ESTRO 2024 - Abstract Book

S3022

Physics - Autosegmentation

ESTRO 2024

Conclusion:

ACo was able to successfully identify errors to gold-standard and model- or data-driven (epistemic or aleatoric) uncertainty regions as low confidence, for internal and external autosegmentations. False-negatives, which are clinically more concerning, were rare (with the exception of lenses for which this MRI-based model showed poorer performance). False-positives near regions of true-positive error were discounted as they are not clinically relevant in the human-operator QA scenario. Remaining false-positive regions were generally associated with low-contrast regions or artefacts, hence representing genuine regions of uncertainty, despite agreement of AS with gold-standard. ACo derived confidence maps could serve as a model-agnostic, independent, per-patient, quality assurance tool, increasing clinical confidence in autosegmentation, potentially reducing editing time and improving efficiency, whist enhancing patient safety.

Keywords: Confidence, Autosegmentation, uncertainty

References:

1. Claessens M, Oria CS, Brouwer CL, Ziemer BP, Scholey JE, Lin H, et al. Quality Assurance for AI-Based Applications in Radiation Therapy. Semin Radiat Oncol. 2022;32(4):421-31.