Austrian Physicians Identify Barriers to AI Breast Diagnostics

The Journal of Medical Internet Research has published a study by Marlene Kritz, Erol Holawatsch, Doris Anita Behrens, and Walter Hyll (J Med Internet Res 2026;28:e80274; DOI 10.2196/80274) examining barriers, facilitators, and intention to use AI for breast cancer diagnosis. The authors ran a cross-sectional, embedded mixed-methods survey of 46 Austrian physicians, comparing clinicians who already use AI with those who do not. Quantitative items were based on the Technology Acceptance Model (TAM) and its extensions, open-ended responses were analysed with conventional content analysis, and the two strands were integrated via joint displays. Ordinary least squares regressions identified predictors of attitudes, intention, and perceived likelihood of future use.

Among the 46 participants, about half (50%) reported current AI use. Both users and non-users expressed strong intentions to adopt AI in future, but cited different obstacles: non-users most often pointed to limited access (81%), high costs, and lack of training, while users emphasised weak integration with existing systems and trust concerns. In the regressions, perceiving more facilitators was associated with more favourable attitudes (B=0.83, p=.017), stronger intention (B=1.32, p=.014), and higher perceived likelihood of future use (B=1.56, p=.001). AI-related skills predicted intention (B=1.00) and likelihood (B=1.16), and colleagues' positive views predicted attitudes and intention. Conversely, perceiving more barriers lowered intention (B=-0.84) and likelihood (B=-1.48), and being aged 50 or older was linked to more negative attitudes (B=-1.11, p=.002).

The design pairs a theory-driven TAM instrument with qualitative coding, a common pattern for early-stage implementation research. For practitioners, the value is in mapping abstract acceptance constructs (perceived usefulness, ease of use, social influence) onto concrete workflow, training, and trust concerns. The single-country sample of 46 limits statistical power and external validity, so the estimates are best read as hypothesis-generating rather than definitive population effects.

The authors frame adoption as a staged process: early-stage users may benefit most from improved access and training, while experienced users need support for workflow integration and trust-building. Useful follow-ups include replication in larger, multicentre samples, links between clinician attitudes and objective adoption or patient-outcome metrics, and whether demographic gaps such as the age effect persist across other health systems.