| Data Privacy |
Risk of leakage of sensitive patient data (ultrasound images, diagnostic records) during AI model training/operation. |
Adopt federated averaging to avoid cross-center data transmission; Implement end-to-end encryption (GB/T 35273–2023) and access control for medical data; - Use edge computing for local model updates without cloud data upload.
|
Xu et al. (2025), Schweihoff et al. (2021), and Weaver (2016)
|
| Allocation of Responsibility |
Ambiguity in liability for adverse outcomes (e.g., needle deviation) caused by AI algorithm errors or improper operator use. |
Enhance algorithm transparency (disclose model training data sources and decision logic); Establish a dual accountability mechanism: physicians for operational compliance, developers for algorithm safety; - Embed real-time error logging to trace failure causes (e.g., AI drift vs. operator misoperation).
|
Andriollo et al. (2024), Kuratani et al. (2022), and Weaver (2016)
|
| Patient Informed Consent |
Difficulty for patients to understand complex AI technology, leading to inadequate informed consent. |
Use simplified visual aids (e.g., AI workflow diagrams) and case examples to explain AI’s role; Develop standardized consent forms outlining AI limitations, potential risks, and alternative treatments; - Provide verbal explanations in non-technical language to ensure comprehension.
|
Weaver (2016)
|
| Algorithmic Bias |
Continuous learning algorithms may exacerbate structural biases (e.g., underrepresentation of obese patients) in training data. |
Integrate diverse training datasets (including special populations like BMI > 35 patients); Conduct periodic bias audits of AI models (e.g., quarterly assessment of segmentation accuracy across demographics); - Involve ethicists and patient representatives in algorithm design reviews.
|
Diplock et al. (2023), Xu et al. (2025), and Wu et al. (2025)
|
