S1596
Physics - Autosegmentation
ESTRO 2026
Conclusion: The accuracy of deep learning-based
autosegmentation of mastication-related structures in HNC was little affected by the presence of different levels of dental implant-induced CT artifacts in the training datasets. Keywords: mastication structures, artifacts Digital Poster 4196 Evaluation of an MRI-based auto contouring prototype for OAR and target delineation in brain Nazanin Rahnama 1 , Stephanie Tanadini-Lang 1 , Riccardo Dal Bello 1 , Vaisakh Nappady Joy 2 , Alina Elter 2 , Matthias Guckenberger 1 , Nicolaus Andratschke 1 1 Radiation Oncology, University Hospital Zurich, Zurich, Switzerland. 2 Cancer Therapy Imaging, Siemens Healthineers AG, Forchheim, Germany Purpose/Objective: Accurate and reproducible target (GTV) and organ-at- risk (OAR) delineation is critical for high-quality radiotherapy planning. While manual delineation of Targets on CT images remains the clinical standard, OARs are increasingly auto segmented by auto contouring tools, in the recent era. At the same time, MRI offers superior soft-tissue contrast that could improve delineation accuracy and reduce inter- observer variability. In this technical note, we present the results of a multi-disciplinary review of contour quality and potential clinical applicability of a prototype MRI-based auto contouring tool. Material/Methods: 27 patients were randomly selected from our institutional database that included previously treated 240 brain metastases. The AI-based prototype auto contouring tool (Siemens Healthineers) was applied to post-contrast 3D T1-weighted MRI to generate automatic contours for both OARs and GTVs. The clinical usability of each auto-generated contour was assessed by an expert radiation oncologist. A 4-point score (1 = not usable, 4 = clinically usable without edits) was used to rate the quality of each contour.Also, for all structures the geometric values of Dice Similarity Coefficient (DSC) and Mean Distance to Agreement (MDA) between the clinically treated contours and auto contoured structures were extracted from MIM (MIM Software Inc., Cleveland, OH).
Results: The average ± 1 standard deviation V3k was 3.1±2.8 cm3 for all 170 cases. The corresponding numbers for T1-T4 were 3.1±3.1, 4.0±2.8, 1.1±1.8, and 3.1±2.9 cm3, respectively.All models underestimated volumes by about 15% compared to the ground truth. Typically, mean HD95 was 2-3 mm, and mean MSD was below 1 mm in all cases (Table 1). Interestingly, the left-hand side masseter and medial pterygoid had smaller mean HD and HD95 for all models compared to their right- hand side counterparts. For all structures, the differences between T1-T3 (102 training cases each) were limited. The differences between T1-T3 and T4 (136 training cases) were also limited but T4 had the same or slightly better geometric metrics in 23/32 situations (8 structures x 4 metrics, Table 1).
Made with FlippingBook - Share PDF online