ESTRO 2025 - Abstract Book

S4394

RTT - Treatment planning, OAR and target definitions

ESTRO 2025

Our findings suggest that PBT is a feasible treatment for vulva cancer, with plans that are robust to daily set up variations on set. Given the rarity of this cancer type and the promising results, further investigation is warranted to establish the clinical utility of PBT.

Keywords: Vulva, Protons, Verification

3937

Digital Poster Multi-disciplinary inter- and intra-observer variability for manual and AI-based contouring of thoracic organs at risk. Keeva Moran 1 , Ciaran Malone 2 , Viola Kelly 2 , Edel Smith 1 , Niamh Carroll 1 , Maeve Keys 1 , Pierre Thirion 3 , Sarah Barrett 4 , Lorna Keenan 1 , Nazmy Elbeltaggi 1 , Brendan McClean 1 , Ruth Woods 1 , Gerard G Hanna 4 , Samantha Ryan 1 1 St. Luke's Radiation Oncology Network, St. Luke’s Hospital Beaumont, Dublin, Ireland. 2 St. Luke's Radiation Oncology Network, St. Luke’s Hospital James's, Dublin, Ireland. 3 Radiation Therapy School of Medicine, Trinity College Dublin, Dublin, Ireland. 4 Trinity St. James’s Cancer Institute, Trinity College Dublin, Dublin, Ireland Purpose/Objective: Manual delineation of organs at risk (OARs) in radiation therapy treatment planning is a labour intensive process, subject to inconsistencies arising from inter- and intra-observer variability. The expertise level of the observer may also impact OAR delineation consistency. Artificial intelligence (AI) can reduce contouring times and may increase observer agreement. This study aims to assess the impact of AI-based delineation on multi-disciplinary inter- and intra-observer variability in comparison to manual delineation of thoracic OARs. Material/Methods: Three participants, a novice radiation therapist, experienced radiation therapist, and experienced radiation oncologist, produced manual contours and amended AI-generated contours for 40 randomised structure sets using 10 thoracic CT scans (10 manual and 10 AI-generated sets with 5 of each set duplicated twice to assess intra observer variation). Observers were blinded to duplication and given randomised orders for manual and AI generated contours, with sufficient intervals to minimise memory effects. Assessing inter-observer evaluation, observers contoured 10 CT scans manually and then edited the AI-generated contours (generated using the MVision AI TM ) on the AI CT scan sets. Selected OARs were the heart, oesophagus, and spinal canal, resulting in 120 contours. Volume comparisons were performed across individuals and between manual and AI-based groups. Analysis was conducted using a Python script to generate surface Dice similarity coefficients at a 2 mm threshold, Hausdorff distance and average surface distance metrics. Results: When AI was used as a starting point, the standard deviation in volume differences significantly decreased across all patients and OARs for both inter- and intra-observer variability, with respective p-values of (p<0.001) and (p<0.001). Inter-observer measures, including the surface Dice score at 2 mm, average Hausdorff distance, and average surface distance, showed significant improvements with AI segmentation as a starting point (all p<0.001; see Table 1). Similar improvements for all OARs were observed for the intra-observer analysis, with improved agreement across all OARs and metrics (all p<0.002; see Table 1) using AI-based contours as a starting point. Notably, the maximum deviations both between observers and within each observer significantly decreased.