S559
Clinical – Head & neck
ESTRO 2026
2013;85(2):415-420. Humbert-Vidan L, et al. IMPACT: a deep learning tool for mandibular parcellation and ORN assessment. Radiother Oncol. 2025;196:110304. Schreiber JJ, et al. Hounsfield units for bone mineral density assessment. J Bone Joint Surg Am. 2011;93(11):1057-1063. Annane D, et al. Hyperbaric oxygen for jaw radionecrosis: ORN96 trial. J Clin Oncol. 2004;22(24):4893-4900. Keywords: osteoradionecrosis, Hounsfield Unit,
basal, and ramus/condyle for left and right sides). On the ORN diagnostic CT, voxels were classified as ORN- positive (within clinical contour) or ORN-negative (remaining mandible). Cross-sectional analysis (n=47) compared HU distributions between ORN-positive and ORN-negative regions using paired statistical tests, controlling for systemic factors. ROC analysis determined discriminatory thresholds. Longitudinal analysis (n=45) calculated mean delta HU for each IMPACT subregion from intermediate follow-up to ORN diagnosis. Spatial concordance assessed whether subregion(s) with ORN involvement matched those with greatest HU decline, quantified using Cohen's kappa. Results: ORN-positive regions demonstrated significantly lower mean HU compared to ORN-negative regions (1545 ± 139 vs 1774± 70HU, p < 0.001 by both paired t-test and Wilcoxon signed-rank test, Cohen's d = 1.88) (Figure 1A). HU reduction was observed in 31/47 patients (66%) using a >5% decrease threshold (Figure 1C) (mean fold-change = 0.87, representing 13% HU decrease). Voxel-level ROC analysis (n=2,847,056) yielded AUC = 0.866 (95% CI: 0.865-0.866) with an optimal threshold of 1136 HU (sensitivity 85%, specificity 78%) (Figure 1B). Longitudinal analysis revealed no significant trajectory differences between eventual ORN and non-ORN regions (delta HU: -64 ± 152 vs -45 ± 177, p = 0.46, d = -0.12), with low spatial concordance (40% agreement), indicating ORN-specific radiographic changes manifest near clinical diagnosis rather than gradually over preceding months. Conclusion: ORN-affected bone demonstrates a distinct radiographic profile, with mean HU values 228 units (13%) lower than healthy bone. Temporal analysis indicates changes manifest near clinical diagnosis rather than gradually, suggesting ORN develops as a late-stage threshold phenomenon. This supports intensive surveillance imaging in months immediately preceding expected ORN onset rather than uniform long-term monitoring. Quantitative HU thresholds enable automated detection of ORN-affected bone regions on post-treatment surveillance imaging, supporting development of clinical decision support tools. These imaging biomarkers may guide targeted interventions such as hyperbaric oxygen therapy, antifibrotic agents, or surgery. References: Tsai CJ, et al. Osteoradionecrosis and radiation dose to the mandible. Int J Radiat Oncol Biol Phys.
Digital Poster Highlight 1197
Personalised prediction of occult nodal involvement in oral cavity squamous cell carcinoma using hidden Markov models
Noemi Bührer 1 , Roman Ludwig 1 , Yoel Pérez Haas 1 , Sandrine Werlen 2,3 , Dorothea Barbatei 4 , Olgun Elicin 5 , Matthias Dettmer 6,7 , Philippe Zrounba 8 , Roland Giger 2,3 , Vincent Grégoire 4 , Adrian Schubert 2,3 , Oliver Riesterer 9 , Sonja Stieb 9 , Annika Sütsch 9 , Panagiotis Balermpas 1 , Jan Unkelbach 1 1 Department of Radiation Oncology, University Hospital Zurich, Zurich, Switzerland. 2 Department of ENT, Head and Neck Surgery, Inselspital, Bern University Hospital, University of Bern, Bern, Switzerland. 3 Head and Neck Anticancer Center, Bern University Hospital, University of Bern, Bern, Switzerland. 4 Department of Radiation Oncology, Centre Léon Bérard, Lyon, France. 5 Department of Radiation Oncology, Bern University Hospital, University of Bern, Bern, Switzerland. 6 Institute of Pathology, Klinikum Stuttgart, Stuttgart, Germany. 7 Institute of Tissue Medicine and Pathology, Bern University Hospital, University of Bern, Bern, Switzerland. 8 Department of Head and Neck Surgery, Centre Léon Bérard, Lyon, France. 9 Center for Radiation Oncology, Cantonal Hospital Aarau, Aarau, Switzerland Purpose/Objective: The elective clinical nodal target volume (CTV-N) in head and neck squamous cell carcinoma (SCC) is currently defined based on prevalence of involvement by lymph node level (LNL) [1]. This approach often recommends extensive irradiation of both sides of the neck, lacking personalisation to the patient-specific risk depending on the individual clinical diagnosis. Previous work presented a hidden Markov model (HMM) of bilateral metastatic progression in oropharyngeal SCC [2], allowing for risk estimation of microscopic involvement based on a patient-specific diagnosis. This study extends the model and its personalised risk predictions to oral cavity SCC. Material/Methods: The HMM treats each LNL as a binary random variable (healthy or involved). Metastatic progression is
Made with FlippingBook - Share PDF online