Scholarlink

Machine learning for medical imaging (ML4MI) is increasingly adopted in hospitals to support diagnostic and workflow processes, yet aligning such systems with clinical practice, governance constraints, and regulation remains challenging in practice. This paper reports an empirical study of how requirements engineering (RE) is currently performed in ML4MI projects. Based on ten semi-structured interviews with stakeholders from hospitals, research, and industry in Denmark, we conducted a grounded-theory–inspired thematic analysis to examine how requirement-related decisions are made across development, validation, deployment, and maintenance. Our findings show that RE is rarely expressed through formal specifications; instead, it is embedded in iterative experimentation, stakeholder collaboration, and institutionally constrained validation and monitoring activities. We identify recurring tensions related to data access, accountability for system updates, and the interpretation of non-functional requirements such as trust, explainability, fairness, and usability. The paper discusses implications for making existing RE practices more explicit in order to support traceability and accountability in hospital-based AI systems.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Requirements Engineering Challenges in Developing Machine Learning Systems for Medical Imaging in Hospitals

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Amanda Garde Vallentin,
  • Natascha Rylander Bech,
  • Elda Paja

摘要

Machine learning for medical imaging (ML4MI) is increasingly adopted in hospitals to support diagnostic and workflow processes, yet aligning such systems with clinical practice, governance constraints, and regulation remains challenging in practice. This paper reports an empirical study of how requirements engineering (RE) is currently performed in ML4MI projects. Based on ten semi-structured interviews with stakeholders from hospitals, research, and industry in Denmark, we conducted a grounded-theory–inspired thematic analysis to examine how requirement-related decisions are made across development, validation, deployment, and maintenance. Our findings show that RE is rarely expressed through formal specifications; instead, it is embedded in iterative experimentation, stakeholder collaboration, and institutionally constrained validation and monitoring activities. We identify recurring tensions related to data access, accountability for system updates, and the interpretation of non-functional requirements such as trust, explainability, fairness, and usability. The paper discusses implications for making existing RE practices more explicit in order to support traceability and accountability in hospital-based AI systems.