Establishing a Method to Detect and Predit Osteoradionecrosis using Computed Tomography

Zaphanlene Kaffey, MS (Advisor: Clifton Fuller, MD, PhD)

Osteoradionecrosis (ORN) of the mandible is a severe late complication of radiotherapy for head and neck cancer, affecting up to 20% of patients and causing chronic pain, pathologic fracture, and functional impairment that can persist for years after successful cancer treatment. Despite its clinical significance, ORN lacks objective, standardized imaging-based tools for detection, severity classification, or early risk stratification. This dissertation develops and pre-qualifies CT-based imaging biomarkers for mandibular ORN surveillance across three aims.

Aim 1 established the baseline for clinical radiographic ORN detection through a prospective blinded multi-disciplinary reader study involving 20 physicians across five specialties reviewing 85 imaging sets from 30 patients with confirmed ORN. No specialty demonstrated diagnostic performance significantly exceeding chance on CT alone after correction for multiple comparisons (AUC range 0.47 to 0.56). Paired CT and orthopantomogram imaging significantly improved performance across all specialties (AUC range 0.79 to 0.98, p < 0.001), but inter-rater agreement for ClinRad severity staging remained in the slight range, with median Fleiss kappa values of 0.22, 0.13, and 0.05 for stages 0/1, 2, and 3, respectively.

Aim 2a characterized the Hounsfield unit (HU) signature of manifest ORN in 47 patients with clinically confirmed mandibular ORN using voxel-wise HU extraction from diagnostic CT. Mean HU was significantly lower in ORN-affected bone compared to contralateral healthy bone (1545 vs. 1774 HU, p < 0.001, Cohen's d = 1.88), with 100% patient-level directionality. Voxel-level receiver operating characteristic analysis yielded AUC 0.866 (95% CI: 0.865 to 0.866), with an optimal detection threshold of 1136 HU.

Aim 2b tested whether CT radiomics extracted from post-RT surveillance imaging improved ORN prediction beyond dose-based normal tissue complication probability (NTCP) modeling in a matched case-control cohort of 260 patients (50 ORN-positive, 210 controls). Delta HU analysis did not significantly differentiate future ORN from non-ORN subregions at the surveillance timepoint (Cohen's d = -0.12, p = 0.46). However, PyRadiomics texture features extracted from the surveillance CT (3 to 12 months post-RT), combined with an institution-specific NTCP model incorporating D30%, V70Gy, dental extraction history, and smoking status, achieved AUC 0.75 compared to 0.68 for NTCP alone (p = 0.046). No significant improvement was observed at the simulation CT or diagnostic CT timepoints.

These findings establish that CT-based imaging biomarkers can detect manifest ORN with good discriminative performance and that surveillance CT radiomics significantly improves ORN prediction overdose-clinical modeling alone. This work constitutes a pre-qualification of CT imaging as a surveillance biomarker platform for mandibular ORN, supporting prospective multi-institutional validation.

Advisory Committee:

• Clifton Fuller, MD, PhD, Chair
• Laurence Court, PhD
• Judy Gichoya, MD
• Rebecca Howell, PhD
• Stephen Lai, MD, PhD
• Jullian Pollard-Larkin, PhD