The Global Reading Room: Purchasing a Radiology Artificial Intelligence System

Your radiology group has decided to purchase and clinically implement a radiology-focused artificial intelligence (AI) algorithm. The group has nominated you to lead the effort. How do you approach this process?

Merel Huisman

Radboud University Medical Center, Nijmegen, The Netherlands [email protected]

I would choose a goal-oriented methodic approach. I would assemble a small multidisciplinary team and define clinical needs and predefined targets to measure success, focusing on workload reduction without quality compromises. Budget-wise, the tool's total costs must not surpass radiologist task expenses.

Next, the team would evaluate the available tools, ensuring they are evidence-based and comply with current and upcoming regulations (e.g., the European Union AI Act). After narrowing down to two top solutions, structured vendor interviews would be conducted to discuss technical compatibility, workflow integration, scalability, and service and postmarket surveillance and to explore price models versus our expected volume. Additionally, we would assess potential for scientific collaboration and obtain quotes, accounting for technical debt. The decision about which tool to purchase would be made by the team, ensuring that all aspects are considered and that the best fit for our clinical needs is selected. After training end users, a shadow deployment period would follow, emphasizing safety and uniform adoption. After 6 months, we would conduct a quantitative and qualitative evaluation to determine whether we have met our predefined targets. This evaluation would inform any necessary adjustments or confirm the decision for long-term implementation, ensuring the AI tool improves our radiology services.

Felipe C. Kitamura

Dasa, São Paulo, Brazil [email protected]

My organization's experience in integrating radiology-focused AI was a daunting yet exciting task. Rather than rushing to procure AI solutions, the organization tried to identify the nuanced issues that it could potentially address for the local practice, a process riddled with uncertainty. Engaging with stakeholders proved to be a Herculean feat, navigating through diverse opinions from the information technology (IT) and legal departments, senior management, and frontline physicians, each with their own priorities and concerns. The arduous task of vendor evaluation demanded meticulous scrutiny; yet, amid budget constraints and conflicting features, perfection remained elusive. Despite best efforts, validating the chosen AI product using the institution's data faced stumbling blocks, highlighting the complexities of real-world information sources. Collaborating with radiologists to seamlessly integrate AI into existing workflows was a balancing act, often met with resistance and skepticism. Determining how to distribute benefits among stakeholders and manage associated costs felt like a tightrope walk, fraught with uncertainty. Establishing key performance indicators for ongoing monitoring of AI's impact proved challenging but crucial to sustain long-term buy-in. Through this journey, I learned that people management, perseverance, and adaptability are the true hallmarks of progress in enhancing patient care and radiology practice.

John Mongan

University of California San Francisco, San Francisco, USA [email protected]

I would begin by calling on (or creating) a multidisciplinary AI governance team, consisting of representatives of clinical users of the AI, IT services, informatics, and (if available) AI experts. The first task would be to determine whether the proposed AI effectively solves one of the group's specific clinical or business problems. If not, then no purchase should be made. If so, then the team evaluates whether the benefits provided by the AI exceed the risks and costs that it incurs. Benefit considerations include the percentage of cases to which the AI will apply, the magnitude of impact on each affected case, and the effect on radiologist efficiency. Risk and cost considerations include probability, detectability, and correctability of AI errors; negative impact on radiologist performance; creation or worsening of health care disparities; and total implementation and maintenance costs (not just licensing cost). If the ratio of benefit to cost and risk is favorable, then the team designs the implementation and integration with existing systems to minimize risk and maximize efficiency. A crucial part of this process is development of an AI performance–monitoring plan that includes performance thresholds below which the AI will be deactivated.

Masahiro Yanagawa

Osaka University Graduate School of Medicine, Osaka, Japan [email protected]

The social implementation of AI in radiology groups requires careful consideration of its functions, accuracy, scope, usability, and cost-effectiveness. Although AI can reduce physician workload and improve diagnostic quality, seamless integration into clinical practice is required. It is critical to view AI as a tool to assist physicians, not replace them, and to prioritize functionality over complexity. For example, specialized AI for lesion detection may be more beneficial than multifunctional systems in screening facilities. As AI becomes more integrated, legal liabilities may arise from discrepancies between physician judgments and AI results. Explainable AI is essential for qualitative diagnosis, providing transparency in the diagnostic process. Although saliency maps provide insight, clinicians must understand their limitations. The term “human in the loop” refers to a design or operational approach in which human judgment, intervention, or oversight is integrated into an automated system or process ¹. This concept is commonly used in multiple fields, including AI and automation. Adopting a human-in-the-loop approach recognizes the strengths of both AI and human judgment, maximizing benefits while addressing limitations. Improving AI literacy through guidelines and seminars is crucial to AI's effective and ethical use and promotes the correct use of AI alongside medical education.

Acknowledgment

J. Mongan thanks Marc Kohli (Department of Radiology and Biomedical Imaging, University of California San Francisco) for his contributions in codeveloping the governance process described in his response.

Notas

Footnote Provenance and review: Solicited; not externally peer reviewed.