S1595
Physics - Autosegmentation
ESTRO 2026
Pathology Unit, Sahlgrenska Academy, University of Gothenburg, Gothenburg, Sweden. 3 Region Västra Götaland, Department of Otorhinolaryngology, Head and Neck Surgery, Sahlgrenska University Hospital, Gothenburg, Sweden. 4 Radiation Physics, Department of Hematology, Oncology, and Radiation Physics, Skåne University Hospital, Lund, Sweden. 5 Medical Radiation Physics, Department of Translational Medicine, Lund University, Lund, Sweden. 6 Department of Medical Physics and Biomedical Engineering, Sahlgrenska University Hospital, Gothenburg, Sweden Purpose/Objective: Deep learning-based methods are state-of-the-art for autosegmentation. However, CT image quality may be affected by dental implant-induced artifacts commonly found in patients treated for head-and-neck cancer (HNC). The purpose of this study was to evaluate if the accuracy of autosegmented mastication-related structures is affected by CT artifacts in the training data set. Material/Methods: A total of 170 HNC patient cases treated 2017-2021 at Sahlgrenska University Hospital, Sweden, containing ground-truth contours were available for training and testing. To quantify CT artifacts (Figure 1), the volume of voxels with CT numbers >3000 HU (V3k) was calculated for each case. The cases were then sorted according to V3k, and every fifth case (n=34) was set aside as test set. The remaining 136 cases were used to define the training sets for four autosegmentation models (T1-T4). The training set for T1 was defined by randomly selecting 102 cases, for T2 by selecting the 102 cases with the largest V3k, for T3 by selecting the 102 cases with the smallest V3k and for T4 by selecting all 136 cases.All models (nnU-Net, ver. 2.2.1) were trained on four bilateral structures (masseter, lateral/medial pterygoid muscles, temporomandibular joint) using 3D full-resolution mode with five-fold cross-validation. Hyperparameters were automatically selected by nnU-Net. Model performance was assessed on the test set using geometric metrics (volume difference, the DICE similarity coefficient [DSC], the 95th percentile and the maximum of the Hausdorff distribution [HD95 and HD], and the mean surface distance [MSD]).
Conclusion: This study demonstrates the possibility of using ABAS to enable a simulation-free workflow for OOIs’ auto- segmentation in MR-Linac. Benchmarks of OOIs, including urethras segmented by ABAS on MR-Linac images, were established. However, contour editing by ROs remains necessary. Further investigation is needed to apply the ABAS model to intra-patient OOIs segmentation. Keywords: Autosegmentation, Prostate cancer, MR- Linac
Digital Poster 4137
Deep learning-based autosegmentation of mastication structures in head and neck RT: effects by dental implant-induced artifacts on model performance Allegra Gasparetto 1 , Louise Mövik 1 , Lisa Tuomi 2,3 , Mruga Gurjar 1 , Christian Jamtheim Gustafsson 4,5 , Caterina Finizia 2,3 , Caroline Olsson 1 , Niclas Pettersson 1,6 1 Medical Radiation Sciences, Institute of Clinical Sciences, Sahlgrenska Academy, University of Gothenburg, Gothenburg, Sweden. 2 Institute of Neuroscience and Physiology, Speech and Language
Made with FlippingBook - Share PDF online