A focus group design was selected for the study [35,36]. The inherent group dynamics of the focus group format allow for deep insights into the nature of sensitive and difficult-to-grasp topics, such as acceptance and trust regarding complex health-related technological developments [37]. This study used high-stakes personal health scenarios to capture often rather implicit assumptions, beliefs, values, and expectations. The study design and reporting followed the COREQ (Consolidated Criteria for Reporting Qualitative Research) guidelines (Checklist 1) [38].

Patients were recruited with the help of patient organizations and self-help representatives who forwarded the invitation to participate in the study to their members, allowing them to register for the study on their own initiative. In addition, calls were made via social media, and flyers were distributed at relevant events. The participants were selected according to the principle of “purposive sampling” [39]. The inclusion criteria considered were that participants had undergone preventive medical checkups and/or were currently under the care of a general practitioner or specialist. They also needed to be proficient in German and at least the age of 18 years.

Focus groups were conducted from November 2022 to June 2023 in the state of North Rhine-Westphalia, Germany. They were conducted until saturation of arguments was reached [40]. Each focus group consisted of 4 to 6 participants. Participants were asked to complete a short questionnaire on demographic data, experiences with radiological examinations, and affinity for technology using the German version of the Affinity for Technology Interaction Scale [41]. All focus groups were held as face-to-face sessions. No relationship was established between participants and researchers prior to the focus groups. The participants were informed about the project’s objectives, the interdisciplinary research team, and the participating research institutions. The empirical study was conducted by a senior researcher (psychologist) with extensive experience in qualitative social research methods. She was assisted by a junior researcher (psychologist).

At the beginning of the focus groups, participants received standardized information on AI–based diagnostics in radiology (Multimedia Appendix 1), as well as information on the purpose and procedure of the focus group. A semistructured, pilot-tested guide served as the framework for the discussion in all focus groups. Prior to the study, interviews were conducted with representatives of patient organizations and medical professional associations to identify relevant topics and develop the guide in a manner appropriate for the target group. The interview guide comprised key questions and additional questions.

To start the discussion, the participants were asked the following questions: “What opportunities and risks do you see?” and “What would need to be guaranteed for you to endorse the use of AI in radiology? Are there circumstances under which you would reject its use?” In the main part of the discussion, key questions were introduced through personal involvement scenarios. The first scenario described the following situation: “Suppose your physician is unsure whether you are developing a serious illness, such as multiple sclerosis, and would like to refer you for an MRI scan to detect the disease as early as possible and, if necessary, initiate treatment.” Participants were then asked, “Would you choose a radiologist who does or does not use AI? What would your decision depend on? Which AI function would you feel most comfortable with when evaluating images or making a diagnosis?” The second scenario described a situation in which an MRI report had been generated using AI and contained indications of abnormalities that suggested the presence of a serious disease, such as multiple sclerosis. Participants were subsequently asked, “Would you trust the AI diagnosis? Under what circumstances? Who would you trust more, the physician or the AI, if your physician doubts the AI diagnosis?” In the third scenario, participants were asked to imagine a development in which AI is used extensively in diagnostics. The following key questions were asked, “Could this change your physician-patient relationship, or would it remain unaffected? What expectations do you have of your physician in relation to this future development?”

For all key questions, participants were encouraged to explain the reasons for their views. The key questions were visualized to stimulate thinking and argumentation. All focus groups were facilitated by the author in a neutral role. A “funnel design” was used [42]. In each group, 1 research assistant noted nonverbal information on the interaction among the participants. Each focus group lasted around 2 hours, was audiotaped, and transcribed verbatim.

Data analysis was performed using structured qualitative content analysis according to Mayring [43], as well as framework analysis [44]. Both methods are research techniques for the systematic analysis of manifest and latent communication content. The strength of structured qualitative content analysis lies in its theoretically grounded, rule-based categorization. When combined with matrix-based framework analysis, comparisons can be made across themes and cases, enabling patterns to be identified. This allows for a broad and in-depth analysis of the data, particularly when dealing with extensive datasets—as in this study.

Deductive and inductive categorization were combined using the software MAXQDA 2022 (VERBI Software GmbH). Deductive categories formed the basic framework, which was supplemented with inductive categories. Conflicts arising between the 2 approaches, such as redundancies and the resulting ambiguities in the category system, required revisions, standardizations, and re-examinations based on the raw data. The resolution of such conflicts was managed through a high degree of flexibility in the creation of the category system and transparent, traceable documentation of all disagreements and adjustments. Any disagreements arising during the development of categories were discussed within the research team until consensus was reached. In addition to the author, another coder from the research team was involved to ensure quality. There was good interrater reliability (κ=0.73).

In the first step, both raters read all transcripts, wrote brief summaries of what had been discussed, and made comments on the text passages that appeared relevant to them. In the second step, main thematic categories were developed deductively based on the literature review and the guide. A total of 4 focus groups were rated by the 2 raters independently of each other, the results were discussed, and the preliminary categories were modified and added to. The entire material was then rated based on the main categories. In the third step, the inductive development of subcategories was carried out. In several rounds of discussion, the entire category system was broken down, and the entire material was then rated independently. In the fourth step, statement matrices were organized according to topics, and, finally, documents were created on the basis of quotations.

This focus group study with patients was part of an extensive study that also included focus groups with physicians and a representative survey of the German population to investigate attitudes toward radiological diagnostics using AI. The study was approved by the Committee for Ethics in Research of the Forschungszentrum Jülich (FRAIM.20221005). Written informed consent was obtained from all participants prior to the empirical study. Data were pseudonymized and stored in a protected digital environment. Participants received €100 (US $117) for their participation.

A total of 7 focus group sessions were conducted with 34 adults aged between 23 and 85 (mean 49.06, SD 17.08) years. Fifteen participants were female, and 19 were male. A majority (30/34, 88.2%) reported Abitur, the German equivalent of a high school diploma allowing university studies, as their highest level of general education. A total of 4 (11.8%) participants reported completing 10 years of formal education (German “Realschul-Abschluss, Mittlere Reife,” or equivalent). The majority of participants (25/34, 75.8%) were employed. Almost all reported having a regular family physician (32/34, 94.1%) and having undergone at least 1 radiological scan (32/33, 97%). The overall score for the 9-item Affinity for Technology in Action Scale indicated a relatively high affinity for technology among the sample (mean 4.14, SD 1.04; Cronbach α=0.94).

The following results highlight the opportunities and risks of using AI in radiological diagnostics from the patients’ perspective, as well as the circumstances that can foster its acceptance and trust. Citations of participants’ statements are provided with the focus group number (FG) and participant number (#). Original statements have been translated into English for publication.

The following aspects emerged during the focus group discussions: (1) efficiency, data integrity, and quality assurance; (2) data privacy, transparency, and explainability; (3) the role of physicians; (4) trust in physicians versus AI; (5) trust-building functions of physicians; and (6) the physician-patient relationship.

In general, all participants agreed that the use of AI offers opportunities for more efficient and accurate diagnostics. However, a more efficient workflow and shorter reading times, thanks to AI, were only rated positively if these improvements shortened waiting times for appointments in radiology, enabled diseases to be detected earlier, and gave physicians more time for their patients.

The participants considered the potential of using AI to reduce human errors and achieve a high level of accuracy in diagnostics to be positive, particularly in the case of abnormalities that are difficult to identify:

In addition, participants feared that the increasing commercialization of the AI market would lead to companies basing their actions purely on economic aspects. This could not only result in insufficient validation of data and discrimination against certain patient groups due to a lack of balance and diversity in the datasets, but also in algorithms being manipulated by companies to suit their own interests:

Careful quality assurance, both in the development and approval of AI algorithms and in their ongoing application, was highlighted as one of the fundamental prerequisites for acceptance (Table 1). Participants emphasized that high data quality, in terms of scope, diversity, and continuous updating, is crucial and that this must be ensured through quality standards and regulation. They also acknowledged that science plays an important role in building trust in the quality of the algorithms. The participants considered it essential that the accuracy of diagnostics using AI is better than without AI:

Some participants found it contradictory to place higher demands on technology than on physicians in order to trust the technology.

Most participants considered existing data protection in Germany to be sufficient for the use of AI in radiology. However, a minority of participants expressed extreme discomfort with the collection and storage of data for the development of AI and its application. These participants stressed the great importance of preserving the freedom of choice.

Ensuring transparency in the use of AI was important to all participants. As technical nonprofessionals, the participants did not feel able to assess the quality of AI in diagnostics. Given the potentially serious consequences of misdiagnosis, the types of training data used should be disclosed, and the accuracy of the algorithm should be specified and clearly communicated. The reputation of the companies that develop the software was considered a good indicator of the software’s trustworthiness. Although transparency was seen as a necessary prerequisite for the acceptance of AI use, participants were unsure whether this would help build trust in AI diagnoses if they were personally affected.

Participants expressed contrasting attitudes toward the explainability of AI diagnoses. One group considered the “black box” nature of AI to be unproblematic as long as superior performance was proven. This argument did not convince the roughly equally large group that found the “black box” nature of AI problematic. For these participants, the process by which AI arrives at a result must be reconstructible in order to trust an AI diagnosis, clarify misdiagnoses, and assign responsibility.

The framework conditions that patients demanded for the use of AI in radiological diagnostics are listed in Table 1. The quotes are typical examples of the respective categories.

Meeting the technical requirements and framework conditions for the use of AI does not automatically lead to a willingness to choose a radiologist who uses AI.

In a situation where a suspected serious neurological disease, such as multiple sclerosis, needs to be clarified radiologically, all but 3 participants indicated that they would choose a radiology practice that uses AI. However, this preference depended on the interpretation and evaluation of the clinical relevance of abnormalities and the final diagnosis remaining the responsibility of radiologists in consultation with the referring physicians.

With the exception of 1 person, who preferred automated AI in all cases, the participants felt strongly that AI should only be used as a tool in the hands of an experienced physician:

The participants considered an independent medical assessment to be important for trusting the AI diagnosis. This was particularly evident in cases where the attending physician doubted the AI’s diagnosis.

If abnormalities were found that required treatment, the AI diagnosis did not automatically lead to a willingness to undergo treatment. Only 1 participant mentioned that they would trust the AI diagnosis and follow the treatment. On the other hand, only a minority trusted the physicians’ judgment unreservedly in such cases. The majority pursued a safety strategy and wanted to eliminate uncertainty by seeking a third opinion. An AI-supported procedure could also be used again, but in the vast majority of cases, a physician was considered necessary to review the AI result again.

One of the reasons for the important role of the physician is the requirement to check the AI output. Many participants demanded not only a plausibility check but also a verification of the AI output through an additional independent assessment. For most participants, the decisive factors were characteristics and skills that were attributed only to physicians and considered indispensable for ensuring well-balanced results based on individual cases (Table 2). In their opinion, physicians have a holistic view of human beings, which enables them to classify the clinical picture in terms of the patient’s entire biography and take into account individual factors that may be relevant to the diagnosis.

It was also argued that physicians can use their intellectual abilities to deviate from routines. Their intuition, creativity, and flexibility enable them to do so. It is precisely these qualities of physicians that are highly valued for recognizing abnormalities beyond the specific issue at hand and for diagnosing rare diseases.

At the same time, however, participants were skeptical that the trust-building functions of physicians would be realized. In their view, the use of AI could lead to a decline in the competence of radiologists. The greatest risks were seen in physicians’ overconfidence in AI output and insufficient attention to, and monitoring of, the results:

In addition, the participants assumed that the combination of AI and radiologists is only a transitional stage toward automated AI. The participants viewed this development with great concern:

It was important to the participants that the social and emotional aspects of the physician-patient relationship were preserved (Table 3). This is not just about communication but rather about experiencing empathy, a reciprocal exchange relationship, and the feeling of being in good hands. It was crucial for participants to know that physicians would continue to take responsibility for their patients even when using AI. It is important to them that physicians make decisions based on the well-being of their patients. For patients, this means that physicians act not only on the basis of their knowledge but also on the basis of values. In addition, physicians should explain to patients what AI is, how it is used, how its results should be evaluated, and, above all, what the added value of an AI-supported diagnosis is.

Most participants were concerned about the threat that the physician-patient relationship might suffer:

They were skeptical that relieving physicians’ workloads through greater efficiency using AI would actually result in more time for patients. Rather, they fear that dehumanization and deindividualization in health care will increase.

The results of this study show that patients are generally open to the use of AI in radiological diagnostics. This is consistent with a number of other studies that demonstrate openness to AI in other medical fields among various population groups [45,46]. Patients recognize the potential for more accurate, faster diagnostics based on expanded medical knowledge, while emphasizing the importance of quality assurance not only during approval but also during ongoing application.

However, the in-depth discussion in this study shows that the advantages mentioned will only increase acceptance of AI if, at the same time, the health care system is set up in such a way that the use of AI leads to real added value for patients. Patient-centered care was one of the most important factors that need to be met for patients to endorse the use of AI. In particular, this means that physicians should use the time saved through the use of AI to provide more personalized care for patients. Rationalization effects and cost savings—such as faster patient flow (ie, more patients in the same amount of time), staff savings, and even the replacement of radiologists by AI—were major reasons why patients would reject its use. Furthermore, the expected increasing commercialization of AI development and application was associated with negative impacts on patient care. These aspects have received too little attention in previous studies.

This study suggests that it is not only technological features—in particular, high performance and reliability—that are decisive for the acceptance of AI; rather, the quality of the entire sociotechnical system plays a role.

This is particularly evident when it comes to the specific willingness to choose a radiologist who uses AI if there is a suspicion of a serious disease that requires radiological clarification. Patients are uncertain in this regard. In such a situation, patients place high demands on the role and function of physicians in particular. As already shown in other studies, it is very important to patients that AI findings are checked by physicians [47,48].

Additionally, this study was able to expand previous findings by showing that what is required is not merely a plausibility check, but rather an independent diagnosis that broadens the diagnostic context. The question is not whether AI is used as a first or second opinion; rather, physicians should contribute specific skills and competencies to the diagnostic process that, from the patients’ perspective, AI lacks, such as intuition, flexibility, and creativity. Only then, in the participants’ view, will the individuality of each case be given appropriate consideration and abnormalities beyond the specific question for which the AI had been trained will be discovered. The inability of AI to integrate isolated imaging results within the broader context of the patient’s living situation is also discussed in the radiology community as a significant limitation of today’s AI systems. It is frequently emphasized that the more nuanced expertise of radiologists is indispensable in the diagnostic process [49].

However, patients are concerned that these very functions of the physician, which promote trust, could be weakened by overreliance on AI results and the increasing deskilling of physicians. This is all the more concerning as the trend toward automated AI is both expected and feared. This contrasts with the fact that radiologists themselves consider the replacement of radiologists unlikely [50].

The results illustrate that patients are highly uncertain about whether they would actually opt for AI diagnostics if they were personally affected. This is also indicated by the finding that patients tend to trust physicians more than AI results if the physician doubts the AI results.

The results of this study show that it is not only the competence of physicians that matters, but also the physician-patient relationship. This result confirms previous findings that the maintenance of the familiar physician-patient relationship is decisive for patients [51,52]. The in-depth discussions in the focus groups clarified which characteristics of the physician-patient relationship are important when using AI. An empathetic relationship and the physician’s responsibility to care for the patient’s well-being are required. In addition, patients emphasized that physicians should take on an advisory role with regard to the pros and cons of using AI in individual cases. This also includes being able to provide information about the quality of AI. Various studies have shown the importance of transparency in the use of AI and the desire of patients to be informed about whether AI is being used [53,54]. The results of this study reveal, in addition, that physicians, as trusted mediators of information, should play a central role in dispelling concerns and fears. These factors in the physician-patient relationship could build trust in the diagnostic process. Trust can promote acceptance of the use of AI [55,56]. However, AI also has the potential to disrupt person-centered physician-patient relationships [57,58]. This study demonstrated that this is precisely what patients are concerned about when it comes to the use of AI in radiological diagnostics.

The findings of the study, and in particular the important role of the physician-patient relationship, arise within the context of Germany’s highly developed health care system, which is characterized by high standards. As a rule, equal access to new technologies is guaranteed, with costs being shared within the system in accordance with the principle of solidarity. Ethical and regulatory considerations are taken into account prior to the implementation of new technologies. In health care systems in other countries that are structured differently, patients might consider other factors to be relevant for acceptance and trust, or weigh them differently.

This study examined patients’ perspectives on the use of AI in radiological diagnostics using an innovative focus group method. Unlike many previous studies, which tended to use a more abstract approach, this study used scenarios as stimuli to gain insights into patients’ reasoning regarding the use of AI in situations that affect them personally. This approach expands current research by focusing on patients’ motives in their decisions for or against a radiologist who uses AI. A limitation of the study could be that the patients had a high level of education and a strong affinity for technology. Both of these factors could lead to a more positive attitude toward AI compared to the general population [59]. Although several channels were used to reach participants and no prior knowledge of AI was required for participation, self-selection and recruitment via patient organizations and self-help groups may have introduced some selection bias. Therefore, the results may have limited applicability to patients with a low socioeconomic status and low digital and health literacy. However, this does not detract from the validity of the argumentation patterns identified in this qualitative study. Nevertheless, validating the results with a more heterogeneous sample in terms of socioeconomic status and digital literacy might further strengthen the conclusions.

This study shows that patients are generally open to the use of AI in radiological diagnostics, depending on the design of the sociotechnical system. It is not only the performance of AI that is decisive; extensive quality assurance measures are necessary, as is, above all, the integration of AI into the health care system in a way that ensures real added value for patients. At the same time, the study reveals that in cases of personal involvement, the decision to use AI—beyond the perceived abstract advantages and disadvantages—depends significantly on the trust-building role of the physician in the diagnostic process. Consequently, framework conditions for the use of AI are required that avoid constraints preventing physicians from acting independently and degrading them to “messengers of AI.” Overconfidence in the AI result and deskilling must be prevented. The integrity of the physician-patient relationship should be preserved. On this basis, patients do not feel “at the mercy” of AI but rather “in good hands” in the diagnostic process guided by the physician. This can help to overcome existing uncertainties. Acceptance and trust can be promoted by strengthening the trust-building role of physicians when AI is used.