ESTRO 2024 - Abstract Book

S2995

Physics - Autosegmentation

ESTRO 2024

The Dice similarity coefficient (DSC) and Hausdorff distance (HD) along with the 95 percentile (HD95) were calculated to compare the model predictions with the original label map. We generated three datasets with 20, 40, and 48 patients to evaluate the relevance of different training cohort sizes.

Results:

The ensemble of 2D and 3D models proved to produce the best results for each training size. For the dataset of 20 patients, the test predictions achieved a DSC of 0.795 ± 0.076 with a HD95 of 3.2. The 40 patient cohort managed 0.808 ± 0.065 and 3.2 respectively and for the cohort of 48 patients, the DSC was 0.818 ± 0.060 with a HD95 of 3.0. Qualitative assessment of the CTV revealed the cranial extent of the CTV (e.g. inclusion of clip material) and the posterior border to the rectum as frequent areas of false segmentations.

Conclusion:

Already with a small cohort of 48 patients, we could reach a DSC which is close to an interrater variability reported in the literature. As the model statistics will likely improve with increased cohort size extending the training set is intended to further optimize the model. However, it still needs to be assessed how the predicted CTVs can be of clinical use and if the network can be included into the workflow seamlessly.

Keywords: Postate, salvage, CTV

References:

(1) Pra AD, Dirix P, Khoo V, et al. ESTRO ACROP guideline on prostate bed delineation for postoperative radiotherapy in prostate cancer. Clin Transl Radiat Oncol. 2023;41. doi:10.1016/j.ctro.2023.100638

202

Digital Poster

Evaluation of foundation models for semi-automated organ delineation in MR images

Krisztian Koos 1 , Eszter Szabo 1 , Muhammad Luqman Hakim 2 , Eszter Katalin Kiss 2 , István Megyeri 1 , Lehel Ferenczi 2 , László Ruskó 2 1 GE HealthCare, Science and Technology Organization, Szeged, Hungary. 2 GE HealthCare, Science and Technology Organization, Budapest, Hungary

Purpose/Objective:

Interactive segmentation models [1-3] have the advantage over the auto-segmentation models that the user can affect the contour with manual prompts. Recently, the Segment Anything Model (SAM) [4] was released that is the