Abstract
This article examines the role of uncertainties in autosegmentation for radiotherapy, which has shown promise in improving consistency, accuracy, and efficiency over traditional manual delineation. Autosegmentation, however, introduces challenges due to uncertainties that arise from factors such as limited training data and models. A nuanced understanding of these uncertainties is essential, as they directly impact clinical decisions and patient outcomes. We explore the sources of uncertainties and how they may be quantified, highlighting the impact on different levels of the segmentation process and discussing the implications of imperfect predictions. Practical applications are proposed, such as uncertainty maps to guide manual adjustments and flagging mechanisms to support quality assurance. We finally address the limitations of these approaches, including computational burdens and risks of information overload. Meaningful uncertainty quantification in autosegmentation holds significant potential to enhance clinical workflows, build trust in artificial intelligence-based tools, and ultimately improve patient care in radiotherapy.