ESTRO 2025 - Abstract Book

S2556

Physics - Autosegmentation

ESTRO 2025

Geometrical accuracy of the rectum auto-contours was comparable to recent studies [4]. Although median DVH metrics were similar, variation was seen for individual patients highlighting the importance of clinician review. The accuracy of rectal bleeding prediction did not depend on the contour method for this cohort. The addition of dose accumulation and more clinical factors may reduce noise in the dataset.

Keywords: Auto-contouring, dosimetry, toxicity modelling

References: [1] D. Dearnaley et al, “Conventional versus hypofractionated high-dose intensity-modulated radiotherapy for prostate cancer: 5-year outcomes of the randomised, non-inferiority, phase 3 CHHiP trial,” Lancet Oncol, Aug. 2016. [2] S. L. Gulliford et al, “Parameters for the Lyman Kutcher Burman (LKB) model of Normal Tissue Complication Probability (NTCP) for specific rectal complications observed in clinical practise,” Radiotherapy and Oncology, Mar. 2012. [3] K. P. Burnham, D. R. Anderson, “Multimodel inference: Understanding AIC and BIC in model selection,” Nov. 2004. [4] M. Kawula et al, “Dosimetric impact of deep learning-based CT auto-segmentation on radiation therapy treatment planning for prostate cancer,” Radiation Oncology, Dec. 2022.

4266

Digital Poster Conditioned patient-specific auto-segmentation in MRI-guided radiotherapy for pancreatic cancer: a substitute for image registration Mehdi Shojaei 1 , Björn Eiben 1 , Jamie R McClelland 2 , Simeon Nill 1 , Alex Dunlop 1 , Arabella Hunt 3 , Brian Ng-Cheng-Hin 3 , Uwe Oelfke 1 1 Joint department of physics, The Institute of Cancer Research and Royal Marsden NHS foundation trust, London, United Kingdom. 2 Department of medical physics and biomedical engineering, UCL Hawkes institute and University College London, London, United Kingdom. 3 Radiotherapy department, The Royal Marsden NHS foundation trust, London, United Kingdom Purpose/Objective: Pancreatic cancer poses challenges for radiotherapy due to anatomical changes and proximity of critical organs-at risk (OARs). MR-Linacs enable adaptive treatment by capturing daily images [1] and using image registration to propagate reference contours (Prop-ROIs), followed by manual adjustments within a two-centimeter margin of planning target volume (PTV+2cm), ideally in 20 minutes. However, image registration often introduces errors, and severe anatomical changes can prolong the process, highlighting the need for more accurate and efficient solutions. Traditional deep-learning models could improve efficiency but rely on large, high-quality datasets, which are scarce in MR-Linac pancreatic cancer [2]. To overcome these limitations, we propose CondPSeg, a conditioned patient specific segmentation framework with a novel structure-guided deformation-based augmentation pipeline (sgDefAug) that simulates plausible daily anatomical variations to streamline workflows and enhance accuracy. Material/Methods: Four patients, each with five session images, were included. OARs (small bowel, duodenum, kidneys, large bowel, liver, spinal canal, spleen, and stomach) were manually contoured and verified by two clinicians. After preprocessing, sgDefAug was applied for data augmentation, generating 20 variants of first session image by simulating subsequent session images through anatomically plausible OAR scaling and displacement (each axis ≤40mm). Then nnU-Net was trained on these images with a conditioned approach [3]. Each input/output was

where I n and L n were image and label for session n , and I m was image for session m with

model predicting A separate model was trained for each of the 4 patients. Evaluation was performed on the