ESTRO 2024 - Abstract Book

S4487

Physics - Machine learning models and clinical applications

ESTRO 2024

Ten core challenges were identified associated with the DGRT workflow: reducing workload, optimizing workflow, decision making, presentation of AI information, EPID measurement accuracy, AI verification, ideal preparation, understanding of treatment errors, increasing confidence, and identifying trends in treatment errors. As our goal aimed at presenting human-interpretable information, we focused on presentation of information to narrow the scope of the solution. In the resulting DGRT.AI interface, users are provided with a list of all fractions that were flagged by the software. General information about the patient is limited to a minimum of relevant information, e.g., the treatment fraction, the AI-identified error, and the AI certainty about the identified error. By clicking on the fraction, information relevant to the identified treatment error is provided. In the example of tumor regression, the most relevant information needed to confirm the AI error classification is anatomical information, so the planning-CT and CBCT scans are shown. If desired, the user can find additional information, such as dose information, gamma analysis, dose recalculation based on log files, or a 3-dimensional rendering of the dose. The MP can use this information to decide what the most suitable next step is with regard to the identified error. The most novel and important features of DGRT.AI include: the AI-based error detection, a decision making aid that follows from the AI; and an AI certainty metric, the option to evaluate the AI classification using explainable AI, including an option to provide feedback in case of an incorrect error identification to build a large, high quality database for further AI model improvements (Figure 2).