ESTRO 2025 - Abstract Book

S3448

Physics - Machine learning models and clinical applications

ESTRO 2025

Results: For 86% of the patients, the FAW generated created clinically acceptable results, confirming that it can heavily reduce the workload by reduced handhoffs and considerable time savings. The FAW took on average 9 minutes (range: 8-10 minutes) in comparison to 87 minutes (range: 70-130 minutes) for the standard workflow. Physicians still selected the conventional plans more often (see Tab.1), primarily based on a better target structure assessment. The automatically segmented targets were consistently smaller and occasionally showed an imprecise demarcation between the bladder and prostate.

Conclusion: Automated segmentation solely based on the simulation CT scan is a restriction for accuracy and also limits the approach to anatomically definable targets. However, for almost 50% of the patients, the FAW was equivalent or better. This result led to our decision to use it as the standard workflow with an option to create an adapted machine learning plan based on manually corrected target volumes and OAR, only creating manually optimized plans as a fallback option.

Keywords: Machine Learning, Fully Automated Workflow