Breast cancer is the most common cancer among women globally and a leading cause of cancer-related morbidity and mortality [1,2]. Early and accurate diagnosis is essential to improving outcomes, yet remains vulnerable to human error, including interpretive variability and diagnostic inaccuracies [3]. In Austria, nearly 90,000 women were living with breast cancer as of 2022 [4], and increasing incidence projections combined with physician shortages underscore the need for scalable and efficient diagnostic tools [5,6].

Artificial intelligence (AI) systems trained on large datasets have demonstrated strong performance in detecting malignancies and reducing diagnostic errors [7,8]. In some screening contexts, AI has even outperformed radiologists in stand-alone mammography detection tasks [9]. Nevertheless, AI is not intended to replace clinicians, and concerns remain regarding overreliance on automated systems without appropriate human oversight, which could have serious clinical consequences [10]. Accordingly, AI is increasingly conceptualized as a decision-support tool that may enhance diagnostic efficiency, accuracy, and workflow while preserving clinician responsibility [11-13].

Despite its potential, the adoption of AI in clinical practice remains uneven across health care systems [5]. International research consistently shows that integration into routine care is hindered by a combination of organizational, technical, and human factors [14,15]. Understanding barriers and facilitators to AI adoption is therefore critical for enabling ethical, effective, and sustainable implementation [14].

The technology acceptance model (TAM) posits that technology uptake is influenced by perceived usefulness (ie, the expected benefits of use) and perceived ease of use (ie, the anticipated effort required), which together inform users’ attitudes and subsequently influence behavioral intention and actual use [16,17]

Extensions of TAM, including TAM2, incorporate subjective norms, capturing the influence of perceived social pressure on adoption decisions [18,19]. The Unified Theory of Acceptance and Use of Technology (UTAUT) further integrates these concepts by identifying by identifying performance expectancy (expected benefits), effort expectancy (expected ease of use), social influence (influence of important others), and facilitating conditions (perceived organizational and technical support) as core determinants of technology acceptance, moderated by individual characteristics such as age, sex, and experience [19].

Across international health technology research, perceived usefulness, usability, and social norms have been consistently associated with clinicians’ willingness to adopt digital innovations [14,20,21]. In radiology and oncology settings, perceived benefits such as improved diagnostic accuracy, workflow efficiency, and enhanced quality of patient care are commonly reported facilitators [22-24]. Further health research found that systems perceived as intuitive and well-integrated into clinical workflows are more likely to be adopted, highlighting the importance of design usability and operational fit [21]. Conversely, persistent barriers include limited infrastructure, high costs, lack of training, workflow disruption, and concerns about increased workload [25].

Qualitative evidence further highlights psychological and professional concerns, including limited AI knowledge, lack of explainability, perceived threats to professional autonomy, and fears of role displacement [26]. In radiology in particular, unfamiliarity with AI tools and concerns about reduced autonomy have been repeatedly identified as salient barriers [21]. Beyond traditional TAM and UTAUT constructs, trust has emerged as a critical determinant of AI adoption [27]. Research suggests that system transparency, reliability, and explainability are integral to building clinician trust, which in turn significantly influences the likelihood of adopting AI-based tools [27].

While AI acceptance has been explored in general health care contexts, its specific application in breast cancer diagnostics remains underexamined, despite growing clinical relevance [9]. Country-specific factors, including regulatory frameworks, digital infrastructure, and workforce capacity, are likely impacting the health care system’s readiness to adopt AI [28]. More targeted research is therefore needed to understand the factors influencing adoption, particularly within national contexts such as Austria [21].

In addition, international studies have identified a range of barriers and facilitators to AI use and suggest that attitudes may vary by level of AI experience [21,26]. However, few studies have directly compared clinicians who currently use AI with those who do not, limiting insight into how hands-on AI experience influences attitudes, perceived barriers, and implementation readiness.

To address the aforementioned gaps, this study investigates AI use in breast cancer diagnostics within the Austrian health care system and explicitly compares physicians with and without direct AI experience.

Specifically, this study has two primary objectives:

These insights aim to inform practical, evidence-informed, and human-centered strategies for integrating AI into breast cancer diagnostics [29].

We used an embedded mixed methods [30], cross-sectional questionnaire design, adopting a pragmatic epistemological stance [31] to prioritize methodological fit with the study aims. The study examined (1) barriers and facilitators to AI adoption among AI users and nonusers using quantitative descriptive analyses complemented by qualitative open-ended responses, and (2) predictors of attitudes, intentions, and perceived likelihood of future AI use using quantitative regression analyses.

This study was conducted as part of a master’s thesis project at the University for Continuing Education Krems, Austria (EH). Eligible participants were practicing physicians involved in breast cancer screening or diagnostics. Exclusion criteria included nonphysicians, retired clinicians, and those not actively engaged in diagnostic work.

Data collection took place over an 8-week period from March to April 2023. A German-language online survey was distributed via predefined professional and institutional channels using email invitations and an electronic flyer. Primary dissemination occurred through the Austrian Medical Chamber (Ärztekammer), complemented by distribution via university channels, social media platforms, and relevant professional networks. Participants accessed the survey via a secure link, and only respondents who consented and met the predefined screening criteria were able to proceed with the survey.

The study received ethical approval from the Ethics Committee of the University for Continuing Education Krems (EKGZ 55/2021-2024). All participants provided informed consent electronically before accessing the survey. Participation was voluntary, and respondents could withdraw at any time without consequences. Data were collected anonymously, with no personally identifiable information recorded. Secure data storage procedures were applied to protect confidentiality. No financial compensation was provided for participation.

Quantitative survey items were adapted from previously validated TAM [16,20] instruments and informed by relevant digital health research [21]. Items were tailored to the context of breast cancer diagnostics [16,19]. The first section captured sociodemographic data, AI awareness, and usage. Sex was assessed via self-report; sex or sex assigned at birth was not collected. Attitudes toward AI were measured using two 7-point Likert items averaged into a composite score; additional items included perceived AI-related skills, perceived usefulness, and perceived attitudes of colleagues. The second section addressed perceived barriers (eg, limited availability, high costs, usability, data concerns) and facilitators (eg, efficiency, diagnostic precision, stress reduction) using multiple-choice and open-ended items. The final section measured future intention and likelihood of AI use with a single 7-point Likert item, followed by optional free-text elaborations. Item phrasing was adapted to reflect whether participants were current AI users or nonusers.

Qualitative data were collected through open-ended questions embedded in the online survey, which invited participants to elaborate on perceived barriers and facilitators to the use of AI in breast cancer diagnostics. Data collection was limited to the survey format, and no additional qualitative follow-up (eg, interviews) was conducted. The full wording of all the translated survey items is provided in Multimedia Appendix 1.

A total of 53 participants completed the survey. Seven participants were excluded: 6 participants because they were not currently working in breast cancer diagnostics and one participant due to missing demographic and professional information. The final sample comprised 46 participants. Of these, most (40/46, 87%) reported being aware of AI technologies for breast cancer diagnosis, and approximately half (24/46, 52%) indicated they currently use AI in their diagnostic practice. The sociodemographic characteristics of the final sample are summarized in Table 1.

Responses to barrier items were available for 44 participants, and responses to facilitator items for 41 participants. Qualitative data were derived from open-ended responses provided by 41 physicians (56% AI users). The qualitative analysis yielded 21 codes, comprising 12 facilitator categories and 9 barrier categories related to AI adoption. An integrated overview of qualitative categories and corresponding quantitative frequencies is presented in Tables 2 and 3. Additional details on qualitative categories, codes with illustrative quotes, and aligned quantitative data are provided in Multimedia Appendix 3.

Among participants who completed the barrier items (n=44), most (40/44) reported at least one perceived barrier to AI use. Conventional content analysis identified 9 codes pertaining to barriers, which were organized into 4 overarching categories. Table 2 presents a joint display of perceived barriers to AI adoption, integrating quantitative response frequencies with qualitative categories and codes.

Infrastructural, financial, and technical limitations were among the most frequently cited barriers to AI adoption. Availability was reported as a significant concern, particularly among nonusers: 81% (17/21) of nonusers indicated that AI tools were not accessible in their workplace, compared with only 22% (5/23) of users (P<.001; z=3.879). Qualitative responses elaborated on these findings, with nonusers noting, “The software is not available in the department,” or “AI tools are not yet integrated in my workplace.” High costs were also raised as a concern, reported by 43% (9/21) of nonusers and 35% (8/23) of users, though this difference was not statistically significant. Several participants described AI as being “too costly to implement here,” underscoring the perceived financial barriers in resource-constrained environments.

In addition to cost, technical integration was reported as a key limitation by 32% (14/44) of participants, with significantly higher prevalence among users (9/23, 39%) compared with nonusers (5/21, 24%). Qualitative accounts contextualized these concerns: several AI users noted that current systems were not fully compatible with existing software, adding rather than reducing workflow burden. As one user explained, “AI doesn’t align well with our current software and therefore adds to the workflow.”

When asked about motives for future use, several participants emphasized that seamless integration and workplace availability are essential preconditions for future use. As one nonuser noted, “AI must be readily available and smoothly integrated to be adopted in daily practice”. Collectively, these results highlight how cost, resource access, and system compatibility may act as gatekeepers that must be addressed to enable broader and sustained AI adoption in clinical settings.

Both quantitative and qualitative data indicated that access to training and ease of use are critical factors influencing AI adoption. Among current users, 35% (8/23) indicated limited user-friendliness as a challenge. In contrast, 33% (7/21) of nonusers reported a lack of knowledge with AI as a barrier. Qualitative responses provided additional context. One nonuser noted, “I haven’t used AI because it’s still a new concept and training hasn’t been offered,” highlighting the absence of structured onboarding opportunities. This category was also reinforced with responses to the question about what would motivate future use: nonusers, in particular, emphasized the need for training to build confidence and capability.

Among AI users, concerns about IT security (26% vs 10%) and data protection (17% vs 10%) were slightly more prominent compared with nonusers, though these differences were not statistically significant. Qualitative responses supported these findings, with 3 AI users raising concerns about patient privacy and data security. For example, one respondent noted, “AI brings up data protection issues for patients”, suggesting that security considerations may become increasingly important as AI use expands.

Challenges related to workflow disruption and trust in AI outputs were primarily reported by current AI users. Although AI was generally perceived to enhance productivity, 11% (5/44) of participants reported a lack of effectiveness as a barrier. This concern was elaborated upon in open-ended responses, which described how AI systems occasionally generated false alarms or required additional verification steps, thereby limiting potential time savings. For example, one user explained, “AI sometimes creates extra work due to false alarms”, highlighting the potential for AI to unintentionally increase workload.

Trust in AI outputs also emerged as an important issue. Several participants described a need to double-check AI-generated results. One user stated, “I trust AI, but I double-check its suggestions”, suggesting that partial trust may reduce the efficiency benefits of AI by necessitating manual oversight. These findings indicate that both technical and psychological barriers, specifically, limited system integration and cautious trust, may attenuate the intended benefits of AI in clinical workflows.

High costs, limited availability, and training needs were identified by both users and nonusers as key barriers to AI adoption. Nonusers more frequently cited access and capability-related challenges, whereas users highlighted issues related to system integration and workflow disruptions as barriers.

Across both groups, improved diagnostic accuracy emerged as the most frequently mentioned facilitator of AI use in breast cancer diagnostics. Users emphasized several clinical benefits, including improved patient care and increased efficiency. Nonusers also acknowledged AI’s potential to support efficiency but placed greater emphasis on the importance of enabling conditions, such as structured training and workplace accessibility, to support future adoption. Addressing both shared and group-specific factors will be essential for developing targeted implementation strategies and ensuring broader, equitable uptake of AI in clinical practice.

This study aimed to examine the key barriers and facilitators to AI for breast cancer diagnosis among Austrian physicians, with a specific focus on comparing current users and nonusers. We also sought to identify how individual, social, and technological factors predict physicians’ attitudes toward AI, their intention to use it, and their perceived likelihood of future adoption. While most participants were aware of AI technologies, only approximately half had used such technologies in their clinical practice.

Key facilitators included perceiving benefits in terms of diagnostic accuracy, professional growth, workflow efficiency, and overall quality of patient care. Barriers included limited access, high costs, integration issues, and insufficient training opportunities. Nonusers reported anticipated benefits, such as improved diagnostic accuracy and workflow efficiency, but identified limited availability as a key barrier to adoption. In contrast, users cited concrete benefits, including enhanced diagnostic precision, quality of care, and productivity, while also expressing some concerns regarding cost, integration with existing systems, and trust in AI-generated outputs.

Regression analyses indicated that perceptions of more facilitators and fewer barriers were strongly associated with more favorable attitudes and higher adoption intentions. Social and demographic factors also played a role: younger physicians and those in environments where colleagues valued AI were more likely to have favorable attitudes and stronger intentions to use AI. Experience with AI and related skills were consistent predictors of intention and perceived likelihood of future use, while older age was negatively associated with adoption likelihood.

Practically, our findings offer valuable insights for health care leaders, educators, and policymakers aiming to support the responsible integration of AI in breast cancer diagnostics. Implementation strategies could benefit from not only highlighting perceived advantages but also addressing specific barriers that appear to differ between current users and nonusers.

As outlined in Table 5, potential strategies may include consideration of demographic differences (eg, age), social norms, and levels of technical readiness. Targeted support, such as age-appropriate training, peer-led initiatives, and structured upskilling opportunities, might help reduce disparities in AI engagement. Findings suggest that adoption may not be a binary decision but a staged process with evolving support needs. Early-stage users may benefit from structural interventions (eg, access and training), while more experienced users may require support with integration and workflow adaptation. This distinction implies that implementation efforts may need to be adapted over time and tailored to users’ varying levels of familiarity and experience with AI.

Furthermore, the influence of social context, particularly perceptions of colleagues’ attitudes, suggests that leveraging peer support and fostering a culture of openness toward innovation may play a role in facilitating adoption. Finally, embedding AI-related content into medical education and continuing professional development may support sustained engagement, particularly among groups less likely to adopt AI initially. These insights, while preliminary, could inform more inclusive and context-sensitive implementation strategies in future work.

Future research could extend our findings through longitudinal studies that track real-world AI adoption over time, as well as intervention studies evaluating tailored support strategies. Comparative research across countries and medical specialties could further clarify how structural, educational, and cultural factors influence implementation [22,38]. In our sample, differences in access, training, and institutional support between users and nonusers highlight the need for flexible, context-sensitive implementation strategies. A governance framework, such as that proposed by Hassan et al [27], may support this by addressing trust-related barriers through structured oversight across the AI lifecycle and by informing policy and institutional priorities that enable more equitable access, investment, and training.

Our findings also support recent calls to expand AI education across the medical training pipeline. For example, Weidener and Fischer [39] emphasized the integration of AI ethics and practical skills in undergraduate curricula. Similarly, Wartman and Combs [40] argued that medical education must evolve beyond the information age to meaningfully prepare future clinicians for AI-enabled practice. Training physicians in AI as part of undergraduate and continuing medical education may enhance adoption by improving relevant skills, confidence, and preparedness for clinical integration, particularly among younger cohorts already engaging with AI informally [39].

Future research could also explore patient perspectives, particularly regarding data privacy, consent, and trust in AI-assisted diagnostics, to support ethical and patient-centered implementation. Although our study focused on physicians, these concerns are especially relevant in breast cancer screening, where patient acceptance may hinge on transparency, perceived fairness in algorithmic decision-making, and how AI is integrated into clinical workflows [41-43].

A key strength of this study is its mixed methods design, which enabled triangulation of quantitative findings with qualitative insights. This approach allowed for a comprehensive understanding of AI adoption by comparing users and nonusers, and by identifying psychological, technical, and social drivers of adoption.

However, some limitations must be acknowledged. First, the sample was relatively small and limited to Austria, which may affect generalizability. Second, the cross-sectional design captures perceptions at a single point in time and does not allow for causal inference. Longitudinal research is needed to track evolving attitudes and behaviors. Third, reliance on self-reported data introduces potential bias, which could be addressed in future studies through objective usage metrics. Fourth, although items were adapted from validated TAM instruments and pilot tested for clarity, formal psychometric validation was not performed. The brief attitude measure and exploratory design, therefore, warrant cautious interpretation of these findings.

Regarding the quantitative analyses, the small sample size and the number of predictors in the regression models may limit statistical power and the stability of estimates. To address these concerns, we conducted extensive sensitivity and robustness checks, including alternative codings of barriers and facilitators, reduced models including only the main predictors, and ordered logistic regressions to account for the ordinal nature of the outcomes. These analyses indicate that the direction and magnitude of the observed effects are largely robust, although statistical significance varies slightly across specifications. Finally, recruitment via professional networks and social media groups meant that the total number of physicians reached was unknown, and a formal response rate could not be calculated. This may introduce self-selection bias and limit the generalizability of the findings beyond physicians with an interest or experience in AI. Despite these limitations, our study offers valuable exploratory insights into the factors influencing AI adoption in breast cancer diagnosis and provides a foundation for future implementation research and policy development.

This study provides timely insight into AI adoption for breast cancer diagnostics among Austrian physicians and highlights several key contributions. First, it focuses on a specific clinical use case (breast cancer diagnostics) within a national health care context, extending prior work that has often examined AI adoption in radiology more broadly. Second, by comparing physicians with and without hands-on AI experience, it shows that barriers, facilitators, and support needs differ by level of AI exposure.

Third, the findings suggest that AI adoption may be better understood as a staged process rather than a binary decision. Early uptake appears to depend mainly on access and training, whereas later-stage use is more affected by workflow integration and trust in AI outputs. This refines existing TAMs by highlighting the role of implementation context across stages of engagement.

Finally, the results have practical implications for implementation. Investment in targeted training, organizational support, and peer-enabled learning may support more equitable adoption. Strategies that account for differences by experience level and demographic factors may be particularly important for achieving sustainable and responsible integration of AI in breast cancer diagnostics.