China issues ethics guidelines for AI medical imaging research

China’s national science and technology ethics committee has published ethical guidelines governing artificial intelligence in medical imaging research, the Ministry of Science and Technology confirmed on 29 August. The guidelines, issued by the committee’s medical ethics subcommittee, are intended to govern algorithm development, clinical validation and deployment of AI systems used to analyse scans such as CT, MRI, PET and ultrasound.

AI medical imaging tools are increasingly used to support diagnosis, treatment planning and shared decision-making between doctors and patients. The guidelines note that such systems handle large volumes of sensitive health data and carry risks including privacy breaches, algorithmic bias, and unclear lines of accountability when AI-assisted outputs contribute to clinical decisions.

The framework sets out six guiding principles: enhancing human welfare, promoting fairness, respecting patient autonomy, protecting privacy and data security, ensuring safety and controllability, and strengthening transparency. Patients’ health interests are to be prioritised throughout, and AI is to remain in a supporting role rather than a decision-making one.

Informed consent is a central requirement. Researchers must obtain written consent from study participants, with provision for dynamic consent where appropriate, and participants retain the right to withdraw at any time. Retrospective imaging data that has already been anonymised and consented to may be exempted from this requirement, subject to ethics committee review.

On accountability, the guidelines state that humans remain the final responsible party for AI system design, deployment and use. Diagnostic and treatment decisions must be made by clinicians drawing on professional judgement and the patient’s specific circumstances, with doctors retaining ultimate control and responsibility for medical decisions.

The guidelines also address data governance in detail, covering data security across the full lifecycle of collection, transmission, storage, analysis, sharing and disposal, as well as requirements around data quality, annotation standards, and de-identification of sensitive health and biometric information. Separate provisions cover synthetic data, requiring that its use be clearly disclosed and not conflated with real patient data or used to inflate reported model performance.

On algorithms, the guidelines call for adequate standards of sensitivity, specificity and precision, alongside explainability measures that allow clinicians and regulators to understand and trust model outputs. Researchers are directed to test for reproducibility across devices, institutions and patient populations, and to actively identify and correct bias arising from imbalanced training data.

The guidelines further prohibit exaggerated or misleading claims about model performance, and call for ongoing ethics training for researchers working in the field.