ESTRO 2024 - Abstract Book

S3080

Physics - Autosegmentation

ESTRO 2024

Conclusion:

This work provides the first, to our knowledge, multi-institute external and internal validation of a 3D Unet trained on a large single institute cohort for the left and right breast CTV auto-segmentation. The metrics used for comparison showed an overall agreement and model transferability for all centers, except Institute 9. Institutes 10, 7, 1 showed a larger variability in some metrics, though globally comparable with the internal test set. In general, predicted contours are slightly larger than clinical contours, except for Institute 10. These achievements have paved the way for generating a Benchmark 3D Unet that includes datasets from transferable institutes with potentially application in automatic segmentation, possibly driving clinician during contouring phase.

Keywords: ai, segmentation, external validation