The participants developed 2 separate ideas of extramural health checks: in one, they envisioned an environmentally sustainable, electric bus touring across the country, in a similar setup as the current population screening for breast cancer prevention in the Netherlands. Such a bus enables screening for a wide variety of potential diseases, which potentially benefits from large-scale screening programs. When AI detects a high-risk scan, the scan is forwarded to a radiologist, who decides whether to follow up. Within this scenario, it is possible to focus on the scan, or to combine the screening with blood or any other test, thereby combining data from multiple streams into a single risk assessment.

The other idea was focused on an extramural imaging, diagnosis, and treatment facility, which can, for instance, be located in a shopping mall. Citizens can visit the facility for a self-initiated check or after being invited due to specific risk factors, such as age. AI is subsequently used for fast diagnosis and forwards data to a radiologist if follow-up is required. As a consequence, there is still value in human intervention and responsibility. This scenario requires data to be shared with hospitals, for example, through electronic health records. A step further is to fully automate the diagnostic workflow by using large models that use large retrospective datasets and individual risk factors to screen for patients at risk and subsequently immediately treat correctly identified patients.

Participants emphasized how important they consider trust and responsibility when thinking about AI and radiology. Questions raised by a social scientist and patient representative included, “Is it responsible to leave decisions to computers?” (g1-s2) and “Who should be responsible for this?” (g1-s1). The ownership of the process and of the data was questioned. According to participants, AI can have added value in this scenario, including alleviating burdens from hospital staff, improving patients’ quality of life with an earlier start of the treatment, and equal access to health care.

Moreover, participants suggested that additional value can be calculated in added quality-adjusted life years, which can be translated into societal gains. The primary benefit of these scenarios may be that early diagnosis reduces long-term health care costs by shifting care from acute to preventive services. For funders, for example, the government, of large-scale AI-supported imaging-based screening programs, this shift may represent potentially favorable cost-benefit ratios where upfront investing may yield substantial savings and productivity over time. Additionally, commercial gain is possible, for example, for companies providing the bus with diagnostic hardware and software. Participants discussed various types of business models. Software companies responsible for evidence generation and validation can charge subscription- or use-based per license. Companies supplying scanners often operate on a purchase or lease model, and decentralized centers or scanner-on-tour suppliers could cover costs via direct-to-consumer supply or health insurance. For quality, competition between companies was regarded as crucial, and hospitals could consider purchasing an intramural diagnostic platform that supports multiple products from different vendors. This creates an overarching platform with competing products. Companies developing the software suites and the AI application that can be integrated into them should be prime movers here. Participants discussed the potential for the program to eventually become institutionalized, similar to breast cancer screening. Business models were seen as increasingly promising, especially once such programs began to generate large-scale data that can be leveraged to train data-hungry models.

Participants agreed that AI could offer an opportunity for the field that is in need of change, with the number of scans being requested rapidly increasing.

Participants envisioned a future in which AI could more and better support the role of medical professionals. Decentralized health care facilities can play a large role in the health care system and help decrease the growth of private companies offering whole-body scans, which was deemed undesirable as disparities in income and wealth may translate into unequal access to health care services, potentially leading to disparities in health outcomes.

Participants’ discussions shed light on a couple of frictions. First, they questioned whether the model should be organized in a centralized or decentralized manner, what the implications of decentralized testing would be for referring people to a hospital, and how data collected may or could be shared. Decentralized centers would need to build support within the broader network of health care institutions, hospitals, companies, and so on, for data sharing; they needed correct licensing, and interoperability must be addressed. Access and management of big data can create suspicions among patients, as they may think: “your data no longer belongs to you” [g1-s2].

Another potential friction arose around new role divisions and associated responsibilities. Laboratory technicians may outsource more tasks and perform fewer tasks themselves, radiologists may be hired on a consultancy (freelance) basis, and new roles may emerge (such as radiology assistants). This could increase the distance between the radiologist and the referring physician, which could complicate responsibilities and task distribution. A third friction point regarded sustainability:

This question was answered later on:

On the one hand, there was a need to manage this rising demand, with patients starting to demand scans for more certainty

versus the question of whether you need to better manage this demand. This can be done either by enlarging the role of AI and maximizing the use of the capacity of available scanners or by downsizing the demand by asking whether a scan is actually necessary, which is a decision AI can support in this scenario.

A group of participants chose to build their scenario around a concrete use case. The participants quickly decided to further develop their scenario around the idea that in the future, it becomes possible to predict, by analyzing either MRI data or only multimodal data, whether a citizen later in their life is likely to develop AD. This is achieved by screening citizens from a certain age group for the disease, either upon invitation or upon the citizens’ request (similar to the previous screening scenario). The benefit of using AI in very early stages to detect or predict AD, participants mentioned, can be that (future) patients are helped and trained to live independently or with minimal help for longer. This approach should improve the quality of life of patients, alleviate some of the societal burden of caregivers for these future patients, and may lead to reduced costs:

Explainability was discussed as an important model aspect for patients. A patient representative argued that the model should be able to explain how it arrived at a certain decision. However, a participant with a background in pathology contextualized that this can be very task dependent. The researcher gave an example that was more related to her own field of expertise:

The participants also briefly touched upon the business landscape, indicating that having multiple vendors would be a desirable aspect:

Participants found it important that implementing AI would be a meaningful addition to radiology, benefiting the individual patient. Thorough validation of the AI product can lead to improved trust in its use. Trust can also be enhanced through (human) explanation of the output. This means that either the algorithm should be able to explain its own output or radiologists should provide an explanation when a textual one is unavailable. Furthermore, patients should always remain in charge of their data and consecutive algorithmic outputs. That is, the data acquired upon clinical indication and the output of the algorithms should be shared only upon their voluntary consent. The participants also believed that individual patients or a society that shares personal data should monetarily benefit from commercial products. In other words, according to the participants, commercial parties should not be allowed to earn profits over products forever and profits should eventually flow back to society in some form, given that the AI product is essentially enabled by voluntarily provided personal data.

The participants argued that AI should deliver a product that radiology currently cannot. That is, true added value is achieved when AI detects what a radiologist cannot see, or when AI predicts future events that are not yet visible as pathology at the time of imaging. Participants discussed that the role of AI in this scenario is to assist with screening. Furthermore, predicting disease before it can manifest can lead to improved quality of life, specifically for AD. The participants highlighted that this is especially the case due to the human and emotional side of a progressive disease such as AD.

The participants discussed the need for increased scanning capacity as a potentially negative consequence of AI. Granting all citizens the opportunity for screening for early detection of AD is costly, partly due to purchasing the scanners. However, using this additional scanning capacity can lead to improved quality of life for the patients, a lower societal burden for informal caregivers, and lower health care costs if cost-effectiveness of the screening is proven. The risk of incidental findings increases, though, which needs to be balanced with increased benefits of early detection. Whereas the industry employee argued that regulations can sometimes be experienced as burdensome, the patient representative found the regulations justifiable with respect to privacy, indicating that without strict regulations the patient representative may not share personal data at all.

Finally, the participants discussed potential consequences concerning insurance and autonomy: if AI predicts a future diagnosis such as AD, an insurance company may want access to those results, given that they may want to assess and price coverage accordingly. Also, the question was raised whether, “will they [some authority] say automatically that you move to an Alzheimer’s [care] house and you have to sell your home?” [g3, s5]. According to the participants, these aspects were undesirable, and these open questions warranted some serious thought.

Seemingly, trust in AI represents a critical trade-off. On the one hand, trust could enable AI to live up to its potential as a stand-alone technology to effectively address complex tasks, such as medical diagnosis or large-scale hospital workflow optimization. On the other hand, patient representatives feared risk for automation bias. Additionally, participants discussed that excessive reliance on AI systems can diminish human supervision, increasing the clinician’s risk of automation bias and the erosion of critical decision-making capabilities. To break out of this trade-off and enhance trust, according to the participants, would often necessitate the adoption of interpretable and explainable AI systems, which radiologists can eventually check. Thus, achieving an optimal balance between trust and skepticism is essential to mitigate risks while harnessing the transformative potential of AI in radiology.

Implementing AI was considered a shared responsibility among stakeholders, each contributing distinct perspectives and expertise. Policy makers play a central role in establishing regulatory frameworks that guide legal-ethical boundaries for AI development and deployment. Industry players produce and scale practical AI applications within the established boundaries set by the aforementioned policy makers. However, when it comes to clinical decision-making, the participants foresee that medical specialists and health care workers will ultimately be responsible for the health and safety of their patients, and clinicians also feel that way. To reach this goal, it could be beneficial for medical specialists to cooperate with more technically oriented hospital personnel to perform quality assurance tests, particularly to ascertain that AI performs equally well on local datasets compared with how industry markets the performance of their products. Furthermore, according to the participants, there might be a conflict of interest when industry performs its own quality assurance. Thus, local testing may be a critical requirement for safe deployment and responsible use of AI in radiology and, according to the workshop participants, should be performed by independent hospital staff.

Given the rapid development of AI, radiological diagnosis is increasingly shaped by algorithmic output. These outputs are often treated as static and definitive, similar to finalized radiology reports. The static nature of AI’s output presents 2 interrelated shortcomings. Radiological diagnosis is often not a fixed product but a dynamic process. It relies on expert interpretation and is influenced by evolving clinical information, such as neurological examinations and pathology results. Several participants noted that diagnosis often emerges through iterative reasoning and collaboration with other medical specialists. AI systems, which typically generate preanalyzed findings or diagnostic probabilities, risk oversimplifying this complexity. Participants suggested that integrating multimodal data could help AI better reflect current clinical practice. In addition, static outputs could lead to a greater demand for diagnostic certainty, potentially leading to unnecessary follow-up: by detecting more abnormalities and including findings that may never become clinically relevant, AI could lead to more scans, consultations, and interventions. A potential solution to both shortcomings that was suggested by participants is a “human-in-the-loop approach,” where the radiologist remains responsible for interpreting AI outputs within the broader clinical context. Collaborative frameworks in which AI complements rather than replaces human expertise were seen as essential. These should include clear protocols for follow-up, shared decision-making with patients, and the development of transparent and explainable AI systems to support trust and appropriate clinical action.

Regulations play a dual role in technological innovation, acting both as a potential constraint and as a necessary enabler for progress, particularly in a fast-paced field such as AI in radiology. On the one hand, industry employees indicated that overly rigid or premature regulations can restrict innovation by imposing burdensome compliance requirements. On the other hand, patient representatives argued that well-designed regulatory frameworks are essential to ensure safety, fairness, and public trust—key requirements for adopting novel technologies. For example, as indicated by patient representatives, robust protection of patient privacy not only follows an ethical imperative but also creates a secure environment for patients to share their data. In other words, effective regulations are required for patients to feel safe when consenting to sharing their personal data, which is the cornerstone for all AI developments. Adaptive and dynamic regulatory approaches, capable of evolving alongside technological advances, were thus considered crucial for maintaining this balance. Following our workshop results, participants regarded effective regulation as an enabler rather than an obstacle to innovation, because it provides the foundation for trust necessary for groundbreaking technologies to thrive responsibly.

The integration of AI in radiology presents significant economic opportunities and challenges. Following initial investments in hardware infrastructures and software packages, there is a wide variety of potential applications on a hospital level, societal level, and vendor level that could yield a net economic benefit. Participants emphasized that AI could significantly improve operational efficiency within hospitals. By automating routine tasks, such as appointment scheduling and administrative workflows, AI has the potential to reduce labor costs and streamline internal processes. These improvements may lead to better resource allocation and increased throughput, contributing to overall cost-effectiveness in clinical operations. At the societal level, the economic benefits of AI in radiology are diverse and potentially far-reaching. Participants pointed to several ways in which AI could contribute to more efficient health care delivery and improved population health. For example, timely diagnosis may enable earlier initiation of therapy, which can improve outcomes and reduce the need for prolonged or intensive treatment. Similarly, more personalized care, which could be supported by AI’s ability to integrate and analyze complex data, may help avoid unnecessary interventions and enhance the overall effectiveness of health care services. These developments could contribute to long-term cost savings and a more sustainable use of health care resources. However, participants also cautioned that increased diagnostic sensitivity might lead to more incidental findings, which could raise demand for follow-up care and specialist input, potentially increasing overall health care expenditures. According to some participants, the responsibility for initial economic value assessments of AI systems lies with the vendors, who should estimate whether their solutions offer sufficient value before deployment. However, participants emphasized that independent researchers should conduct postdeployment evaluations. Similar to quality assurance processes, this separation is essential to avoid conflicts of interest and ensure unbiased assessments of economic performance.

Unintended consequences from AI might arise when the decision becomes binding rather than advisory. This decision could drastically affect a patient’s life. In the AD case, the participants discussed the extreme example of mandatory compliance with the AI decision for future patients when AD is predicted. Furthermore, if (by obligation) shared with insurers, these screening programs could have negative consequences for monthly health insurance premiums, thus penalizing individual citizens financially for factors beyond their control, which raises privacy, fairness, and proportionality concerns. Or, in the extreme case of extrapolating this issue, insurance companies may increase costs for patients at risk or deny (future) patients access to (additional) health and property insurances at all. Regulation should therefore be in place to prevent such a scenario from materializing. Moreover, patients have the right to not know. The participants agreed that this moral dilemma might be quite difficult for an AI to weigh, where there needs to be a good balance between medical necessity and patients’ rights and preferences. To counter this unintended consequence, human intervention might be the only solution.

Finally, it was discussed that AI is an energy-demanding technology. Combining AI with radiology for screening means that scarce resources might be used for citizens who may not (yet) express any disease symptoms. Thus, the participants discussed potential benefits for health care and the economy of AI in radiology should also be carefully weighed against the potential environmental impact. The monetary gain and gain in quality of life after the introduction of AI to radiology should therefore be proportionate to the environmental impact of additional interventions, such as screening in combination with AI.

In our analysis, we observed trust in AI as a complex topic the participants linked to, for example, performance and explainability. The participants perceived a paradoxical relationship between trust and effectiveness: the more trust they were willing to place in AI, the more AI could potentially operate as a stand-alone technology in radiology and fulfill its promise of improving clinical outcomes. Yet, this very trust introduces ethical tensions. As described by Singh et al [77], the ethical paradox in medical practices arises when efforts to enhance health care through AI risk diminishing the humanistic aspect of care, such as empathy and interpersonal interactions, thereby potentially provoking skepticism and decreasing trust. Reflecting on scenarios 1 and 2, AI could be deployed for triaging or identifying patients at risk. As such, AI applications may allow clinicians to focus their limited time on more complex and emotionally sensitive cases, potentially improving overall effectiveness of care and patient well-being, which were expectations that were reported earlier as well [37]. However, fully trusting AI may lead to dependency on AI or potential loss of human oversight. Similar concerns were also shared during the cocreation workshop. Patient representatives, among other stakeholders, confirmed that a human-in-the-loop was, in their opinion, still required. This idea was also shared in earlier research [63,78-80], although in other reports, patients have also indicated that autonomous AI could be acceptable on the condition that AI outperforms a radiologist [18]. Yet, patient representatives and other stakeholders seemed to have a positive attitude toward AI-assisted triaging, especially because the ultimate decision remains with a radiologist. The explainability of the AI-based results of the triaging could further help gain clinicians' confidence [53]. Discussions suggest that following these conditions, applications of AI in radiology may contribute to enhancing efficiency without compromising safety or trust. Based on the links between effectiveness, trust, explainability, and the humanistic aspect of care, we recommend designing AI systems that support clinicians, ensuring human-led decision-making and communication, while leveraging AI for effectiveness.

Participants in the workshop highlighted that there should not be conflicts of interest in the case of responsibilities. In other words, in environments where people must act responsibly to, for example, safeguard patients’ health, one should not be incentivized to act differently. For this reason, some participants emphasized that generating evidence for AI in early stages of development and real-world testing of AI should be performed by 2 distinct parties. Whereas industry was seen as responsible for product (co)development and early evidence generation, real-world implementation, thorough (local) testing, and longitudinal evaluation should be performed by independent hospital personnel, potentially in collaboration with more technically oriented personnel such as clinical physicists or technical physicians [81]. Furthermore, there was a seeming distrust of commercial parties to perform their own prospective evaluation. Earlier research focused on how to perform postinstallation evaluation [22,26,82-84]. However, this research did not dive into who should do so and what the possible impact on interconnected themes such as trust could be. Specifically, industry involvement should be limited in prospective evaluation, such that thorough evaluation can be performed with as little conflict of interest as possible. Our research demonstrates preliminary evidence that independent evaluation of AI in a clinical setting is pivotal to enhancing trust, and that trust can be achieved only by proper distribution of responsibilities along the innovation, installation, and evaluation chain. Therefore, we recommend that evaluation of AI is performed by hospital staff or independent external parties.

The proposed function of the AI within the radiology system varied within our workshop, ranging from screening to acting as a copilot for the radiologist. AI tools in medical imaging have traditionally functioned as narrow, outcome-focused instruments, for example, detecting polyps, segmenting tumors, or flagging anomalies in isolation [85]. However, radiology is fundamentally a process of evolving clinical interpretation, where patients’ histories, laboratory results, physical examinations, and ongoing changes in condition contextualize each series of images [86,87]. Multimodal data integration could enable AI to move beyond static predictions and could facilitate evolving model outputs [88]. Furthermore, the inclusion of large language models (LLMs) could provide narrative understanding, enabling AI to synthesize diverse inputs into coherent clinical insights [89-91]. In turn, a radiologist can check the output of the LLM or interact with the LLM, which could lead to further explainability and could foster trust, which also links back to emerging theme 1. This convergence transforms AI from a tool that offers answers into one that participates in reasoning. Importantly, this shift could allow AI to align more closely with the clinician’s cognitive workflow, adapting dynamically as new data come in [89-91]. It could also open the door to more explainable and human-centered AI systems that support, not replace, clinical judgment, as was also proposed earlier [63,78-80]. Thus, we recommend prioritizing iterative and interactive AI systems where radiologists can validate and refine AI-generated insights, ensuring that AI acts as a reasoning partner rather than a black box tool.

Moreover, the participants deemed the integrability of AI tools within existing hospital systems as a facilitator for successful adoption, as was also reported in earlier research [59]. Integrating these capabilities into unified platforms, rather than isolated tools, could thus further aid in the commercialization of AI in radiology. This process of platformization supports scalability, continuous learning, and interoperability [69], potentially enabling a shift from a one-off solution to comprehensive decision support ecosystems that are clinically and economically sustainable. Finally, participants regarded having multiple vendors as beneficial to the health care system, encouraging competition in either performance or costs, potentially benefiting patients. To support this, platforms should remain vendor-neutral, allowing easy integration from multiple providers. Whereas earlier research recommends four components: (1) (automatic) system triggering, (2) retrieval of relevant imaging series, (3) application of algorithm pipeline, and (4) availability of the results in Picture Archiving and Communication System of dedicated viewing software [92], we recommend conducting more research in the direction of cooperative guidelines that allow vendors to provide innovative solutions while remaining flexible for competition, potentially including regulatory sandbox environments, adaptive regulations, and value-based procurement.

In some discussions, the role of regulations came forward. While industry employees regarded regulations as a potentially impeding innovation, patient representatives regarded regulations as a requirement to share data. This paradoxical friction was also described in a recently published paper. According to Singh et al [77], the solution seems to be to strike the right balance between data accessibility and the obligation to safeguard patient confidentiality and to adhere to regulatory frameworks. This finding, previously reported among medical practitioners, was also identified in patient representatives, further confirming and reinforcing its broader relevance. The data paradox is quite complex: on the one hand, patient representatives in our workshop mostly preferred to retain ownership of their data. On the other hand, industry requires a diverse and up-to-date dataset, such that the AI products they develop generalize sufficiently well on various populations and do not suffer from data drift [40,46,51]. In light of this data paradox, future work should explore how hospital-private sector collaborations can be structured to enable responsible data sharing, with strong privacy safeguards, transparency, and mutual benefit for both public and private stakeholders. Recent global frameworks, such as the World Health Organization’s regulatory pillars for AI in health [93], the EU AI Act’s classification of medical imaging AI as “high-risk” [94], and international consensus initiatives such as FUTURE-AI [26], emphasize that striking this balance is not only desirable but also essential for compliance and trustworthiness. These frameworks advocate transparency, human oversight, and robust governance, which can guide collaborations toward ethical and legally sound data sharing.

During the workshop, participants raised the question of whether it was desirable for companies to earn money using voluntarily provided individual patient data without patient populations benefiting. A competitive landscape was described by the participants that could solve these issues. First of all, multiple vendors compete for the best product. The participants argued that this aspect should lead to better health outcomes for individual patients, linking to earlier argued ethical justification of the use of patient data if companies contribute to public health benefits [95]. In other words, the profits earned using patient data are reinvested in the companies to improve existing products or to create new products, which benefit both public and private stakeholders. Second, vendors can compete for the lowest price, which benefits the health care system as a whole, as price competition can lead to lower health care costs. At the same time, a diverse vendor landscape may help mitigate risks associated with vendor lock-in [96], ensuring that health care providers retain flexibility in choosing and switching between solutions as technology evolves and needs change. However, without more studies rigorously evaluating AI systems in real-world clinical settings, it is difficult to substantiate claims of added value [74]. We therefore advocate for more studies that evaluate AI in a real-world setting and that include multiple angles of value creation, such as decreased reporting times, added number of relevant incidental findings, and the prevention of unnecessary follow-up scans or interventions. This should lead to elaborated business models for both vendors (suppliers) and hospitals (buyers). Upon positive results, these studies could further facilitate AI adoption in clinics and provide vendors with useful insights about the possibility to generate revenue with their products. One example comes from Brix et al [11], who found that AI-based image reconstruction could lead to improved patient throughput, thereby obviating the need for an additional scanner. Such insights can help vendors to demonstrate the added value of their AI solutions, which may justify premium pricing and unlock new business models, especially when validated in real-world clinical environments.