Session

Overview: This session extensively explores the critical aspects of ensuring the reliability and effectiveness of artificial intelligence (AI) applications in the field of radiotherapy, covering various dimensions of quality assurance. The focus extends on topics other than planning and contouring: synthetic CT generation, clinical use of synthetic CT, patient-specific quality assurance (both offline and online), and quality assurance (QA) of outcome prediction. Addressing the challenges and importance of quality assurance in synthetic CT generation, the session delves into two key aspects: patient-specific QA and machine QA. Patient-specific QA involves ensuring the accuracy of dose calculations, especially in situations where conventional CT scans are unavailable. This will be approached through a review of proposed approaches in the literature, along with exclusion criteria. The second aspect involves ensuring sufficient image quality and addressing concerns like image distortion. Existing recommendations on MR QA are explored, encompassing comprehensive workflow testing specifically designed for MR-only radiotherapy. Additionally, issues related to data transfer and end-to-end verifications will be addressed. The session transitions to the crucial topic of patient-specific QA, covering both offline and online approaches. Strategies for verifying the accuracy of AI-driven patient-specific QA are proposed and discussed. The final segment explores challenges and advancements in using AI models for outcome predictions, along with validation methodologies for predictive algorithms: how to ensure that projected treatment outcomes align with actual patient responses. We delve into the implications of these predictions on treatment planning and decision-making, emphasizing the importance of robust quality assurance measures. In conclusion, the session aims to provide a comprehensive overview of the multifaceted challenges associated with ensuring the quality assurance of AI in the radiotherapy workflow. The central question addressed is: are we ready to use these tools in clinical practice? To what extent?