Diffusion kurtosis imaging to predict tumor hypoxia and recurrence in HPV-associated oropharyngeal cancer

Introduction: A recent phase II trial demonstrated the efficacy of hypoxia-guided chemoradiation (CRT) de-escalation in human papillomavirus-associated oropharyngeal cancer (HPV+ OPC). 18F-fluoromisonidazole (18F-FMISO) positron emission tomography (PET), used to measure hypoxia, is not widely available, limiting use of hypoxia as an imaging biomarker in research and clinical care. Diffusion kurtosis imaging (DKI) quantifies non-Gaussian water diffusion, especially at higher b-values of diffusion weighted (DW) data. DKI captures microstructural complexity, reflects integrity of cell membranes and cellularity, and may serve as a surrogate for hypoxia. Methods: DW data were obtained from a phase II trial of hypoxia-guided CRT de-escalation in patients with T0-2/N1-N2c HPV+ OPC. Patients received pre-treatment 18F-FMISO imaging and serial MRI scans from pre-treatment through week 4 of CRT. Longitudinal nodal volume was collected. DKI-derived features Dapp and Kapp (mean values, standard deviation, and skewness) from modeling of the multiple b-values DW data were extracted. Associations with pre-treatment hypoxia and nodal recurrence were evaluated using Wilcoxon rank-sum tests. Random forest (RF) models integrating volume and DKI-derived features were developed to predict hypoxia and recurrence, with feature importance assessed using SHAP values. Results: 90 of 158 trial patients with longitudinal DW-MRI exams were included. Pre-treatment 18F-FMISO PET demonstrated normoxia in 25 patients (27.8%) and hypoxia in 65 (72.2%). Pre-treatment tumor volume was significantly higher in hypoxic tumors than normoxic tumors (mean 25.7 vs. 11.3 cm³, p < 0.001), while pre-treatment DKI-derived features did not differ significantly. Eight patients had nodal recurrence, all in the de-escalated cohort. An RF model incorporating pre-treatment DKI features and tumor volume predicted pre-treatment hypoxia with good discrimination (AUC = 0.77), outperforming a model using volume alone (AUC = 0.65, p<0.001). SHAP analysis identified tumor volume, Dapp skewness, Dapp mean, and Kapp mean as the most meaningful predictors. Longitudinal tumor volume in weeks 2-4 was significantly higher in patients who recurred, while longitudinal DKI features did not differ significantly. The model using week 3 volume and DKI features had the best performance (AUC = 0.79). SHAP analysis identified volume, volume change from pretreatment to week 3, Kapp skewness, and mean Dapp as the most meaningful predictors of recurrence. Conclusion: Changes in cell architecture and metabolism caused by hypoxia may be associated with restricted movement of water molecules in complex tissues. DKI and volumetric measurement demonstrated good predictive ability for both hypoxia and recurrence risk in HPV+ OPC. These findings support further prospective validation of DKI for assessing tumor hypoxia and recurrence risk.