ESTRO 2025 - Abstract Book

S2422

Physics - Autosegmentation

ESTRO 2025

613

Digital Poster Assessing the ETHOS contouring efficacy --- An AI study in the wild.

Frank Van den Heuvel 1,2 , Lia Vugts 1 , Fiere Claassen 1 , Patricia Brouwers 3 , Veronique Coen 3 , Marielle van Hinsberg 1 , Inge Jacobs 3 , Isabelle de Marco 4 , Astrid Poland 4 , Doetsje de Vries 4 , Barbara Wachters 3 , Lockie Yuen 1 , Cees Griep 3 1 Medical Physics, Zuidwest Radiotherapeutisch Instituut, Vlissingen, Netherlands. 2 Oncology, Unversity of Oxford, Oxford, United Kingdom. 3 Clinicians, Zuidwest Radiotherapeutisch Instituut, Vlissingen, Netherlands. 4 RTT, Zuidwest Radiotherapeutisch Instituut, Vlissingen, Netherlands Purpose/Objective: On-line Adaptive Radiation Therapy (OART) has become clinically viable, facilitated by vendors such as ETHOS TM (Varian Inc), MRIdian TM (ViewRay Inc), and Radixact TM (Accuray Inc). ETHOS TM , in particular, relies on artificial intelligence (AI) for segmentation, yet its implementation remains opaque, with limited literature backing its efficacy. This study aims to quantify ETHOS's performance and assess its clinical validity. Material/Methods: We analyzed 21 consecutive prostate carcinoma patients treated with a 9-field IMRT plan delivering 60 Gy in 20 fractions. At each fraction's start, contours for the prostate, seminal vesicles(SV), bladder, and rectum were created and approved before importing the existing planning contours. Radiotherapists then adapted these contours with clinician and physicist oversight to optimize the treatment plan, resulting in 5,000 to 6,000 images and files per patient. We produced two contour sets (AI-generated and clinically used) for analysis. Quantifying the similarity using a Dice Similarity Coefficient (DSC), pixelised Hausdorff distance, and volume ratios. Additionally, the pre treatment verification CBCT was used to estimate the time needed for the procedure. Results: Figure 1a displays the DSC for a single representative patient, indicating slight differences in bladder and prostate values, but larger discrepancies for the SV and rectum. Figure 1b shows the volume ratios of clinical to planned contours, revealing that while the prostate is consistently well represented, SVs exhibit significant variability, underscoring the need for an adaptive approach.