Sage Journals: Discover world-class research

Abstract

This paper follows the reaction of the radiology profession to artificial intelligence (AI). We examine the effort of radiology as a powerful medical specialty to maintain its professional jurisdiction while allowing AI's disruption. We study the discursive work of radiologists as evident in their academic publications. Our results suggest that radiologists hold simultaneously multiple perspectives in regard to AI, which allow them to be both conservative and innovative in their relations to it: accept it, subordinate it, reject it and surrender to it, all the same time. These perspectives are: (a) to integrate AI tools and skills into the radiology profession by cooperating and coproducing with AI experts while preserving the core values and structures of the radiology profession; (b) to absorb AI into radiology as (yet another) technology, subordinating it to radiologists’ authority; (c) to fight-off the threat made by AI by undermining the legitimacy and capabilities of AI in radiology and strengthening professional boundaries and (d) to assimilate the radiology profession into the field of AI. These perspectives enable radiologists as a powerful medical specialty to engage in a rhetorical dance with the equally powerful AI specialty and challenge techno-optimistic approaches to innovation.

Keywords

Institutional maintenance professions artificial intelligence radiology

Introduction

The entrance of artificial intelligence (AI)¹ into expert labor is becoming a major topic in the study of professional work (Brayne, 2017; Browder et al., 2022; Christin, 2017; Faulconbridge et al., 2021; Goto, 2022; Lange et al., 2019; Lebovitz et al., 2022). In low and middle-status occupations, workers are more easily subjected to algorithmic management (Kellog et al., 2020; Möhlmann et al., 2021). However, in high-status professions, experts have an advantageous position (Suddaby and Viale, 2011), where they benefit from their competitive abstractions (Abbott, 1988), their tacit knowledge (Polanyi, 1966; Lebovitz et al., 2021) and sophisticated resistance power (Christin, 2017; Goto, 2022). All these complicate techno-optimistic projections of AI automating or replacing expert-professional labor (Susskind and Susskind, 2015). The medical profession is of special interest in this context as it is both powerful and territorial regarding its expertise domains, even in the advent of commercialization and increasing power of other stakeholders in healthcare (Parry and Parry, 2018; Timmermans and Oh, 2010). In this article, we focus on the reaction of the medical profession of radiology as AI enters its territory and aims to disrupt it.

The theory of institutional work, and specifically the aspects of maintenance work in professions (Currie et al. 2012; Dacin et al., 2010; Scott et al., 2000; Zilber, 2009), can contribute to the discussion of AI's disruption of expert-professional labor. The notion of maintenance in this framework refers to the “supporting, repairing, or recreating” of professions as institutions (Lawrence and Suddaby, 2006: 230) and is perceived as essential for the stability over time of social structures (Dacin et al., 2010). Institutional theory has repeatedly shown that professionals are well-versed in preserving their powerful positions in organizational and institutional environments (Adler and Kwon 2013; Murray, 2010). Specifically, Suddaby and Viale (2011) argue that in many cases, professionals are the ones who shape new fields of action, create identities, rules and standards, redefine a field's boundaries and reproduce its hierarchies and social capital (Currie et al., 2012; Malsch and Gendron, 2013). Hence, even in the face of the wide institutional disruption generated by AI, the power of radiology as a medical profession to reshape the imaging jurisdiction to its benefit should not be underestimated.

As studies of AI's disruption of expert labor show, maintenance work varies between different professional fields and is highly dependent on regulatory environments (Brayne and Christin, 2021), the level of professionalization of the profession, the particularized sensemaking of the technology (Goto, 2022; Lebovitz et al., 2022) and organizational factors (Brayne, 2017; Goto, 2022). In addition, the type of task at hand is also important. While quantitative “number crunching” and high-frequency trading have led to AI dominance in finance expert labor (Hansen, 2021; Lange et al., 2019; MacKenzie, 2021; Pasquale, 2015), meta-quantitative professions, which require non-numeric high integration, assembling, and cognitive skills based on associations, such as language comprehension (Goto, 2022), remain less vulnerable to current day AI technologies. However, even meta-quantitative and professionalized professions cannot reject AI completely due to its reported improvements and the institutional pressures for its acceptance (Brayne, 2017; Ribes et al., 2019).

In radiology specifically, improvements in computer vision based on deep learning (DL) algorithms have been described by prominent AI experts as a case of complete automation of a medical task by the technology (Hinton, 2016). AI experts have argued that the technology they produce will be able to replace radiologists’ work and diagnose imaging scans autonomously, eliminating the need for an expert radiologist performing this task in hospitals (Kim et al., 2021). This claim has understandably ensued a professional struggle (Abbott, 1988) between radiologists and AI experts over the imaging professional jurisdiction. The institutional pressures to use AI in expert labor, forces radiologists to engage with this disruptive technology while maintaining their expert position and traditional power structures.

In these circumstances, the institutional order is an ongoing process and thus “unfinished” (DiMaggio, 1988: 12). Therefore, institutions in general and professional projects in our case, require ongoing work and exchanges based on support, corrections or reactions among stakeholders (Zilber, 2009). Such mechanisms are needed for dealing with change and uncertainty and are crucial for recreating professional stability (Lawrence and Suddaby, 2006).

The maintenance work of radiologists presented in this article is thus dualistic and complex (DiMaggio, 1991); it is conducted through spinning a web of scientific arguments regarding AI experts and technologies, creating a rhetorical dance of integration, subordination, rejection and assimilation at the intersection of the two types of expertise. The dance metaphor (cf. Ortmann and Sydow, 2018; Wallenburg et al., 2021) encapsulates the constant movement of attraction-rejection dynamic between radiology and AI, the efforts to decipher the impact of the technology on radiologists’ work, and the containment of various voices in the professional community in the face of uncertainties involved in change. As previous studies of professional maintenance show (Murray, 2010; Wright et al., 2017, Zilber, 2009) intersections at the boundaries of powerful professions require sophisticated maneuverings and movements, both rhetorical and practical. For radiologists, the rhetorical dance in their scientific argumentations enables them to integrate the AI expertise with their expertise, while maintaining their professional boundaries.

The profession of radiology and its use of technology

Radiologists are responsible for reading and interpreting graphical imaging scans, produced by several types of technological instruments, mainly Roentgen, computed tomography (CT), magnetic resonance imaging (MRI) and ultrasound. In the modern hospital, almost every admitted patient requires the opinion of the radiologists (Jalal et al., 2019). Radiologists have been in the forefront of medical-technological change since the late 1960s (Barley, 1995). Compared with other medical specialties, they have little patient contact (Francis, 2008), and they face less malpractice claims (Chockley and Emanuel, 2016), which allows them the high status of a technological-medical, yet clinical, sub-specialty.

The social sciences focus on radiology mainly as a medical sub-specialty that relies heavily on technology (Barley, 1986; Burri 2008). Barley (1986) has shown that the introduction of a new technology into a radiology department in two hospitals may result in the restructuring of the division of labor between physicians-radiologists and technicians, and that the professional use of technology is a localized matter, not a universal one. Burri (2008) has shown the radiologists use visualization technology as a means to demonstrate professional skills and power, increase one's reputation, and renegotiate identity. This use, Burri argues, allows radiologists to achieve symbolic capital in their boundary work, namely their distinction from other medical professions. Rystedt et al. (2011) showed that the radiological expertise is inherently practical, highly situated and domain-specific. Yu and Levy (2010) demonstrated how the power of radiological associations and their ties to national institutions, is determinant in halting or submitting to globalization ambitions of private companies. These accounts of radiological expertise, its situated and socially constructed nature, point to the central role played by radiologists and their professional institutions in shaping the use of AI in their profession.

Radiology is now at the forefront of AI in medicine (Pesapane et al., 2018), as most FDA-approved AI devices (over 70%) are targeted at radiology (Zhu et al., 2021). Recently, Lebovitz et al. (2022) have observed radiologists using AI devices in hospitals and showed that opaque AI impacts routine radiology tasks to become non-routine, increase uncertainty and lead radiologists to use AI devices differently than their producers had anticipated, frequently dismissing their results. Lebovitz et al. (2021) also point to the tacit knowledge embedded in radiological expertise, missing from the labeled data used in AI devices. This missing knowledge results in AI's underperformance in clinical settings. Therefore, even in the face of AI institutional disruption, the agentic role played by the profession of radiology challenges the feasibility of complete automation of radiology tasks, in spite of AI experts’ rhetoric (Hinton, 2016) and managerial desire for algorithmic control (Kellogg et al., 2020). We therefore argue that despite engineering and managerial attempts to automate radiology tasks, as a medical profession, radiology is well-equipped to harness AI to its benefit and maintain its professional boundaries, privileges and power structures.

While ethnographic fieldwork provides us with on-the-ground assessments of technology-in-use (Bader and Kaiser, 2019; Brayne, 2017; Christin, 2017; Goto, 2022; Lange et al., 2019), the expertise-related reasons and justifications for the use (or dismissal), acceptance or rejection, of AI are mainly evident in the scientific discussions of the profession. As the ultimate legitimacy-generating mechanism in expert labor (Abbott 1988), peer-reviewed scientific journals provide professions with the thoroughly thought of reasoning and justifications, backed with empirical research, data and theoretical reasoning, which direct professional conduct. These scientific discussions are then used by practitioners as abstractions to justify their daily practices (Stevens et al., 2018). We therefore turn to the scientific publications in radiology to study this medical profession’s reaction to AI.

Research design and methods

Our study of professional maintenance focused on published reactions of radiologists that appeared as articles in top radiology academic journals. As the venue of formal discussion in established professions, peer-reviewed academic articles are an institutionalized arena to investigate professional perspectives. Read widely by the members of the profession, these publications institutionalize new knowledge and perspectives regarding professional boundaries. In addition, top journals have a high level of quality assessment and peer-review selection procedures. This assured us that we can sample the reactions and perceptions of top professionals in the field of radiology. Thus, our study examines the radiology profession's complex reactions to AI as they appear in 60 academic publications in the field of radiology.

We used two databases to select our sample. First, we collected articles with Google Scholar using the search words: “machine learning” and “radiology.” A Google Scholar search provides a bird eye's view of academic publications. Figure 1 below shows the significant increase in publications containing the search terms “machine learning” and “radiology” over the last decade.

Figure 1.

Number of articles with the terms ML and radiology for 2010–2020.

At the time we searched for these data (December 2020), there were about 355,000 results for this search. From this list, we reviewed the first 250 results sorted by relevancy, focusing on articles published in leading radiology journals, such as American College of Radiology, British Institute of Radiology, Canadian Association of Radiologists Journal, European Radiology, Pediatric Radiology, Abdominal Radiology, Clinical Radiology, Current Problems in Diagnostic Radiology, Journal of Magnetic Resonance Imaging and more. This range of journals allowed us to sample radiology journals from different countries and various sub-specializations. Based on their abstracts, we distinguished between articles reporting only technological applications of AI in radiology with no discussion of their justifications for choosing ML for to perform radiological studies (n = 184) and articles discussing opinions and justifications (n = 66), with or without an empirical study to back these opinions. Of the articles discussing opinions we randomly selected 30 articles as our first sample set.

Second, we consulted a librarian and the JCR (Journal Citation Report) ranking of academic journals, to distinguish prominent venues of publication in radiology. We then focused on the journal Radiology, (IF 29.16 for 2021) as one of the highest-ranked radiology journals and a host of the ongoing professional debate on AI. There were 50 results for the search “machine learning” in this journal since 2016 and we randomly sampled another 30 articles and added them to the Google Scholar sample as a second sample set, giving the two databases, Google Scholar and JCR, the same number in our sample data.

We viewed our dataset of radiology articles discussing AI as discursive arena where authors and their readers shape their views of AI. Following grounded theory principles and an interpretive approach (Creswell and Miller, 2000; Saunders et al., 2018; Strauss and Corbin, 1997), we engaged in close reading of the 60 articles, focused on 171 quotations, which are paragraph-level arguments radiologists make regarding AI, and identified categories in these arguments (Pope et al., 2000). We identified the main arguments independently and then compared our classifications to reach inter-judge reliability (Leiva et al., 2006). We labeled segments of articles first by their topic, for example, when radiologists discussed the interpretability of ML models, we categorized “Explainability,” or when they discussed the reaction of patients to automated systems, we categorized “Patients.” Next, we compared elements within general categories to discern the different types of discursively constructed relations between radiologists and AI. For example, in the category “Ethics” the phrase: “AI is based on unethical data agreements and privacy breaches” was compared with “AI will allow for more ethical service to patients.” In the general category “Patients” the phrase “patients would object to an algorithm reviewing and diagnosing their imaging scans,” was compared with “patients would behave as consumers, navigating between machine and human diagnosis and choosing their preferred option.” This type of comparisons allowed us to discern different themes in the perspectives of radiologists regarding AI (Gioia et al., 2013). As relational perspectives (relations of radiologists to AI), these themes represent perceptions of both “us” (radiologists) and “them” (AI experts/ AI technology), the relations (and power relations) between the two groups, and radiologists’ ideas regarding the appropriate responses of their profession to AI.

The below outline of these institutionalized perspectives is by no means exhaustive; more perspectives may be forming in other venues, not covered by the data presented here, and new outlooks may also evolve over time, when radiologists are better acquainted with AI, and as AI develops and proves (or disproves) it is worth and influence in radiology. These perspectives are also not completely distinct or mutually exclusive. Rather, they are a prism through which the different types of institutionalized reactions to AI can be observed (Oliver, 2004).

Results

Results reveal a complex web of perspectives of radiologists regarding AI. This complexity suggests that professional maintenance in the field of radiology entails a constant maneuvering between interrelated perspectives, even in a single article, by a single author. This rhetorical dance of argumentation suggests that radiologists’ maintenance work is not necessarily strategic or intended, as no author wishes to appear inconclusive. Rather, the complexity of arguments suggests that maintenance work is done through an ongoing discussion, that enhances professional boundaries. As in prior studies of maintenance work in professional fields, this complexity is typical (Brayne and Christin, 2021; Goto, 2022; Malsch and Gendron, 2013; Micelotta and Washington, 2013) and is aimed at preserving traditional power structures while allowing the field to enjoy the prestige and benefits of innovation and technology.

It is possible to crudely divide the approaches of radiologists into “positive” (in regards to AI) versus “negative” (in regards to AI) (Kim et al., 2021); “cooperate” versus “compete” (Anteby et al., 2016) or “connect” (with AI) versus “protect” (from AI) (Noordegraaf, 2020), as has been suggested in recent discussions in the sociology of professions. However, we offer a more nuanced understanding of radiologists’ positions regarding AI and show that in their institutional work, radiologists not only simultaneously perform boundary work and networking with the AI field, but a dance-like maneuvering between relational perspectives. Thus, the results below show that interprofessional relations, as all social relations, are a prismatic dance, never merely a movement of protection/competition versus connection/cooperation (Alvehus et al. 2021).

Below are the four themes that emerged in our analysis of the data: (a) integrate—radiologists aim to work together with AI experts and technologies, viewing both specialties as essential and equal; (b) absorb—radiologists work with AI experts and technologies but view them as subordinate and as (yet another) limited technological tool; (c) fight-off and delegitimize—radiologists attack the expertise, efficiency and legitimacy of AI in radiology and (d) assimilate—radiologists view their profession as assimilating into the AI field. Table 1 shows some examples of the categories that emerged in our analysis, and the different perspectives for each category. This spectrum of perspectives for each category demonstrates how complex the relations are at the intersection of radiology and AI. For example, in the category of standardization, radiologists view AI as lacking standards in development and clinical application and at the same time as enabling better standardization of radiological expertise, or even offering standards where there are none. The table visually exemplifies the rhetorical dance that radiologists perform in their scientific discussion of AI.

Table 1.

Examples of the different perspectives of radiologists regarding AI.

	Integrate	Absorb	Fight-off and delegitimize	Assimilate
Structural relations	AI is exciting; Radiologists should work together with other disciplines on the development of AI.	AI is narrow, each model performs only one task, and therefore AI is a tool for radiologists.	AI is a very limited tool; it misinterprets images, has not been clinically trialed and is incapable of dealing with clinical complexity.	AI outperforms radiologists and should supervise radiologists.
Standards	AI allows for standardization, objectivity, and quantification of radiology.	AI can elevate radiologists’ status with its quantification techniques.	AI research is not standardized, regulated, is biased, over fits its training data, and is generally statistically inadequate.	With AI, radiology will stop being a qualitative, subjective profession.
Efficiency	AI will enable efficiency, cost reduction and workflow optimization.	AI models should be used as a second opinion.	AI will not be cost effective and efficient for hospitals. Cost of computing is high and there are special organizational hazards such as adversarial algorithms.	Resistance really is futile. AI has the support of powerful social forces such as technological companies.
Patients	Service to patients will improve with the integration of AI into radiology.		Patients would object to their images being interpreted by AI.	Patients should be considered as consumers.
Liability		Algorithms can identify cases radiologists miss and help avoid malpractice claims.	Radiologists are the legally responsible agent.	Human not in the loop is the next logical step.
Ethics		Radiologists should supervise the ethical implementation of AI.	The ethics of AI development is questionable: Illegal agreements, privacy breaches.
Skills and expertise	There should be an integration of radiology and data science skill sets.	Radiologists should learn AI in order to supervise it.	Radiologists should focus on their training and ignore AI as yet another limited CAD system. The AI expertise is heavily depended on radiologists’ skills, specifically for training algorithms.	Radiology practice will change dramatically.
Explainability	Radiologists can help with some black box problems in AI.	Radiologists can pour meaning into obscured AI models.	AI is a black box, parameters and features are meshed up.
Institutions	Joint institutions should be established such as associations, databases and hospital centers.	Radiology associations should supervise AI development in radiology.		Open competitions are the only way to advance radiological knowledge.

The integrate and fight-off perspectives were the most frequent, with 54 quotes in each. We found 37 arguments for absorb and only 26 which argue to assimilate. Each argument entails explanations by its author(s) on how and why radiology should integrate with, absorb, fight-off or assimilate into the AI expertise. These varied justifications exemplify the sociological complexity of embedding AI in radiology, the distinctiveness of these two professional cultures, and the attraction/ rejection dance radiologists perform with the AI technology. In the following lines, we discuss each perspective and the interactions between them.

Integrate

Integration means the continued cooperation of two professional groups while each group preserves its core values, training, skills and interpretations of tasks and problems. In the case of AI in radiology, rather than closing-off the boundaries of the profession, some radiologists seem to embrace the new technology and expertise viewing them as exciting and beneficial in terms of knowledge production, standards, efficiency and service to patients. In this framework, AI in radiology is perceived as reducing workload (Wong et al., 2019), standardizing and promoting quantification in radiology (Cochon et al., 2019; Hricak, 2018). This opening of boundaries in the profession is done specifically by adjusting the knowledge base of radiology to include and assist the AI expertise. We will discuss here three aspects of this integration: (a) learning the language and methods of AI; (b) integrating medical knowledge into AI systems and (c) establishing new shared institutions for knowledge development.

Learning the language, methods and skills. In their academic discussions, radiologists make an effort to learn the language of AI, read into it, as their new “dance partner” (Ortmann and Sydow, 2018). They teach each other about AI methods. For example, Erickson et al. (2017), Moore et al. (2019) and Nichols et al. (2019) review definitions, terminology, approaches and methods in AI, in what seems as a genuine effort of the radiology community to learn about AI. Even critical voices (e.g. Pesapane et al., 2018; Al’Aref et al., 2019) first present in their articles the definitions, terminologies, approaches and methods. This effort to learn points to the willingness of radiologists to integrate AI into their profession. However, this integration does not include an internalization of the AI's community ethos of machine superiority over human experts; radiologists hold their fortress of experts in imaging.

Integrating medical knowledge into AI systems. A second form of integration effort is done by finding ways radiologists can contribute to the development of AI in radiology. Ideas about cooperation and integration are many. Radiologists can help AI become less opaque (Akselrod-Ballin et al., 2019), annotate images for the training of AI (Chan and Siegel, 2019), and help create databases for AI (Tsai et al., 2021). In cancer screening, radiologists offer to contribute their knowledge to the systems:

Other groups have reported the inclusion of qualitative semantic features such as nodule location, cavitation, and calcification… Hence, additional work is required to integrate these radiologist-crafted features to analyze their importance in our cohort… clinical translation as a cancer screening tool will require careful planning to integrate the human and machine interpretations together in decision support mode. (Beig et al., 2019: 791–792)

These voices promote a welcoming approach within the radiology profession to the promises of the new technology, offer the AI expertise assistance in “reading” into the radiological expertise (Ortmann and Sydow, 2018), and encourage participation in developing the technology. This integration, again, points to the importance of the human-radiologists expert in developing AI for radiology.

Building shared institutions. A third form of integration is promoted by radiologists when new institutions are established in both the professional and the organizational levels. Shared professional associations for radiology and computational methods are being founded, while hospitals establish AI centers. One example for this institutional integration is the efforts of the European Society of Medical Imaging (EuSoMII), formally established in 2016, to “connect radiologists, radiology residents, data scientists and informatics experts” (EuSoMII website).

Institutional cooperation also includes the establishment of “data centers” in hospitals (Dreyer and Geis, 2017: 715). These new institutions reflect a field-level effort to establish hybrid organizational forms, where a space is created for new tasks and power structures to emerge. When radiologists willfully establish shared institutions with data scientists and other IT workers, and cooperate with them for the development of AI, they seem to create a new field of medical-radiological data science, radiomics, with its related identities and careers. However, the establishment of such organizational forms does not eliminate the tension between the two work cultures, which use different criteria to establish “ground truth” (Lebovitz et al. 2021, 2022).

Finally, efforts to integrate radiology and AI knowledge are done in shared innovative competitions, where radiologists create annotated databases and release them to the online AI community:

RSNA and Society of Thoracic Radiology (STR) collaborated to develop the RSNA International COVID-19 Open Radiology Database (RICORD). This database is the first multi-institutional, multi-national expert annotated COVID-19 imaging dataset. It is made freely available to the machine learning community as a research and educational resource for COVID-19 chest imaging. (Tsai et al., 2021: 204)

This acceptance of competitions as a way to create new scientific knowledge and the close cooperation of radiologists and the AI community indicate on the one hand a successful integration. On the other hand, as will be discussed below, open competitions represent an adoption by radiologists of the work culture of the AI community, allowing non-physicians to work on the diagnosis of imaging data.

In sum, integration efforts with the AI community in the form of learning, knowledge exchange and shared institutions promote an opening of professional boundaries. At the same time, this integration is done with consistent distinction of the two scientific cultures.

Absorb

The absorb perspective means that radiologists view AI as a technology and expertise that will be incorporated into their profession, but not in a disruptive or revolutionary way. Rather, some radiologists view the incorporation of AI into radiology as a slow, gradual process, similar to the incorporation of previous technologies, and as subordinated to radiologists’ authority. Here, radiologists preserve their authoritative position by viewing AI as a narrow and limited tool. In this perspective, thus, AI is understood as a technology that should be supervised by radiologists. This subordination is done as follows:

Framing AI as narrow. First, radiologists point to AI's known Achilles’ heel, namely the absence of “general AI” and the fact that thus far computer scientists are only able to build “narrow AI.” In this vein, radiologists frame AI as performing partial and fragmentary tasks and criticize these limitations.

The majority of (if not all) deep learning techniques developed to date are unitaskers. They address a single type of image or modality and a single disease entity. There are likely situations where this is perfectly appropriate and useful, but in the long run we cannot have a plethora of different independent schemes providing a multitude of “opinions” that the radiologist needs to somehow sift through to make sense of. (Krupinski, 2018: 854)

Krupinski reminds us that even when AI outperforms a radiologist, it does so in a very specific task. This means, that to incorporate AI into radiology, may result in a plethora of different tools, each dedicated to a different radiological task, and resulting in inefficiency. Others argue that AI offers narrow solutions to narrow problems, and is mostly successful in games, and not in complex realities:

Board games such as “Go” focus on a very “narrow” artificial intelligence task where a winning vs losing status can be assessed, whereas medical imaging is associated with far greater amount of ambiguity, and a larger variety of features, classifications, and outputs. It is also likely that thousands of “narrow” algorithms based on separate large, well-annotated databases will be required for a computer to begin to compete with a radiologist for comprehensive diagnostic assessment of even a single modality covering a single anatomical region of the body… Although possible in games with simple defined rules such as Go or Chess, analogous self-reinforcement learning is not so easily attainable in radiology, given the lack of a simple set of rules of the “radiology game” to allow this sort of self-play. (Chan and Siegel, 2019; 1–2)

Framing AI as good for games, and then ironically referring to the “radiology game” rhetorically contrasts the frivolous and immature tech culture with the serious and responsible medical profession. In this perspective, radiologists are open to integrating AI tools in their work, but view them as inferior, narrow, game-like tools.

Subordinating AI as a limited technology. When viewed as a tool, rather than an alternative expertise, radiologists welcome AI into their profession, mainly to assist in the simple cases. They aim to incorporate AI into their profession as yet another computer aided detection (CAD) limited tool (Chan and Siegel, 2019), and do not view it as an exciting, superior technology or a threat. Instead, radiologists subordinate AI by criticizing its performance. As this senior radiologist comments:

Current data indicate that in general even the best machine learning systems do not yet perform at the level of a radiologist… The use of radiomics and machine learning to predict patient outcomes is still in its infancy. (Summers, 2019: 1987–1989)

This author frames AI as being in its “infancy” and points to the superiority of the radiological expertise over the technology. In this perspective, radiologists do not close-off the boundaries of their profession to the technology. However, they frame it is as being under-developed. This way, they can preserve their authoritative position, while gaining from the aura of innovation (Burri, 2008; Malsch and Gendron, 2013).

Similarly, from an authoritative position, radiologists aim to learn AI in order to know its limitations and supervise it:

As ML algorithms become more embedded in clinical practice, radiologists will need to expand their understanding of these methods and what functions the models were created to perform. More importantly, it will be imperative to understand the limitations of these tools and models in the patient care continuum. (Halabi et al., 2019: 503)

Here, radiologists frame their role as the ones that “know better” than the systems, and are able to supervise them. In other words, radiologists holding this perspective reject one of the core values of the AI community, that of “computer knows best” (McDougall, 2019).

In addition to arguments about performance and expertise, radiologists preserve their authoritative position also in terms of ethical considerations. They argue that since they (and other physicians) are the responsible and liable agents, they must supervise AI:

As machine learning enters state-of-the-art clinical practice, medicine thus has the immense obligation to ensure that this technology is harnessed for societal and individual good, fulfilling the ethical basis of the profession. (Darcy et al., 2016: 552)

Thus, absorbing AI into radiology and subordinating it to radiologists’ authority is done by a rhetorical dance of opening the boundaries of the profession to the technology, yet not as an equal, disruptive expertise to integrate with, but rather as a narrow and limited tool, which requires supervision.

Fight-off and delegitimize

While they allow AI into their profession, radiologists also directly challenge AI and “disrupt disrupters” (Hargrave and Van de Ven, 2009: 120). They protect the boundaries of their expertise, separate their core knowledge and skills from those of AI experts and undermine the legitimacy of AI in radiology. We will discuss here five major arguments radiologists make against the incorporation of AI in radiology:

Standardization. First, radiologists attack the lack of standardization of AI in radiology. For example:

Because of the rapid growth of this area, numerous published radiomics investigations lack standardized evaluation of both the scientific integrity and the clinical relevance. (Pesapane et al., 2018: 5)

Radiologists are concerned regarding the evaluation of AI in academic publishing, and even more, regarding the lack of standardization through clinical trials:

While sensitivity, specificity and ROC comparisons between algorithms and clinicians on test data sets certainly add validity to algorithm performance, the gold standard for any new methodology applied in a clinical setting relies on comparable or superior performance in a randomized clinical trial. (Nichols et al., 2019)

The authors require that AI will standardize according to the clinical trial practice in medicine. The clinical trial gold standard is in striking opposition with the above-mentioned open competition standards of the AI community: while randomized clinical trials require long-term evaluation of sensitivity and specificity of tools on a well-defined, randomly assigned cohort of patients, open competitions use different accuracy measurements, and mainly compare algorithms to physicians on given labeled datasets, returning to the “ground truth” problem of labeling (Lebovitz et al., 2021), and without clinical long term follow up on patients. Thus, while participating in competitions, and even providing labeling for algorithms, radiologists cast a shadow on the validity of these endeavors in clinical settings.

Lack of explainability and a “black-box” technology. A second argument against AI is that it is a “black-box” technology, which does not produce explanations for its decision-making process:

One of the biggest issues in deep learning is the so-called black box problem (i.e. in deep learning the algorithm finds the rules itself, but often without leaving an audit trail to explain its decisions)… Especially in medicine, where accountability is important and can have serious legal consequences, it is often insufficient simply to have a good prediction system. This system should be able to articulate itself in a certain way tracing a report to explain its decisions. (Sollini et al., 2018: 7)

Additionally, an AI model may diagnose a statistical-probabilistic chance of cancer based on an image and the patient's medical record, without pointing to the tumor location in the image (Akselrod-Ballin et al., 2019); or, an AI model may detect small aneurysms, not detected by expert radiologists, but miss the big and obvious ones (Flanders, 2019: 196).

This type of erratic and opaque behavior of AI, while viewed as a shortcoming in radiology and management literature (Lebovitz et al., 2022) and as a reason for radiologists not to clinically trust AI, is glorified in the AI literature (Breiman, 2001) and is seen as a positive, trans-human feature of this technology, not as a bug. This difference in the two professional-scientific cultures, although allegedly mitigated by attempts to produce “explainable AI,” is actually persistent, and represents a deep scientific divide between the creators of AI and the rest of the scientific community (Cox, 2001; Lange et al., 2019; Vesa and Tienari, 2020). Thus, the powerful capabilities of AI to perform expert labor, such as radiological diagnosis, are constantly challenged by its inherent opacity.

Ethics. Third, radiologists doubt some of the ethical standards guiding the implementation of AI in their profession. For example, the following excerpt presents doubts regarding the legal standing of database generation in one case involving DeepMind, a famous AI company owned by Google:

Developing AI requires access to large volumes of data. In the UK, this has been challenging particularly within the NHS. The panel found the agreement between DeepMind and the Royal Free Hospital Trust to be illegal and contained “deficiencies” as it did not safeguard patient data. (Wong et al., 2019: 142)

These concerns regarding medical databases involve considerations of patients’ privacy and autonomy, as well as bias and health disparities. The following excerpt is concerned with the use of medical information without patients’ informed consent:

There is also a major ethical concern with the development and use of augmented intelligence systems. How should consent be obtained from patients for the use of their imaging data? Should permission be obtained from the patients at the time of imaging that their imaging data may be used to train an algorithm? Further, how should data privacy be addressed? Where should this data be stored, when data hacking is an ongoing problem for some of the most secure systems worldwide? (Jalal et al., 2019: 11)

Informed consent is the basis for medical ethics since the Nirenberg trials after World War II; forsaking it in big data research is a concern among medical practitioners and ethicists (Rothstein, 2015). Other authors are worried about biased databases that will lead to biased AI:

Machine learning, despite being data-driven, can often be riddled with traditional biases. While harmless in most commercial settings, these can become problematic in the high-stakes healthcare space, where precision is of the utmost importance. (Al’Aref et al., 2019: 1985)

The authors, well acquainted with medical research, raise the concern of unmitigated bias in AI. They point to sampling, observer and database bias, and to the alleged absence of procedures to mitigate these biases in AI, as opposed to medical clinical trials.

These ethical considerations are part of a wide concern in healthcare regarding the implementation of AI systems into work practices (Siala and Wang, 2022). Specifically, they point to a growing problem in the ligations between the two professional cultures, medical and big data research, namely how to distinguish legitimate research practices from professional misconduct. Therefore, efforts to integrate and open the boundaries of the radiology profession to AI tools are done with a persistent delegitimizing of their ethical compatibility with medical standards.

Statistical inadequacy of ML-based AI. When fighting-off AI, radiologists point specifically to the statistical inadequacy of many AI implementations. As one review article summarizes:

Feature type, selection, and classifiers vary among studies, patient sample sizes tend to be relatively small, test/validation datasets are lacking in most cases, and the image types used to extract features are variable. Moreover, it is unclear which indices should be used, what they represent, and how they are related to the underlying biological mechanism. (Sollini et al., 2018: 7)

Another article doubts the feasibility of relying on automated tools to perform diagnosis, considering the particularized nature of medicine:

Given any task, it is then merely a matter of experimentation to determine the optimal model with the greatest potential for generalizability. Theoretically, the lowest attainable error is known as the Bayes error rate. This ceiling on performance exists because most phenomena studied in the natural world is permeated by noise. Consider, for example, two patients characterized by identical clinical parameters; while it may be reasonable to assume that such individuals will, on average, experience similar clinical outcomes, it is impossible to make such a claim with absolute certainty because the system being approximated is probabilistic rather than deterministic. (Al’Aref et al., 2019: 1978)

Specifically, another known Achilles heel of AI technologies, namely the problem of overfitting a model to the training data, is mentioned by radiologists as a statistical limitation of this technology (Pesapane et al., 2018; Moore et al., 2019; Kahn, 2017). One review article even argues that the prevailing over-fitting phenomena in AI prevents it from giving reproducible results:

Our results demonstrate performance inconsistency across the data sets and models, indicating that the high performance of deep learning models on one data set cannot be readily transferred to unseen external data sets. (Wang et al., 2020: 796)

Again, while participating in the development of AI in the integrate perspective, radiologists doubt the statistical validity of AI and point to the disconnect of this technology from biological mechanisms.

Cost-effectiveness and efficiency of AI. Fifth, radiologists question the cost-effectiveness of AI in radiology. Considering the attempts to de-professionalize medicine in contemporary hospitals, and transforming it into a corporate profession (Ritzer and Walczak, 1988), this argument carries special importance; it means that physicians “go after” AI on grounds of cost-effectiveness, and not of expertise. For example, radiologists preach administrative wariness of more computing costs:

… more complex tasks require more computing power. Uncertainty remains surrounding the degree of processing required to run these advanced programs and hospitals may not be prepared for additional network requirements. (Wong et al., 2019: 142–143)

Anchoring the debate in past experience with technology, some radiologists argue that the incorporation of AI in radiology may harm healthcare systems:

Physician skepticism of health care informatics is justifiable given the poor design of so many electronic hospital records systems, the additional clerical burden assumed by physicians, the barrier created between physicians and patients, and the contribution of EHRs to physician dissatisfaction and burnout. (Moore et al., 2019: 515)

These challenges to the ethics, expertise and effectiveness of AI point to the rhetorical dance and elaborate maintenance work of radiologists, based on distinct values of the two scientific cultures. Reliable standardization versu chaos; explainability versus opacity; ethical and responsible public service versus market-driven, capitalist approach; trustworthiness versus overfitting and modest saving versus technological extravagance are at the core of these distinction efforts, which are done while establishing shared associations, hospital data centers or participating in competitions and providing labels for AI.

Assimilate

While in the larger part of their discussion, radiologists view their expertise as superior, powerful and enduring, and AI as an exciting partner or a limited tool, on the discourse margins, automation of radiology is considered a prospective option, a future that should be discussed. Assimilation means that radiologists are anticipating a profound change in their profession and expertise—in the way knowledge and practices are produced and the professional identity of those who are “allowed” to practice image diagnosis. Although none of the radiologists viewed their expertise as redundant at present time, when they discuss automation as a possibility, they acknowledge an upcoming change in the radiological expertise in the face of AI's disruption. For example, some argue that the human expert may soon be replaced by an AI system: “The idea to dismiss mammograms that are categorized as very likely normal without any human reader interpretation is the logical next step” (Geras et al., 2019: 251). Affiliated with both radiology and data science centers, the authors embrace full automation as a viable option in radiology, yet pointing to AI's limitation at the end of their article and reserving clinical evaluation for physicians. These rhetorical maneuvers suggest a centrifugal force operating in the radiology profession (Montgomery and Oliver, 2007), initiating outward-directed networking with the AI community's set of practices, norms and believes of machine superiority, due to an understanding of the role of AI and its supporters in twenty-first century expertise. However, as Montgomery and Oliver (2007) show, these efforts of assimilation may actually reinforce domain claims, as the profession stretches its boundaries to include new and innovative approaches, techniques and technologies.

This is a challenging maneuver to make. Faced with the capabilities of AI, and the wealth, vigor and youth of the institutional forces driving it, some radiologists view “resistance as futile”:

Machine learning technologies are now deeply embedded in our medical information systems. These methods will ultimately be pervasive in the digital realm of radiology. Resistance really is futile. Despite substantial impediments, there are also tremendous forces in its favor, including powerful and wealthy IT companies such as IBM and Alphabet, along with the great equalizing force of data clouds, high speed networks, and most of all, brilliant young minds born into the digital age. (Choyke, 2018: 139)

Here, the radiologist argues for AI, not on the basis of expertise. Rather, he makes a political evaluation of the forces struggling over the imaging jurisdiction, especially big IT companies and young enthusiastic researchers in the field of AI. In contrast to authors in the above-mentioned fight-off perspective, here the author believes these forces and their methods of “doing things” will be “pervasive in the digital realm of radiology,” changing radiological expertise, and that resistance to these forces “is futile.” In this vein, other radiologists call for opening up the way medical research is conducted and medical knowledge created:

ML and related data science initiatives for medical imaging will succeed with greater access to accurately curated and publicly available data sets. The broad availability of the data set allows individuals with different backgrounds to explore nontraditional solutions, accelerating discoveries at a pace that is more rapid than that of the traditional scientific method. (Halabi et al., 2019: 501)

Halabi et al. (2019) reported an imaging diagnosis competition organized by the Radiological Society of North America (RSNA). While they “read” the way knowledge production is organized in the AI expertise, they call for a change in the way scientific research in radiology is organized; instead of an individual researcher working alone on a database, they open radiological research to non-radiologists and non-experts following the machine learning community's preference for open competitions on open datasets. While these efforts have been mentioned above as integrative, they might carry a profound change in the professional identity, knowledge and skill, and daily practice of the persons performing research in medical imaging, and ultimately, change the way medical scans are diagnosed in clinical settings.

Discussion

The reaction of radiologists to AI's disruption reflects the complexity evident in other expert fields in reaction to AI (Brayne and Christin, 2021; Goto, 2022) and the well-documented complexity of change in professional fields (Malsch and Gendron, 2013; Murray, 2010). Our motivation was to explore the reaction of a powerful and territorial medical profession to the potential major disruption posed by AI expertise. Our results, not surprisingly, depict a range of reactions, all directed toward maintaining medical dominance, even in the face of an equally powerful expertise, such as AI. Our contribution to the literature of institutional maintenance in professional fields lies in the concept of rhetorical dance. Radiologists are able to contain AI's penetration into their profession through maneuvering between different, cooccurring perspectives, each with its own internal logic, but all are a part of larger dance of maintenance work.

In the first perspective, we found, radiologists argue for integration of their profession with the AI expertise—they learn “to read” AI methods and terminology, integrate their knowledge into AI systems, and establish shared institutions with the AI community in what seems as a productive cooperation with an exciting and transformative new technology and expertise.

At the same time, radiologists make certain that AI technology is not seen as automated procedures that may be considered a replacement for their expert labor or discretion for complicated cases, as other experts in other fields do (Goto, 2022; Christin, 2017). They demonstrate their professional power in the second perspective, absorb, when they frame AI as yet another technology, a mere tool for the expert radiologist, and not as disruptive or transformative. Advancing the technology while subordinating it, supports the status of radiologists as technology-orientated physicians, masters of both medicine and technology.

Concurrently, radiologists also engage in a typical professional struggle over the imaging jurisdiction, with their fight-off and de-legitimize perspective. Framing AI as inaccurate, unstandardized and unreliable, opaque, unethical and not cost effective, the radiologists ensure that they maintain their legitimate dominant position in the imaging jurisdiction. They clarify that no other party is able to judge AI's performance and be liable for its decisions and that public opinion, state regulation and hospital management consider radiologists’ discretion as more accurate, standardized, efficient, ethical and responsible than that of AI experts.

Out of the four rhetorical perspectives we describe in our study, the fourth perspective, assimilate, is of special interest in the context of professional maintenance as it depicts an alleged lost battel between radiology as an expert profession facing AI. Here, radiologists open the boundaries of their profession to the AI community's logics of algorithmic abstractions and trust in machine judgment. This way, they can affiliate their specialty with wealthy IT companies and young, enthusiastic AI experts. Radiologists do so because they are aware of the benefits of joining these heterodox forces (Malsch and Gendron, 2013). In assimilating into the logics of the AI culture, radiology gains the image of an innovative profession, inside and outside the medical field, a coproducing partner in the development of advanced technology.

These cooccurring four perspectives allow the radiology profession to undergo a change in its knowledge base and expertise while aiming to maintain its current power structure (Micelotta and Washington, 2013; Montgomery and Oliver, 2007). Following Farjoun (2010), we see these four perspectives of radiologists as interdependent expressions, a rhetorical dance for professional maintenance work. Unlike other concepts related to institutional maintenance, we depict a rhetorical dance as a spectrum of relational perspectives that emerges when looking closely, through an interpretive prism, on the scientific discourse of the profession (Oliver, 2004). Rhetorical dance differs from previous concepts of institutional maintenance such as experimentation (Malsch and Gendron, 2013) that views institutional work as a fragile process, subject to trials and tests in experiments, while trying to extend the professional jurisdiction and apply institutional reproduction to enhance legitimacy. It differs also from the notion of conflictual hybrids (Murray, 2010), where hybrids are emerging from conflict. Rhetorical dance differs as well from the conventional notion of boundary work where only protection is used to maintain the resilience of professional logics and from the concept of micro-level ambiguity in managerial logics, where ambiguity is performed on the ground, but not in formalized in scientific publishing (Currie and Spyridonidis, 2016). Finally, rhetorical dance differs from boundary-spanning activities and networking (Montgomery and Oliver, 2007), where professional boundary maintenance work is based on dual and intermittent centripetal and centrifugal forces. In our paper, as others who use the dance metaphor (Ortmann and Sydow, 2018; Wallenburg et al., 2021), we show that radiologists are able to spin a range of cooccurring arguments, of both attraction to AI, rejection of it, and the moves in between, allowing them to decipher the boundary crossing of external experts that offer alternative expertise, and contain the uncertainty and disruption caused by AI.

The introduction of AI into professions has been described as a vision of innovative automation of expertise. However, a close reflection on experts’ reaction to AI reveals the active and agentic role played by professionals across expert fields, their evaluations of technology and their maintenance efforts, as an important force in processes of innovation (Christin, 2017; Faulconbridge et al., 2021; Goto, 2022; Stevens et al., 2022). We hypothesize that this “one-step integrate, two-steps fight-off, spin and subordinate, turn and assimilate” dance that radiologists perform with AI will persist, and more so, that it is not unique to radiology. While AI experts are aiming to enter powerful professions, and specifically medicine, the battle to maintain the integrity of existing fields will continue.

Future research in the study of AI in expert labor should compare different expert fields’ reactions to AI's disruption and offer a systematic review of resistance strategies. In radiology, research should study the profession's perspectives over time, to examine whether radiologists continue to adhere to mixtures of perspectives or congregate to one.

One limitation of this study is the lack of distinction between radiologists’ perspectives regarding the AI technology itself and the AI experts producing it. This lack of distinction stems from the absence of such distinction in the data; the radiologists do not make this distinction. However, analytically, there is a difference between the technological tools and the experts developing them. This difference should receive research attention. Another limitation is that we present a sample of reactions that appear in the academic journals of radiology. Obviously, additional arguments may be gained through interviews or observations of radiologists conducting their work. Such additional data sources should be applied in future research.

Footnotes

Acknowledgements

The first author would like to thank the Hebrew University of Jerusalem for hosting her as a postdoctoral fellow at the time of the research.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iDs

Netta Avnoon

Amalya L Oliver

Notes

References

Abbott

(1988) The System of Professions: An Essay on the Division of Expert Labor. Chicago: University of Chicago press.

Adler

Kwon

(2013) The mutation of professionalism as a contested diffusion process: Clinical guidelines as carriers of institutional change in medicine. Journal of Management Studies 50(5): 930–962.

Akselrod-Ballin

Chorev

Shoshan

, et al. (2019) Predicting breast cancer by applying deep learning to linked health records and mammograms. Radiology 292(2): 331–342.

Al’Aref

Anchouche

Singh

, et al. (2019) Clinical applications of machine learning in cardiovascular disease and its relevance to cardiac imaging. European Heart Journal 40(24): 1975–1986.

Alvehus

Avnoon

Oliver

(2021) It’s complicated’: Professional opacity, duality, and ambiguity—A response to Noordegraaf (2020). Journal of Professions and Organization 8(2): 200–213.

Anteby

Chan

DiBenigno

(2016) Three lenses on occupations and professions in organizations: Becoming, doing, and relating. Academy of Management Annals 10(1): 183–244.

Bader

Kaiser

(2019) Algorithmic decision-making? The user interface and its role for human involvement in decisions supported by artificial intelligence. Organization 26(5): 655–672.

Barley

(1986) Technology as an occasion for structuring: Evidence from observations of CT scanners and the social order of radiology departments. Administrative Science Quarterly 31: 78–108.

Barley

(1995) Images of imaging. Longitudinal Field Research Methods: Studying Processes of Organizational Change 1(3): 231–337.

10.

Beig

Khorrami

Alilou

, et al. (2019) Perinodular and intramodular radiomic features on lung CT images distinguish adenocarcinomas from granulomas. Radiology 290(3): 783–792.

11.

Brayne

(2017) Big data surveillance: The case of policing. American Sociological Review 82(5): 977–1008.

12.

Brayne

Christin

(2021) Technologies of crime prediction: The reception of algorithms in policing and criminal courts. Social Problems 68(3): 608–624.

13.

Breiman

(2001) Statistical modeling: The two cultures (with comments and a rejoinder by the author). Statistical Science 16(3): 199–231.

14.

Browder

Koch

Long

, et al. (2022) Learning to innovate with big data analytics in inter-organizational relationships. Academy of Management Discoveries 8(1): 139–166.

15.

Burri

(2008) Doing distinctions: Boundary work and symbolic capital in radiology. Social Studies of Science 38(1): 35–62.

16.

Chan

Siegel

(2019) Will machine learning end the viability of radiology as a thriving medical specialty? The British Journal of Radiology 92(1094): 20180416.

17.

Chockley

Emanuel

(2016) The end of radiology? Three threats to the future practice of radiology. Journal of the American College of Radiology 13(12): 1415–1420.

18.

Choyke

(2018) Quantitative MRI or machine learning for prostate MRI: Which should you use? Radiology 289(1): 138–139.

19.

Christin

(2017) Algorithms in practice: Comparing web journalism and criminal justice. Big Data & Society 4(2): 1–14.

20.

Cochon

Kapoor

Carrodeguas

, et al. (2019) Variation in follow-up imaging recommendations in radiology reports: Patient, modality, and radiologist predictors. Radiology 291(3): 700–707.

21.

Cox

(2001) Comment to statistical modeling: The two cultures by L I. Statistical Science 16(3): 216–218.

22.

Creswell

Miller

(2000) Determining validity in qualitative inquiry. Theory into Practice 39(3): 124–130.

23.

Currie

Lockett

Finn

, et al. (2012) Institutional work to maintain professional power: Recreating the model of medical professionalism. Organization Studies 33(7): 937–962.

24.

Currie

Spyridonidis

(2016) Interpretation of multiple institutional logics on the ground: Actors’ position, their agency and situational constraints in professionalized contexts. Organization Studies 37(1): 77–97.

25.

Dacin

Munir

Tracey

(2010) Formal dining at Cambridge colleges: Linking ritual performance and institutional maintenance. Academy of Management Journal 53(6): 1393–1418.

26.

Darcy

Louie

Roberts

(2016) Machine learning and the profession of medicine. JAMA 315(6): 551–552.

27.

DiMaggio

(1988) Interest and agency in institutional theory. In: Zucker

(ed) Research on Institutional Patterns: Environment and Culture. Cambridge: Ballinger Publishing Co, 3–21.

28.

DiMaggio

(1991) Constructing an organizational field as a professional project: US art museums, 1920–1940. In: Powell

DiMaggio

(eds) The New Institutionalism in Organizational Analysis. Chicago: University of Chicago Press, 267–292.

29.

Dreyer

Geis

(2017) When machines think: Radiology’s next frontier. Radiology 285(3): 713–718.

30.

Erickson

Korfiatis

Akkus

, et al. (2017) Machine learning for medical imaging. Radiographics 37(2): 505–515.

31.

Farjoun

(2010) Beyond dualism: Stability and change as a duality. Academy of Management Review 35(2): 202–225.

32.

Faulconbridge

Folke Henriksen

Seabrooke

(2021) How professional actions connect and protect. Journal of Professions and Organization 8(2): 214–227.

33.

Flanders

(2019) Machine learning detection of intracranial aneurysms—will it play in Peoria? Radiology 290(1): 195–197.

34.

Francis

(2008) Finding and maintaining professionalism in radiology. Clinical Radiology 63(1): 15–17.

35.

Geras

Mann

Moy

(2019) Artificial intelligence for mammography and digital breast tomosynthesis: Current concepts and future perspectives. Radiology 293(2): 246–259.

36.

Gioia

Corley

Hamilton

(2013) Seeking qualitative rigor in inductive research: Notes on the Gioia methodology. Organizational Research Methods 16(1): 15–31.

37.

Goto

(2022) Accepting the future as ever-changing: Professionals’ sensemaking about artificial intelligence. Journal of Professions and Organization 9(1): 77–99.

38.

Halabi

Prevedello

Kalpathy-Cramer

, et al. (2019). The RSNA pediatric bone age machine learning challenge. Radiology 290(2): 498–503.

39.

Hansen

(2021) Model talk: Calculative cultures in quantitative finance. Science, Technology, & Human Values 46(3): 600–627.

40.

Hargrave

Van de Ven

(2009) Institutional work as the creative embrace of contradiction. In: Lawrence

Suddaby

Leca

(eds) Institutional Work: Actors and Agency in Institutional Studies of Organizations. Cambridge: Cambridge University Press, 120–140.

41.

Hinton

(2016) On radiology. Presented at the Machine Learning and the Market of Intelligence, Creative Destruction Lab, Toronto. Available at https://www.youtube.com/watch?v=2HMPRXstSvQ.

42.

Hricak

(2018) 2016 New horizons lecture: Beyond imaging—radiology of tomorrow. Radiology 286(3): 764–775.

43.

Jalal

Nicolaou

Parker

(2019) Artificial intelligence, radiology, and the way forward. Canadian Association of Radiologists Journal 70(1): 10–12.

44.

Kahn

Jr (2017) From images to actions: Opportunities for artificial intelligence in radiology. Radiology 285(3): 719–720.

45.

Kellogg

Valentine

Christin

(2020) Algorithms at work: The new contested terrain of control. Academy of Management Annals 14(1): 366–410.

46.

Kim

Koopmanschap

Mehrizi

MHR

, et al. (2021) How does the radiology community discuss the benefits and limitations of artificial intelligence for their work? A systematic discourse analysis. European Journal of Radiology 136(2021): 109566.

47.

Krupinski

(2018) Deep learning of radiology reports for pulmonary embolus: Is a computer reading my report? Radiology 286(3): 853–855.

48.

Lange

Lenglet

Seyfert

(2019) On studying algorithms ethnographically: Making sense of objects of ignorance. Organization 26(4): 598–617.

49.

Lawrence

Suddaby

(2006) Institutions and institutional work. In: Clegg

Lawrence

Hardy

Walter

(eds) The Sage Handbook of Organization Studies. London: Sage, 215–254.

50.

Lebovitz

Levina

Lifshitz-Assaf

(2021) Is AI ground truth really “true”? The dangers of training and evaluating AI tools based on experts’ know-what. Management Information Systems Quarterly 45(3b): 1501–1525.

51.

Lebovitz

Lifshitz-Assaf

Levina

(2022) To engage or not to engage with AI for critical judgments: How professionals deal with opacity when using AI for medical diagnosis. Organization Science 33(1): 126–148.

52.

Leiva

Ríos

FJM

Martínez

(2006) Assessment of interjudge reliability in the open-ended questions coding process. Quality and Quantity 40(4): 519–537.

53.

MacKenzie

(2021) Trading at the Speed of Light: How Ultrafast Algorithms Are Transforming Financial Markets. Princeton: Princeton University Press.

54.

Malsch

Gendron

(2013) Re-theorizing change: Institutional experimentation and the struggle for domination in the field of public accounting. Journal of Management Studies 50(5): 870–899.

55.

McDougall

(2019) Computer knows best? The need for value-flexibility in medical AI. Journal of Medical Ethics 45(3): 156–160.

56.

Micelotta

Washington

(2013) Institutions and maintenance: The repair work of Italian professions. Organization Studies 34(8): 1137–1170.

57.

Möhlmann

Zalmanson

Henfridsson

, et al. (2021) Algorithmic management of online labor platforms: When matching meets control. Management Information Systems Quarterly 45(4): 1999–2022.

58.

Montgomery

Oliver

(2007) A fresh look at how professions take shape: Dual-directed networking dynamics and social boundaries. Organization Studies 28(5): 661–687.

59.

Moore

Slonimsky

Long

, et al. (2019) Machine learning concepts, concerns and opportunities for a pediatric radiologist. Pediatric Radiology 49(4): 509–516.

60.

Murray

(2010) The oncomouse that roared: Hybrid exchange strategies as a source of distinction at the boundary of overlapping institutions. American Journal of Sociology 116(2): 341–388.

61.

Nichols

Chan

HWH

Baker

(2019) Machine learning: Applications of artificial intelligence to imaging and diagnosis. Biophysical Reviews 11(1): 111–118.

62.

Noordegraaf

(2020) Protective or connective professionalism? How connected professionals can (still) act as autonomous and authoritative experts. Journal of Professions and Organization 7(2): 205–223.

63.

Oliver

(2004) On the duality of competition and collaboration: Network-based knowledge relations in the biotechnology industry. Scandinavian Journal of Management 20(1–2): 151–171.

64.

Ortmann

Sydow

(2018) Dancing in chains: Creative practices in/of organizations. Organization Studies 39(7): 899–921.

65.

Parry

(2018) The Rise of the Medical Profession: A Study of Collective Social Mobility. London and New York: Routledge.

66.

Pasquale

(2015) The Black Box Society: The Secret Algorithms that Control Money and Information. Harvard University Press.

67.

Pesapane

Codari

Sardanelli

(2018) Artificial intelligence in medical imaging: Threat or opportunity? Radiologists again at the forefront of innovation in medicine. European Radiology Experimental 2(1): 35–45.

68.

Polanyi

(1966) The Tacit Dimension. New York: Doubleday.

69.

Pope

Ziebland

Mays

(2000) Analysing qualitative data. British Medical Journal 320(7227): 114–116.

70.

Ribes

Hoffman

Slota

, et al. (2019) The logic of domains. Social Studies of Science 49(3): 281–309.

71.

Ritzer

Walczak

(1988) Rationalization and the deprofessionalization of physicians. Social Forces 67(1): 1–22.

72.

Rothstein

(2015) Ethical issues in big data health research. Journal of Law, Medicine & Ethics 43(2): 425–429.

73.

Rystedt

Ivarsson

Asplund

, et al. (2011) Rediscovering radiology: New technologies and remedial action at the worksite. Social Studies of Science 41(6): 867–891.

74.

Saunders

Sim

Kingstone

, et al. (2018) Saturation in qualitative research: Exploring its conceptualization and operationalization. Quality and Quantity 52(4): 1893–1907.

75.

Scott

Ruef

Mendel

, et al. (2000) Institutional Change and Healthcare Organizations: From Professional Dominance to Managed Care. Chicago and London: The University of Chicago Press.

76.

Siala

Wang

(2022) SHIFTing artificial intelligence to be responsible in healthcare: A systematic review. Social Science and Medicine 296(114782): 1–15.

77.

Sollini

Cozzi

Chiti

, et al. (2018) Texture analysis and machine learning to characterize suspected thyroid nodules and differentiated thyroid cancer: Where do we stand? European Journal of Radiology 99: 1–8.

78.

Stevens

Wehrens

De Bont

(2018) Conceptualizations of big data and their epistemological claims in healthcare: A discourse analysis. Big Data & Society 5(2): 2053951718816727.

79.

Stevens

Wehrens

Kostenzer

, et al. (2022) Why personal dreams matter: How professionals affectively engage with the promises surrounding data-driven healthcare in Europe. Big Data & Society 9(1), 20539517211070698.

80.

Strauss

Corbin

(eds) (1997) Grounded Theory in Practice. Thousand Oaks, CA: Sage.

81.

Suddaby

Viale

(2011) Professionals and field-level change: Institutional work and the professional project. Current Sociology 59(4): 423–442.

82.

Summers

(2019) Are we at a crossroads or a plateau? Radiomics and machine learning in abdominal oncology imaging. Abdominal Radiology 44(6): 1985–1989.

83.

Susskind

(2015) The Future of the Professions: How Technology Will Transform the Work of Human Experts. Oxford: Oxford University Press.

84.

Timmermans

(2010) The continued social transformation of the medical profession. Journal of Health and Social Behavior 51(1_suppl): S94–S106.

85.

Tsai

Simpson

Lungren

, et al. (2021) The RSNA international COVID-19 open annotated radiology database (RICORD). Radiology 299: 204–213.

86.

Vesa

Tienari

(2020) Artificial intelligence and rationalized unaccountability: Ideology of the elites? Organization 29(6): 1133–1145.

87.

Wallenburg

de Graaff

Bal

, et al. (2021) Dancing with a virus: Finding new rhythms of organizing and caring in Dutch hospitals. In: Waring

Denis

Pedersen

Tenbensel

(eds) Organising Care in a Time of Covid-19: Implications for Leadership, Governance and Policy. London: Palgrave Macmillan, 121–138.

88.

Wang

Liang

Zhang

, et al. (2020) Inconsistent performance of deep learning models on mammogram classification. Journal of the American College of Radiology 17(6): 796–803.

89.

Wong

Al-Hasani

Alam

, et al. (2019) Artificial intelligence in radiology: How will we be affected? European Radiology 29(1): 141–143.

90.

Wright

Zammuto

Liesch

(2017) Maintaining the values of a profession: Institutional work and moral emotions in the emergency department. Academy of Management Journal 60(1): 200–237.

91.

Levy

(2010) Offshoring professional services: Institutions and professional control. British Journal of Industrial Relations 48(4): 758–783.

92.

Zhu

Gilbert

Chetty

, et al. (2021) The 2021 landscape of FDA-approved artificial intelligence/machine learning-enabled medical devices: An analysis of the characteristics and intended use. International Journal of Medical Informatics 165(104828): 1–7.

93.

Zilber

(2009) Institutional maintenance as narrative acts. In: Lawrence

Suddaby

Leca

(eds) Institutional Work: Actors and Agency in Institutional Studies of Organizations. Cambridge: Cambridge University Press, 205–235.

Nothing new under the sun: Medical professional maintenance in the face of artificial intelligence's disruption

Abstract

Keywords

Introduction

The profession of radiology and its use of technology

Research design and methods

Results

Integrate

Absorb

Fight-off and delegitimize

Assimilate

Discussion

Footnotes

Acknowledgements

Declaration of conflicting interests

Funding

ORCID iDs

Notes

References