ESTRO 2025 - Abstract Book

S2500

Physics - Autosegmentation

ESTRO 2025

Oncology, Princess Margaret Cancer Centre, Toronto, Canada. 6 Dental Oncology, Princess Margaret Cancer Centre, Toronto, Canada. 7 Head and Neck Surgery, The University of Texas MD Anderson Cancer Center, Houston, USA

Purpose/Objective: The ClinRad system 1 , a new classification for osteoradionecrosis of the jaws, includes radiological assessment of bone damage vertical extent, enhancing early detection of mandible or maxilla damage, even in cases with intact mucosa. Bone density and composition influence radiosensitivity, with the alveolar process—rich in spongy bone— being more radiation-sensitive than the denser basal region. Radiation dose distribution to these bones is uneven, making average dose reporting insufficient to capture sub-volume irradiation. Tooth-specific dose assessments could inform safer pre- and post-radiotherapy dental procedures. Image-based data, like CT and dose maps, allow spatial modeling of dose distributions, but auto-segmentation of orodental structures is lacking and essential for sub-volume dose analysis. This study aimed to develop a comprehensive deep learning (DL)-based auto segmentation pipeline encompassing individual teeth, mandible, and maxilla sub-volumes, accounting for both bone composition and laterality. Material/Methods: Individual teeth were manually contoured and numbered following the Universal Numbering System, mandible and maxilla sub-volumes included left/right/central alveolar and basal bone regions with laterality defined by teeth sextants (Figure 1). The RESNET and UNETR architectures (mandible and maxilla) and a two-stage UNETR-based architecture with center of mass detection (teeth) were trained (40 cases) and tested (9 cases). The mean dose (Dmean) values within each segmented class were compared between manual and predicted segmentations.

Results: The UNETR model outperformed the RESNET model across both basal and alveolar regions of the mandible and maxilla, with highest performance obtained on the mandible basal region sub-volumes, particularly the basal central sub-volume with a mean Dice of 0.82 (0.71-0.90). For the maxilla, segmentation of the basal region sub volumes was challenging with all models, as these sub-volumes were largely non-existing or very small. Teeth segmentation performance was best in the central regions with a mean Dice of 0.84 and 0.79 for the upper (6-11) and lower (22-27) teeth, respectively, but was challenging in commonly missing teeth (e.g., molars). No statistically significant difference was observed for Dmean values in any of the sub-volumes or teeth, with average paired two sided t-test p-values of 0.62 (mandible), 0.28 (maxilla) and 0.39 (teeth). Conclusion: The developed DL-based auto-segmentation pipeline achieved robust segmentation performance across mandible and maxilla sub-volumes, especially in the mandible’s basal regions and central teeth regions. This pipeline shows promise in enabling precise sub-volume dose analysis, potentially improving perio-radiotherapy decision-making by