (77.2%) of medical students agreed that AI would revolutionize radiology,^[17] and a further majority (67.7%) believed AI would reduce the demand for radiologists^[19] while this registered neutral when expanded to 20 years.^[20] In addition, a majority of medical students (8/10 on a 10-point Likert scale) stated that AI should be used as a support for evaluating radiological images.^[20] However, in the study of van Hoek et al, they found medical students showed a neutral attitude (median = 1) towards the role of AI in replacing radiologists using a 21-point Likert scale (-10–10: -10 = strongly disagree, 10 = strongly agree).^[20] Understanding the importance of AI, a majority of medical students (80%) agreed that teaching in AI would be beneficial for their careers, and a further majority (78%) believed that students should receive training in AI in their university phase.^[21]

Table 1.

Medical students’ attitudes towards AI in radiology.

First author (yr)	Survey study design	Population	Study results
Pinto dos Santos (2019) (17)	Electronic survey using the SurveyMonkey web-application	Three major German universities (263 respondents)	(1) The majority agreed that AI will revolutionize (77%) and improve (86%) radiology.
			(2) The majority (83%) disagreed that AI will replace radiologists.
			(3) The majority (71%) agreed on the need for AI to be included in medical training.
			(4) Male and tech-savvy students were more confident on the benefits of AI and less fearful of AI.
Gong (2019) (19)	Anonymous online survey	Seventeen Canadian medical universities (322 respondents)	(1) The minority (29.3%) agreed that AI would replace radiologists in the future.
			(2) The majority (67.7%) agreed that AI would reduce demand for radiologists.
			(3) Almost half (48.6%) agreed AI caused anxiety when choosing radiology as a speciality.
			(4) The minority (16.7%) would rank radiology first if it were not for anxiety about AI.
van Hoek (2019) (20)	Online questionnaire using SurveyMonkey platform	Medical students (55 respondents)	(1) AI should be used as a support for evaluating radiological images (Median point: 8, Likert scales: 0–10).
			(2) If AI achieves high diagnostic accuracy, it should be used to evaluate radiological images alone (Median point: 3, Likert scales: 0–10).
			(3) In 20 years, there will be more/fewer diagnostic radiologists (Median point: -2, Likert scales: -10-10).
			(4) In 20 years, there will be more/ fewer interventional radiologists than today (Median point: 2, Likert scales: -10-10).
			(5) The profession of diagnostic radiologists will be endangered in the future by AI (Median point: 1, Likert scales: -10-10).
Sit (2020) (21)	Anonymous electronic survey consisting of Likert and dichotomous questions	Nineteen United Kingdom medical universities (484 respondents)	(1) 49.2% agreed that they were less likely to consider radiology as a career due to AI.
			(2) 27.1% agreed that they were not less likely to consider radiology as a career due to AI.
			(3) Students with no teaching in AI were less likely to consider radiology as a career.
			(4) Understanding of AI: 44.6% agree, 43.4% disagree, and 12.4% were neutral.
			(5) Understanding limitations of AI: 48.3% agree, 30.4% disagree; and 21.3% were neutral.
Bin Dahmash (2020) (22)	Cross-sectional multicenter survey	Three medical universities in Riyadh, Saudi Arabia (476 respondents)	(1) 31% believed that AI would replace radiologists in their lifetime.
			(2) 44.8% believed that AI would minimize the number of radiologists.
			(3) About 50% believed they had a good understanding of AI.
			(4) 22% of the questions were answered correctly when knowledge of AI was tested.
			(5) 58.8% were anxious about the uncertain impact of AI on radiology among the respondents who ranked radiology as their first choice.

AI = artificial intelligence.

4. The role of artificial intelligence in radiological education

Contrary to the argument that AI will replace radiologists, “AI will empower education” is proposed by Duong et al,^[25] suggesting how AI can compensate for the shortcomings of the current apprenticeship model in the medical education.^[26] With the widespread applications of AI in radiological practice, there are many potential ways AI can be used for radiological education. AI can obtain a large amount of data about residents’ education, performance and progress through training, be tailored to individual trainees based on trainees’ learning styles and needs, and make precision education in radiology possible.^[25]

In recent years, several new educational approaches, such as flipped classroom, virtual education, permeated radiology. Incorporating AI into radiological education may provide novel ways for residents to stimulate interest in learning and improve learning efficiency. For example, gaming techniques are increasingly being applied in radiological education programs.^[27] Radiologic trainees will receive rewards through an online platform after completing the activities, such as completing milestone exams, and passing online modules. AI can help trainees track their progress by automatically identifying activities conducted or milestones achieved and recording them. AI may constantly change the case difficulty based on the knowledge and performance of trainees, providing cases best suited to enhance trainees’ learning.^[27] Zhang et al proposes to build trainee models that utilize features automatically extracted from mammography images using a computer-aided education system.^[28] According to the trainee’s previous performance and imaging characteristics of the lesion, including tissue strength, size, location, similarity to adjacent areas, and symmetry with the contralateral side, this AI-aided education system allows for predicting the likelihood of missing a mass on mammography and automatically searching databases of mammograms to select difficult cases for the trainees in an automatic and efficient method. As its algorithm continues to be improved, and the number and complexity of extracted imaging features grows, AI can present residents with optimized cases on the basis of their specific “profile” and their tendency to miss lesions.^[28]

In addition, the radiological reports are the final products produced by the radiologists to communicate clinically possible disease diagnosis to the referring physician (and patient) after the imaging evaluation. In diagnostic radiology, residents must learn many aspects of radiological practice, including writing radiological reports of sufficient clinical cases. AI can personalize radiological learning by tracking resident competency profile and reinforcing challenging topics, acting as an “intelligent tutor.” For instance, AI can initially interpret a case and assign it to a resident whose profile shows potential benefits, then guide the resident to review similar case reports combined with relevant literatures. After discussion and attending review, the case will be included to the teaching file and the competency profile of the resident will be updated.^[25] This “live teaching file cataloging” can be of great use in creating a substantial case-based database and improving the diversity of cases encountered by residents. Increased case exposure and volume with related feedback can facilitate the accumulation of expertise. In addition, many radiological residency programs provide case-specific feedback to residents about their radiological reports and create automated case log and volume analytics feedback to reduce the necessity for manual recording.^[29,30] These feedbacks are designed to give radiological residents the opportunity to review cases they missed or misunderstood, and to provide valuable learning opportunities. Using AI to analyze residents’ performance, there is more significant potential to provide personalized feedback based on daytime rotation performance or off-hours call performance.

5. The challenges of artificial intelligence in radiological education

In the radiological practices, the radiological reports are initially written by residents using many AI tools (such as AI-aided lung nodule detection tool) as AI products of radiology increase rapidly. Although many tedious and repetitive tasks are significantly reduced for residents attributed to AI applications,^[11,16] there is a doubt as to whether it is wise for a resident to delegate these tasks to the software without first understanding them and performing them autonomously. With AI, radiologists can be freed from the tedious work, and they will have more time to interact with patients, participate in multidisciplinary committees, and work on research and data analysis.^{[14,16,24,31]} However, suppose we introduce AI tools to residents prematurely and indiscriminately, without human oversight and without analyzing their strengths or limitations. In that case, inexperienced radiological residents may conclude that AI is “‘better’” than repetitive effort and human intellectual work, thus making them over-dependent on AI tools.^[32] Excessive and uncontrolled automation attributed to AI applications reduces our ability to learn, interpret, and critically analyze. Therefore, we hold the opinion that AI applications for residents may be desirable only when a resident has processed a sufficient number of relevant imaging studies and then can access AI tools. Meanwhile, residents should familiarize themselves with how the algorithms of AI tools work. For example, future residents should be familiar with how, when, and why the AI tool might fail. In addition, another challenge in training future radiologists is teaching trainees to recognize AI errors.^[33–35] Given the nature of AI technology, its errors may be more subtle, unpredictable, and unrepeatable.^[35] The idea that AI tools are more correct than human beings will be more prominent when human practitioners are less confident, which may have a more significant impact on inexperienced radiological residents.^[36] Therefore, in the future, radiological residents should be taught to be vigilant about this automation bias.^[32]

6. Conclusions and perspectives

Recent studies have shown that AI will be used to deliver educational content based on the needs of radiological trainees (so-called personalized precision education), which may achieve greater standardization and harmonization in the acquisition of interpretive skills.^{[10,16,25,37]} Furthermore, as the most tedious and repetitive tasks are freed up attributed to AI, there will be more time for teaching and research.^[37,38] AI will reduce daily tasks (such as acquisition theory knowledge, examination evaluation, and rotation) for both teachers and trainees, and enable them to pay attention to tasks with greater added value (such as learning habits, interviews, research, and communication skills).^[12,14,31]

Whether you like it or not, AI is here to stay. AI will change not only the way we work, but also the way we teach and learn. If we embrace AI as we have embraced other advanced technology in the past (picture archiving and communication system, new imaging modalities, and structured reporting, etc), we will be able to improve the performance of our work and pay more attention to tasks with added value and positively impacting patient care and training medical students and residents. However, AI will be another noninterpretive skill that radiological trainees need to learn. In addition to learning how to interpret radiologic images, current and future radiological trainees need to learn how to interpret AI outputs. They are required to know how AI alters clinical workflows and be sensitive to unusual AI outputs. It is still not clearly defined about the exact role AI will play in the future practice of radiology. However, it will certainly be consolidated into radiological education in the foreseeable future.

Acknowledgments

This work is supported by Zhejiang Provincial Natural Science Foundation of China under Grant No. LY21H180003.

Author contributions

Conceptualization: Chao Wang, Ling Hu.

Data curation: Chao Wang, Huanhuan Xie, Shan Wang, Siyu Yang.

Funding acquisition: Chao Wang.

Investigation: Shan Wang.

Methodology: Chao Wang.

Resources: Huanhuan Xie, Siyu Yang.

Supervision: Chao Wang.

Writing – original draft: Chao Wang.

Writing – review & editing: Ling Hu.