ESTRO 2025 - Abstract Book

S3388

Physics - Machine learning models and clinical applications

ESTRO 2025

The reliability diagrams showed a mostly linear relationship between uncertainty estimation and prediction error (figure2a). The MVE1 approach achieved the best ENCE, followed by DE and standard MCD. Confidence curves decreased for all approaches when voxels with high uncertainty were discarded. The rise near 100% discarded voxels indicated a few confident but faulty predictions, possibly due to pixelation errors at the patient surface. Again, MVE1 showed the most favorable behavior (figure2b).