S1567
Physics - Autosegmentation
ESTRO 2026
Keywords: Pancreatic Cancer, Gross Tumour Volume, MRI-Linac
CBCT-, sCT-, and CT-based models, respectively, with the tubular cauda structure showing the lowest performance across all models. Surface-based metrics, like NSD, demonstrated lower variability, with mean values above 0.85 for all models. In terms of DSC, no significant difference was observed between CBCT- and sCT-based segmentation (p=0.50), while the CT- based baseline performed significantly better than both CBCT- and sCT-based models (p=0.01).
Digital Poster Highlight 2234
Training CBCT auto-segmentation models without segmented CBCT data using cyclic image synthesis Adrian Thummerer 1 , Dinah Konnerth 1 , Stefanie Corradini 1,2 , Claus Belka 1,3 , Guillaume Landry 1,3 , Christopher Kurz 1 1 Department of Radiation Oncology, LMU University Hospital, Munich, Germany. 2 Now at Department of Radiation Oncology, Universitätsklinikum Erlangen, Erlangen, Germany. 3 Bavarian Cancer Research Center, (BZKF), Munich, Germany Purpose/Objective: Segmentation of daily cone-beam CT (CBCT) images is essential for adaptive radiotherapy. However, due to the limited availability of accurately segmented CBCT datasets, deep learning-based auto-segmentation on CBCTs remains far less developed than CT-based auto- segmentation.To address this limitation, we propose using a cyclic image synthesis model to generate synthetic CBCTs (sCBCTs) from planning CTs (pCTs) for training CBCT-based segmentation models using existing pCT contours, or alternatively, to apply a CT- based segmentation model to synthetic CTs (sCTs) generated from CBCTs using the same image synthesis model. Material/Methods: Data from 333 pelvic cancer patients (300 training, 23 validation, 10 testing) were used to train and evaluate an unsupervised 3D CycleCUT image synthesis model [1] for generating sCBCTs and sCTs. Based on this dataset, nnUNet segmentation models [2] were trained using either (i) pairs of sCBCTs with pCT contours (nnUNetCBCT) and applied to real CBCTs, or (ii) trained using pCTs with corresponding contours (nnUNetCT) and applied to CBCT-derived sCTs (see Figure 1). Ground truth CBCT segmentation for the test set were delineated by an experienced radiation oncologist. As a baseline, nnUNetCT models from (ii) were applied directly to pCTs of the test patients. Segmentation performance for bladder, rectum, cauda equina and femoral heads was evaluatd using the Dice similarity coefficient (DSC), 95th percentile Hausdorff distance (HD95), and normalized surface distance (NSD, 3 mm threshold). Statistical differences in DSC among CBCT-, sCT- and CT-based segmentations were assessed using the Friedmann test followed by a post- hoc Nemenyi test (p<0.05). Results: Figure 2 presents boxplots of DSC, HD95 and NSD. The overlap-based DSC metric showed mean values ranging from 0.62–0.91, 0.68–0.89, and 0.58–0.96 for
Made with FlippingBook - Share PDF online