ESTRO 2025 - Abstract Book

S2554

Physics - Autosegmentation

ESTRO 2025

4209

Digital Poster An evaluation of rectum contours generated by commercial deep learning contouring software using geometry, dosimetry and predicted toxicity. Owen McLaughlin 1 , Fereshteh Gholami 1 , Stephen J McMahon 1 , Conor K McGarry 2,1 , Suneil Jain 3,1 1 Patrick G. Johnston Centre for Cancer Research, Queen’s University Belfast, Belfast, United Kingdom. 2 Radiotherapy Physics, Northern Ireland Cancer Centre, Belfast Health and Social Care Trust, Belfast, United Kingdom. 3 Department of Clinical Oncology, Northern Ireland Cancer Centre, Belfast Health and Social Care Trust, Belfast, United Kingdom Purpose/Objective: This study assesses rectum contours generated using a commercial deep learning auto-contouring model and compares them to clinician delineated contours using geometry, changes in dosimetry and toxicity modelling. Material/Methods: This retrospective study involved 308 patients who were treated using 3D-conformal radiotherapy between 2004 and 2010. Computed tomography (CT) images were input into Limbus Contour (v1.8.0b3) to generate an auto contour structure set for each patient. Contours were not edited after their generation. Rectum auto-contours were compared to clinician contours using geometry and dosimetry. Dice similarity coefficient (DSC), Hausdorff distance (HD) and volume difference were assessed. Dose-volume histogram (DVH) constraints from the ChiPP trial (V41%, V54%, V68%, V81%, V88%, V95% and V100% [1]) were compared, and a Wilcoxon signed rank test was used to evaluate statistical significance of differences between approaches. Toxicity modelling to compare contours was carried out using equivalent uniform dose (EUD) calculated using DVHs weighted for grade 1 rectal bleeding [2]. Abdominal surgery (n=61) and atrial fibrillation (n=16) were included in multiple logistic regression models with these factors alone comprising a clinical model. Trained models (80%) were tested (20%) in their prediction accuracy of rectal bleeding (ntotal=124) using area-under the receiver operating characteristic curve (AUC) and change in Akaike information criterion (ΔAIC) [3]. Results: Median DSC (interquartile range (IQR)) was 0.85 (0.09), median HD was 1.38 mm (0.60 mm) and median volume difference was 7.53 cc (9.75 cc). Median DVH differences for auto-contours were found to be small (<1.5%) for all constraints though systematically larger than clinician contours (p<0.05). However, a significant spread was seen for individual patients across all dose constraints, with an IQR of up to 8.0% (figure 1).