S2313
Physics - Machine learning and AI algorithms
ESTRO 2026
Digital Poster Highlight 3908 Pathways towards AI implementation in Radiotherapy Martijn Vroegindeweij 1,2 , Luca M Heising 1,2 , Carol X.J. Ou 1 , Maria Jacobs 2,3 , Wouter van Elmpt 1,2 1 Information Systems & Operations Management, Tilburg School of Economics and Management, Tilburg, Netherlands. 2 Innovation Implementation, Maastro, Maastricht, Netherlands. 3 Strategy & Entrepreneurship, Tilburg School of Economics and Management, Tilburg, Netherlands Purpose/Objective: Artificial Intelligence (AI) has become one of the most anticipated and promising innovations in Radiotherapy (RT), yet the implementation of these systems remains challenging (Bardhan et al., 2025; Thompson et al., 2018; Zapata et al., 2023). In particular, AI systems introduce a probabilistic logic into a highly protocolized and risk averse clinical environment. This study examined which factors hinder or facilitate successful AI implementation in RT, focusing on how organizational, technical, and contextual elements combine to shape implementation outcomes. Understanding these dynamics is critical as AI systems in RT must expand from experimental pilots to clinical use. By identifying the conditions under which AI adoption succeeds or fails, this study provides actionable insights for radiotherapy centers (RTCs) aiming to integrate AI safely and effectively. Material/Methods: We conducted in-depth interviews across 18 RTCs in the Netherlands, involving radiation oncologist, technicians, physicists, and project managers. We coded these qualitative insights into a fuzzy quantitative set, which enabled us to analyse it with Qualitative Comparative Analysis (QCA), a method that identifies multiple combinations of factors leading to success or failure (Fiss, 2011). Multiple subdimensions were combined resulting in five high order dimensions, which can be found in Figure 1. We analysed these dimensions for two AI processes currently used in RT: auto-segmentation and auto- planning.
Results: Out of 18 RTCs, eleven centres reported successful implementation of auto-segmentation, and seven achieved successful auto-planning implementations. Only configurations supported by two or more cases were used to ensure robust results. Interestingly, the same configuration appeared in both successful implementations, suggesting that there is no clear difference when successfully implementing auto- planning or auto-segmentation. Success was consistently linked with strong AI capabilities (transparent and high performing AI systems), organizational capacity (engaged and trained clinicians), clear governance (alignment between system and implementation strategy), and technical alignment (seamless integration with existing systems). Centres lacking one of these dimensions, faced workflow disruptions, clinician resistance, or vendor dependency issues when implementing AI.
Conclusion: Successful AI implementation in RT requires more than advanced algorithms, it depends on aligning
Made with FlippingBook - Share PDF online