Beyond the algorithm: embedding ethics for trustworthy AI in radiology and oncology

Artificial intelligence (AI) in radiology and oncology promises improvements in diagnostic accuracy and efficiency yet introduces complex ethical and societal challenges. Governance efforts frequently rely on high-level principles such as trustworthiness and fairness, which risk becoming ineffective when not grounded in specific contexts. This study presents findings from our work on ethical and societal aspects of AI within the EuCanImage project.

We conducted a multi-method empirical study involving literature reviews, interviews, and workshops with developers, clinicians, and other stakeholders. The study explored how ethical concerns emerge in real-world settings and how they are shaped by institutional, clinical, and sociotechnical dynamics.

Findings indicate that ongoing interdisciplinary involvement is essential to address explainability, accountability, bias, and social impact in radiological AI. The literature review identified four guiding dimensions of trustworthy AI (i.e., explainability and interpretability, trust and trustworthiness, responsibility and accountability, and justice and fairness) which remain difficult to operationalize without concrete procedural guidance. Empirical findings highlight that ethical issues cannot be addressed solely as technical problems or abstract principles. Trustworthiness emerged as relational and co-constructed through interactions among very diverse stakeholders.

We propose a structured, multi-stakeholder AI development pathway that advances from decontextualized, principle-driven ethics toward embedded, interdisciplinary approaches attentive to clinical realities, power relations, and socio-cultural conditions, by strengthening stakeholder engagement for trustworthy AI in cancer care.