Training opportunities of artificial intelligence (AI) in radiology: a systematic review

In this study, we use “training program” as an umbrella term for multiple formats in which a training is organized. It includes on-site and online courses, both educational seminars and webinars, and workshops or lectures. For being included, a program has to (1) be about artificial intelligence or other related topics such as deep learning, machine learning, and learning algorithms; (2) target radiologists as the audience; (3) have specific learning objectives; (4) specify the time and ways of delivery; and (5) have specific instructors. We therefore excluded the presentations, talks, videos, and other materials that did not have one of these criteria.

We searched for online information about the training programs, basically via Google, LinkedIn, and PubMed platforms, using a series of keywords (see Table 1). We also searched on the website of various radiological associations and organizations (see Table A6 in Appendix 1). We checked our initial list of training programs with four experts, who are active in offering AI training in radiology, which helped us enrich the list and find new programs. We initially found 130 potentially relevant programs. We excluded the programs about which limited information was available. Eventually, we collected a sample of 100 training programs, offered in 2019 and 2020 (until end of June), which appeared to cover a wide range of programs in terms of content and offering agents.

We collected qualitative data about each training program by examining their general aspects (e.g., date, duration, and price), content, covered topics, target audience, instructors, offering agents, and ways of legitimizing (see Table 2). The qualitative data on each of these dimensions was archived in an excel file (see Appendix 2) for further analysis.

We used deductive thematic analysis [8] to analyze the data, based on the codebook (Table 2). One coder coded the entire data and a second coder cross-checked the results and resolved the ambiguities. The results were organized in a database (Appendix 2), in which we explored patterns based on cross-tabulation and comparisons between the training programs [9]. We interacted with three experts in the domain of AI and radiology to validate the relevance and richness of the extracted patterns.

The number of AI trainings for radiologists has increased in the last 1.5 years (Fig. 1). Most of the training programs are offered purely online (70%), and only a small fraction of them are offline (13%). Half of these programs are offered for free; for the rest, the average price per program is less than 100 Euro. Around 60% of the programs are shorter than 1 h and only 20% of them take longer than 3 h. Overall, most of the programs are short, stand-alone sessions, which are not part of any larger curriculum (see Tables A1 and A2 in Appendix 1).

The analysis of their content (Fig. 2) shows that a majority of the training programs focus on passive offering of the content about the conceptual aspects of AI such as the basics of machine learning and its potentials for medical practice in general (not necessarily related to radiology). Around half of the programs go further and discuss some potential applications of AI in radiology, e.g., by showing how artificial intelligence can be applied to diagnostic imaging, and how AI can be applied to a particular modality (CT, MRI) or giving examples of AI applications in clinical radiology. However, only a small fraction of the programs practically engages the learners in hands-on practices of working with some tools. There, the focus is on practicing the basics of writing code (e.g., programming in Python) or training an algorithm based on some image data. Few programs focus on learning how to perform the actual medical work with real AI applications (see Table A3 in Appendix 1).

We clustered the topics that these programs cover into (1) medical (how AI can support medical diagnosis), (2) technical (how AI can be developed, tested, and configured), (3) legal/ethical (what are the legal and ethical considerations of implementing and using AI), and (4) managerial (e.g., integration of AI in the radiological workflow, how to best deploy AI and the inherent challenges) aspects (see Table A4 and Fig. A1 in Appendix 1). Overall, the focus of many programs is on a combination of medical and technical topics, sometimes combined with peripheral discussions about the legal/ethical and/or managerial aspects of implementing AI. Although managerial topics are covered in around half of the programs, they are often side discussions.

Although we selected the training programs that are offered to radiologists, most of them (76%) target other medical professionals (e.g., other specialization, GPs, medical researchers) and even non-medical professionals (e.g., data scientists, IT specialists, managers) as well (Fig. 3). In fact, most of the programs are quite generic (e.g., on the basics of AI), thus are promoted as relevant for a wider set of audience than merely radiologists. Only in 7% of the programs radiographers are explicitly included as the audience. When training programs merely focus on radiologists, their content becomes more connected to the medical cases and the integration of AI into the daily workflow.

The instructors have very diverse backgrounds such as “professors of radiology,” “founders of AI companies,” and “researchers” from R&D departments. Half of the programs are jointly offered by the instructors with medical and technical backgrounds (Fig. 4). Only one-third of the programs are offered by radiologists and very few instructors come from the social and managerial domains (see Table A5 in Appendix 1 for an overview of the instructors’ backgrounds). Sometimes, people with medical background offer technical content (e.g., basics of machine learning).

The providing agents (Fig. 5 and Table A6 in Appendix 1) are primarily professional institutions such as radiology associations and societies (59%). Commercial companies such as AI developers and commercial training agents are also active in offering AI. Academic institutes are marginally involved (only 8%).

Since AI is not officially part of the radiology curriculum yet, the providing agents need to use various strategies to legitimize their programs as relevant and valuable for radiologists (see Table A7 in Appendix 1). Our analysis shows that most of the training programs use multiple strategies to legitimize their programs, including (1) showing the importance of the content (90%), (2) appealing to the status of the providing agents and instructors (99%), (3) offering accreditations and certificates for participants (73%),⁵ and (4) showcasing the positive feedback and acknowledgements of the prior participants (47%). Half of the formal recognitions that these programs offer are medically recognized accreditations such as CME, approved by EACME or AACME. As Fig. 6 shows, professional institutions are using the first three strategies to a high extent, whereas academic and commercial institutions mainly focus on legitimizing their programs based on the content and the status of the providing agents.

Our analysis shows that most AI training programs offered to radiologists are primarily short, stand-alone, and fragmented.⁶ Although they are useful for creating awareness and familiarizing radiologists with basic concepts, they still need to be complemented with programs that are coherent and pedagogically integrated into the general radiology curriculum. As we are in the stage of reframing the radiology profession, we need to develop training programs that are based on scientific research and offer an effective learning trajectory for radiologists who want to invest in long-term development of their career. Some initiatives such as the ESR-ETC can serve as guiding frameworks for designing the programs that complement each other. This requires a more active involvement of academic institutes in designing and offering the training programs that are based on scientific research and aligned with pedagogical principles.⁷ When these training opportunities are offered systematically as part of the national radiology curricula, they could be incorporated into professional examinations such as European Diploma in Radiology (EDiR). Here, some regional and international institutions such as ESR can play a strong role by offering relevant courses needed for the EDiR examination.

Our findings show that the majority of the programs focus on passive delivery of content on the basics of AI and its potential impacts on radiology work. As AI gradually finds its way in the daily practice of radiologists, we need to engage learners in practical exercises with real AI applications and learning how to use them effectively and critically at their work. Training programs need to focus more on real use-cases, and (close to) real work settings and applications. Learning about the basics of machine learning, though might offer some basic knowledge, is not enough for radiologists who will be “working with,” rather than “developing,” the AI applications. Learning how to critically use and effectively integrate AI applications into the working routines requires developing new mental and practical skills. Developing these skills requires the involvement of not only medical and technical instructors, but also experts from organizational, legal, ethical, and psychological domains. Above all, younger radiologists need to be supported by programs that enable them to strategically design and develop their professional career for future.

Finally, our study shows that most of the current training is generic in terms of their content and target audience. Although these generic programs are effective for creating some basic familiarity, they are limited to develop knowledge and skills that are specific to the context of radiology work. Further programs can be more specialized on the specific use-cases that radiologists face in using AI and consider the special needs of radiologists (e.g., when residents need to check their examinations with a senior radiologist). Furthermore, we need to develop training programs that are customized and localized to the specific clinical practices and organizational settings. Especially in Europe, the working conditions, legal frameworks, and workflow configurations vary from one country to another. All these differences require customized training programs that go beyond the generic ideas and consider the specific organizational settings (e.g., technological infrastructure), legal frameworks (e.g., privacy regulations), and cultural aspects (e.g., patients’ and referring physicians’ expectations of and trusts in AI applications). Table 3 offers a list of suggestions for developing AI training.

Although we examined a diverse sample of training programs, our focus on LinkedIn (instead of Twitter) may have biased our sample towards Europe [10]. Future studies can also examine the other offerings such as podcasts and knowledge clips, which, although are not officially framed as training programs, are often used by radiologists to learn about AI [11]. In addition, since many training programs are still emerging, future studies need to continuously examine new trends and update our findings. Next to radiologists, other professional groups such as radiographers, clinical physicists, and technical physicians are learning how to use AI for supporting various tasks in the acquisition of the images and pre-processing them. Knowing how they learn to work with AI and what skills they can develop is important to understand how AI can be effectively used in radiology.