2831 - Automated vs. Manual Salivary Gland Segmentation for Xerostomia Risk Assessment in Head and Neck Cancer Patients from the ARTSCAN III trial
Presenter(s)
V. Rogowski1,2, C. Jamtheim Gustafsson2,3, M. Gebre-Medhin4, A. Anghede Haraldsson2, and P. Munck af Rosenschold1,2; 1Medical Radiation Physics, Department of Clinical Sciences Lund, Lund University, Lund, Sweden, 2Radiation Physics, Department of Hematology, Oncology, and Radiation Physics, Skåne University Hospital, Lund, Sweden, 3Department of Translational Medicine, Medical Radiation Physics, Lund University, Malmö, Sweden, 4Oncology, Department of Hematology, Oncology, and Radiation Physics, Skåne University Hospital, Lund, Sweden
Purpose/Objective(s): We evaluated the impact of deep learning (DL)-based auto-segmentation of salivary glands in radiation therapy planning, and its relationship to xerostomia of grade II-III for head and neck squamous cell cancer (HNSCC) patients from the prospective randomized ARTSCAN III trial. The aim of this research was to assess both the geometric accuracy of DL segmentation and to compare its performance against manual segmentation for xerostomia outcome modeling in multi-institutional trials.
Materials/Methods: A total of 298 patients with HNSCC were selected for analysis, with 217 eligible for xerostomia normal tissue complication probability (NTCP) modeling. Parotid and submandibular glands were delineated both manually as part of the trial protocol and automatically using a commercial DL system. Geometric concordance was quantified using Dice Similarity Coefficient (DSC) and 95th percentile Hausdorff Distance (HD95). Dose and volume differences between manual and DL segmentations were assessed for statistical significance using the Wilcoxon test (p<0.05). Univariate and logistic regression models for grade II-III xerostomia were evaluated for ipsi-, contralateral and average parotid gland doses, as well as submandibular gland doses, baseline xerostomia score, age, gender, and randomized treatment arm. Multivariate logistic modelling with forward conditional variable selection was developed to model grade II-III severity of xerostomia at 12 months post-treatment.
Results: DL segmentations demonstrated strong geometric concordance with manual delineations, achieving mean DSC values of 0.81 for both parotid and submandibular glands. Inter-institutional analysis revealed variations in segmentation performance, with DSC values ranging from 0.73 to 0.86 across treatment centers. The statistical analysis demonstrated significant volumetric differences between DL and manual segmentations (p<0.001), with consistently larger volumes generated by DL compared with manual delineations. The DL-segmentations showed a stronger correlation with xerostomia outcomes and were used henceforth. Only ipsi-, contralateral and average dose to both parotid glands were significant variables in univariate logistic modelling, with all other variables non-significant. In multivariate logistic regression models, the average dose to the contralateral parotid remained significant, with a D50 of 19.85 Gy and ?50 of 0.33.
Conclusion: This study demonstrates that DL-based auto-segmentation provides a robust alternative to manual delineation in multi-institutional settings. The DL-based segmentation demonstrated a stronger association with xerostomia outcomes compared to manual segmentation. The contralateral parotid gland mean dose emerged as the most significant predictor in multivariate models. These finding underscore the potential use of DL segmentations to enhance NTCP modeling consistency across multi-institutional settings.