China Issues First Ethical Guidelines for Artificial Intelligence Medical Imaging Research

On August 27, 2026, the Medical Ethics Subcommittee of the National Committee on Science and Technology Ethics, through the Ministry of Science and Technology, published the “Ethical Guidelines for Artificial Intelligence Medical Imaging Research.”

Rather than establishing a product-registration framework specifically for AI medical imaging, the guidelines provide an ethical governance framework for the responsible development and application of these technologies throughout their lifecycle.

The guidelines establish six fundamental principles:

• Promoting human well-being
• Promoting fairness and justice
• Respecting human autonomy
• Protecting privacy and data security
• Ensuring safety and controllability
• Strengthening transparency and trustworthiness

Data Governance: Building an Ethical Foundation for AI Development

The guidelines establish requirements for controlling data throughout its lifecycle, while also addressing data quality, personal-information protection, consent, and the use of synthetic data.

Rigorous controls must apply across every stage of the data lifecycle, from collection and authorization to transmission, storage, analysis, sharing, and destruction.

Data-use agreements should clearly address:

• Intended uses
• Permissions
• Access controls
• Encryption
• Backup and recovery
• Monitoring and auditing

Data use should remain traceable and verifiable throughout the lifecycle.

Data Quality and Annotation

Researchers should ensure the professionalism and accuracy of data annotation and establish consistent standards for:

• Annotator qualifications
• Image findings
• Image annotation
• Image segmentation
• Image quantification

Research teams should also assess the provenance and scientific validity of datasets, including:

• Retrospective versus prospective data sources
• Source distribution
• Potential bias
• Data structure
• Annotation quality

These considerations are important not only for data integrity, but also for the reliability and generalizability of the resulting AI model.

De-Identification of Medical Imaging Data

When medical imaging data containing sensitive personal information, including health or biometric information, must be shared for scientific research or other permitted purposes, appropriate de-identification measures should be applied.

The processed information should not permit the direct or indirect identification of a specific individual and should not be capable of being restored to identifiable information.

Consent and Authorization

The guidelines require written informed consent for AI medical imaging research, with the informed-consent documentation and process subject to ethics committee approval.

Dynamic informed consent may be used when necessary, and research participants should be permitted to withdraw their consent.

An important exception applies to retrospective medical imaging data. Where data providers have already signed informed-consent documents and the data have been anonymized, informed consent may be waived if an ethics committee confirms that the research meets the applicable requirements of China’s rules governing ethics review of life-science and medical research involving humans.

Requirements for Synthetic Medical Imaging Data

The guidelines specifically address synthetic medical imaging data generated through mathematical models, algorithms, or generative AI.

Researchers using synthetic data for training or data augmentation should recognize risks such as:

• Spurious correlations
• Amplification of bias
• Misleading model-performance results

Synthetic data must be clearly identified as such, and the generation process, applicable conditions, and potential risks should be disclosed.

Synthetic data must not be presented as real patient data without disclosure. Researchers must also not deliberately use synthetic data to exaggerate model performance or circumvent ethics review.

2. Algorithm Development and Validation: Demonstrating Reliable Performance

Once appropriate data governance is established, the guidelines place significant emphasis on the safety, explainability, reproducibility, and stability of AI algorithms.

Safety, Explainability, and Performance

The guidelines call for systematic assessment of:

• Algorithm safety
• Explainability
• Reproducibility
• Performance stability

Algorithm performance — including measures such as sensitivity, specificity, and precision — should meet appropriate standards for the intended medical research application.

For deep-learning systems that may function as “black boxes,” researchers should provide an explanation of algorithm interpretability so that physicians and regulatory authorities can understand and have confidence in model outputs.

Reproducibility Across Clinical Environments

The guidance places particular emphasis on demonstrating that algorithm performance is reproducible beyond the environment in which the model was originally developed.

Algorithms should undergo reproducibility testing, with the conditions, specific context, and necessary methods for reproducing results documented.

Validation should consider variation across:

• Different imaging devices and equipment
• Different medical institutions and clinical environments
• Different patient populations

The guidelines further encourage a progression from internal validation and external validation to prospective clinical trials and multicenter real-world research.

The objective is not simply to demonstrate strong technical performance, but to establish that the model remains reliable in actual clinical settings and provides measurable health benefits to patients.

Identifying and Correcting Bias

Researchers should proactively identify potential bias throughout:

• Data collection
• Annotation
• Model training
• Model evaluation

Research teams should establish quantitative bias indicators and monitoring mechanisms and update algorithms or correct bias when necessary.

The objective is to prevent biases in data or algorithms from becoming discriminatory outcomes for people affected by the technology.

3. Clinical Use: Keeping Humans Accountable

The guidelines establish human decision-making as a central principle for AI medical imaging.

AI should occupy an assistive role in decision-making, while humans retain ultimate decision-making authority.

Physicians Retain Final Decision-Making Authority

Final diagnoses and treatment plans must be determined by physicians based on their professional judgment and the individual patient’s circumstances.

The guidelines state that humans must retain ultimate control and responsibility for medical actions.

Patients Should Understand AI’s Role

Patients should also have the right to understand the role of AI medical imaging technology in their diagnosis or treatment, as well as its potential risks and limitations.

This reinforces transparency not only as a technical requirement, but also as part of the patient-facing use of AI in clinical care.

Clear Allocation of Responsibilities

The guidelines call for a collaborative responsibility mechanism covering the full lifecycle of AI medical imaging research.

Researchers, medical institutions, clinicians, technology companies, and relevant regulatory authorities should have clearly defined responsibilities.

Research results and processes should be traceable, auditable, and accountable, covering activities from technology design and development through deployment and application.

The guidelines also require disclosure and management of potential conflicts of interest that could affect the objectivity or fairness of research.

Research teams should establish mechanisms to identify, assess, and mitigate such conflicts across:

• Study design
• Data collection
• Analysis
• Technology translation.

Ethics Review Extends Beyond Initial Approval

The guidelines give ethics committees a continuing oversight role rather than limiting review to the initial research approval.

Ethics review should consider:

• Compliance with applicable laws, regulations, ethical principles, and professional standards
• The scientific and social value of the research
• Whether the research addresses meaningful clinical or public-health needs
• The scientific validity of the study design and sample size
• Transparency, explainability, and responsibility traceability
• The completeness of research records, including reliability testing and validation
• Model generalization across different equipment and clinical environments
• Data provenance, authorization, access, and sharing controls

The guidelines specifically call for follow-up ethics review because of the risks and uncertainties associated with AI medical imaging research. After deployment, review should monitor issues including model performance drift, warning signals, and new problems arising following model updates or iterations.

This means that ethical oversight is intended to continue as an AI system evolves and moves into real-world use.

Action Plan for Overseas Manufacturers

International medical device and software manufacturers planning or conducting AI validation in China should align their existing R&D, data governance, and quality management systems with these requirements through five practical steps:

1. Audit Existing AI R&D and Data Pipelines

Assess current projects and clinical research against the new requirements, focusing on data governance, participant authorization, traceability, de-identification, algorithm validation, bias monitoring, and ethics review.

2. Strengthen Data Governance Controls

Review procedures for consent and authorization management, data provenance, access controls, and data retention and destruction.

3. Build Cross-Environment Validation into Development Plans

Account for variation in imaging equipment, clinical sites, and patient populations when designing AI medical device clinical validation strategies. Establish performance criteria and identify potential sources of variation before clinical or regulatory submission.

4. Formalize Bias and Performance Monitoring

Identify potential sources of bias throughout data collection, annotation, training, and evaluation. Establish quantitative indicators and monitoring mechanisms and define procedures for corrective action.

For deployed systems, incorporate monitoring for performance drift and issues associated with subsequent model updates.

5. Connect Ethics Documentation with Technical and Clinical Evidence

Maintain clear records linking ethics approvals and reviews with data authorization, data-management records, algorithm development, reliability and validation testing, and clinical research documentation.

This can help demonstrate traceability and accountability throughout the research and translation process.

What the New Guidelines Mean for China AI Medical Device Development

The new guidance reinforces a lifecycle approach to ethical governance of AI medical imaging, extending from data acquisition and algorithm development through validation, clinical translation, deployment, and subsequent monitoring.

By incorporating data governance, informed consent, de-identification, algorithm explainability, cross-environment validation, bias management, human oversight, and post-deployment monitoring into development plans at an early stage, manufacturers can reduce potential evidence and documentation gaps by incorporating.

Navigating NMPA expectations requires precision. Cisema guides global life science innovators through every phase of Chinese regulatory compliance, handling compliance audits, Ethics Committee submissions, clinical research strategy, and NMPA registration.

To assess how China’s latest AI medical imaging requirements may affect your R&D, clinical validation, and NMPA registration strategy, contact Cisema today.

Further Information

Explore Cisema's services for medical device registration in China.

References

The "Ethical Guidelines for Artificial Intelligence Medical Imaging Research" were released《人工智能医学影像研究伦理指引》发布