“Lost in Translation”: AI Ethics Frameworks and the Reality of Deployment in Radiology Across Health Systems

Abstract

Ethical frameworks for artificial intelligence in healthcare have proliferated rapidly, yet evidence on how such frameworks operate in practice, particularly across diverse health-system contexts, remains scarce. This thesis examined how a leading multisociety framework for “developing, purchasing, implementing, and monitoring” radiology AI, authored by the American College of Radiology, Canadian Association of Radiologists, European Society of Radiology, Royal Australian and New Zealand College of Radiologists, and Radiological Society of North America, actually operated across twenty‑two deployments in fifteen countries. The cases ranged from university hospitals in Europe and North America to tuberculosis screening programmes, clinics, and teleradiology platforms in India, Kenya, Mozambique, Papua New Guinea, and other low‑ and middle‑income settings. Using document analysis, semi-structured interviews with twelve expert respondents, and a six-domain evaluative rubric derived from the framework, the study assessed the extent to which ethical and practical standards were implemented, adapted, or bypassed across different health-system contexts. We found that framework compliance was systematically different by health-system capacity. In high-income settings with mature digital infrastructure, radiologist-led services, and established regulatory and reimbursement pathways, core framework requirements around problem definition, evidence validation, workflow integration, and governance were broadly achievable; in several cases, the framework correctly supported decisions not to adopt AI at all. In low- and middle-income settings, the same tools were deployed to create basic diagnostic capacity where none previously existed, and foundational framework assumptions such as feasible local validation, stable institutional financing, on-site radiologist oversight, and hospital-centred procurement, were routinely unmet. Governance functions had migrated to donors, AI marketplaces, and teleradiology platforms whose accountability structures the framework does not evaluate. Across both contexts, post-deployment monitoring remained aspirational, and financing emerged as the single most consequential variable shaping implementation, one the framework does not address. These findings indicate that the multisociety framework while detailed lacks actual implementation in resource constrained environments. Four adaptations are proposed: an explicit preconditions domain specifying the minimum institutional conditions under which other safeguards are realistic; treatment of financing and business models as a core governance domain with explicit sustainability requirements; recognition of intermediaries as governance actors whose practices must themselves be evaluated; and disaggregation of human-in-the-loop requirements alongside elevation of equity from a dataset property to an organising purpose. Alongside the framework adaptations, practical steps for LMICs are also presented. This study attempts to conduct a systematic review of a major radiology AI governance framework across both high- and low-income health systems, and advances an empirically grounded basis for adapting global AI governance standards to the settings where diagnostic AI is most urgently needed.