ESTRO 2024 - Abstract Book

S3126

Physics - Autosegmentation

ESTRO 2024

Conclusion:

It was shown how clinically advantageous the use of DL-based models can be in the female and male pelvis RT, outperforming ST and RF algorithms. Their excellent performances in OAR contour correctness and their significantly reduced segmentation times would allow for faster RT workflow and high-quality treatments, possibly automatizing its running in the background. To further validate the capabilities of the DL models, an extension to other districts and a multicenter study are necessary.

Keywords: artificial intelligence, pelvic radiotherapy

2676

Poster Discussion

Auditing clinical usage of OAR deep learning auto-segmentation for every patient

Josh Mason, Jack Doherty, Sarah Robinson, Jack Miskell, Meagan De La Bastide, Ruth McLauchlan

Imperial College Healthcare NHS Trust, Department of Radiation Physics and Radiobiology, London, United Kingdom

Purpose/Objective:

Deep learning auto-segmentation models are typically evaluated in a limited number of patient cases before clinical implementation. Auditing ongoing usage in a busy clinical department with large numbers of independent users is challenging. We investigated the value of automated script based auditing to analyse trends in usage.

Material/Methods:

Deep learning auto-segmentation was implemented for OAR contouring using commercial software for all anatomical clinical sites. Initially the software was evaluated using retrospective data for 35 patients using qualitative scores, geometric and dosimetric measures. Following clinical implementation the software has been used on 750 patients to date. A script compared the deep learning auto-segmentation contour to the final contour approved by the clinician for each OAR in each patient. The surface DICE 1 (3mm tolerance) was calculated for OARs that were edited. The script was run 3 and 6 months after clinical implementation.

Results:

In the initial evaluation, all OARs were scored qualitatively as needing only minor or no adjustment. Some consistent differences with guidelines or local practice were identified. Dosimetric analysis identified a small number of clinically