Toward a human-centric co-design methodology for AI detection of differences between planned and delivered dose in radiotherapy

Introduction: Many artificial intelligence (AI) solutions have been proposed to enhance the radiotherapy (RT) workflow, but limited applications have been implemented to date, suggesting an implementation gap. One contributing factor to this gap is a misalignment between AI systems and their users. To address the AI implementation gap, we propose a human-centric methodology, novel in RT, for an interface design of an AI-driven RT treatment error detection system. Methods: A 5-day design sprint was set up with a multi-disciplinary team of clinical and research staff and a commercial company. In the design sprint, an interface was prototyped to aid medical physicists in catching treatment errors during daily treatment fractions using dose-guided RT (DGRT) with a portal imager. Results: The design sprint resulted in a simulated prototype of an interface supported by all stakeholders. Important features of an interface include the AI certainty metric, explainable AI features, feedback options, and decision aid. The prototype was well-received by expert users. Conclusion/discussion: Using a co-creation strategy, which is a novel approach in RT, we were able to prototype a novel human-interpretable interface to detect RT treatment errors and aid the DGRT workflow. Users showed confidence that the overall design method and the proposed prototype could lead to a viable clinical implementation.