2314 - Deep Learning-Based Auto Contouring of Mandibular Sub-Volumes Based on the ClinRad System for Spatial Localization of Osteoradionecrosis of the Jaw

L. Humbert-Vidan¹, K. A. Wahid², Z. Kaffey³, S. Mirbahaeddin¹, J. Curiel¹, S. Acharya¹, A. Shekha¹, J. M. Rigert¹, C. Dede¹, S. Attia¹, R. He¹, M. Naser¹, S. Y. Lai⁴, C. D. Fuller¹, and A. C. Moreno¹; ¹Department of Radiation Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX, ²Department of Imaging Physics, The University of Texas MD Anderson Cancer Center, Houston, TX, ³University of Texas MD Anderson Cancer Center UTHealth Graduate School of Biomedical Sciences, Houston, TX, ⁴Department of Head and Neck Surgery, The University of Texas MD Anderson Cancer Center, Houston, TX

Purpose/Objective(s): Patients with head and neck cancer (HNC) treated with radiation therapy (RT) are at a lifetime risk of developing osteoradionecrosis of the jaws (ORNJ), a chronic, often debilitating, and progressive radiation-associated orofacial sequelae. Prior staging systems for ORNJ exclude radiographic features or consider the mandible as a homogeneous structure. The ClinRad system, a recently proposed risk-based model that has been endorsed by the ISOO-MASCC-ASCO Guidelines for ORNJ, incorporates radiographic elements based on mandibular sub-volumes that vary in bone composition, density, and radiosensitivity. The purpose of this study is to develop a mandibular sub-volume deep learning (DL) auto contouring model to enable the incorporation of spatial heterogeneities into ORN prediction models. Materials/

Methods: A total of 38 CT simulation scans were used for manual segmentation of 6 mandibular sub-volumes anatomically defined to differentiate alveolar from cortical bone and include laterality: the alveolar sub-volumes (region extending 1.5 cm inferiorly from the mandibular alveolar ridge), the basal body sub-volumes (region extending inferiorly from the interior boarder of the mandibular alveolar process to the border of the mandible), and the basal angle sub-volumes (region extending laterally and distally from the alveolar process to include the angle, ramus, coronoid, condylar process, and head). The mandibular symphysis served as the anatomical landmark for defining laterality of the sub-volumes. The 3D Transformer UNETR and 3D ResNet convolutional neural networks were trained and tested using a 5-fold cross-validation approach, comprising 25 training and 5 validation scans per fold, with 8 scans reserved for testing. The final Dice Similarity Coefficients (DSC) for each mandible sub-volume in the test set were calculated using a majority vote strategy for both UNETR and ResNet models.

Results: The DSC values ranged from 0.74 to 0.89 across different sub-volumes and models, indicating varying degrees of segmentation accuracy. Specifically, the basal angle sub-volumes exhibited the highest segmentation performance, with DSC values consistently above 0.88, while the alveolar sub-volumes demonstrated slightly lower but still substantial DSC scores, ranging from 0.74 to 0.82.

Conclusion: These results highlight the effectiveness of DL models for auto contouring of mandibular sub-volumes. Future work will focus on achieving equal segmentation performance across all sub-volumes and validating our models on a larger and independent dataset. Our approach enables automated spatial and dosimetric information on anatomical heterogeneities within the mandible for more effective modelling of ORNJ and development of clinically relevant decision support tools.