Pre-Radiation CT-Based Radiomic Features Predict Locoregional and Distant Failure in Locally Advanced Head and Neck Cancer

Purpose/Objective(s)

Locally advanced head and neck cancer treatment often involves aggressive multimodal therapy. While outcomes have improved in certain groups (HPV-positive disease), locoregional (LRF) and distant failure (DF) remain persistent problems. Accurate prediction of these risks would provide vital clinical insight; for instance, a patient with low LRF risk could be de-escalated, while a patient at high-risk for DF could be offered systemic options in lieu of aggressive local therapy. In this study we explore the use of CT-based radiomic features (quantitative sub-visual cues) for this purpose.

Materials/Methods

We identified patients with primary or recurrent locally advanced head/neck cancer treated between 2014-2020 who received a definitive or adjuvant RT course (naïve to RT prior). Histologies included squamous cell carcinoma (SCC), poorly differentiated carcinoma (nasopharyngeal or salivary origin), high-risk parotid malignancies, and skin SCC metastatic to the parotid/neck. Patients had at least 12 months of follow-up and charts were assessed for LRF and DF. CT simulation images were extracted from our treatment planning system, and specific contoured regions of interest (ROIs) were selected including the peritumoral region or post-surgical tumor bed (high-risk clinical tumor volume, CTV HR) and involved/elective neck levels (CTV Necks). CT scans were resampled to be isotropic and 3D radiomic features (Gabor wavelet, Hara lick, and CoLlAGe) were extracted from within CTV HR and CTV Necks ROIs. Using a Wilcoxon rank sum test, differential expression for each feature was calculated between patients experiencing treatment failure (LRF and DF) and those with no evidence of disease progression. A similar process was done with baseline clinical features including performance status, tumor stage, age, p16 status and smoking history. Where required, a chi-squared test was used for categorical clinical variables.

Results

95 patients were analyzed (45 definitive, 50 adjuvant). There were 17 LRF (17.9%) and 16 DF (16.8%) events. 304 radiomic features were extracted from CTV HR and CTV Necks (the latter was only performed for patients with bilateral neck contours). 2 features extracted from CTV HR ROIs were significantly associated with LRF and 29 features with DF (p < 0.05). 30 features extracted from CTV Necks ROIs were significantly associated with LRF and 0 features with DF (p < 0.05). There were no statistically significant clinical features associated with LRF or DF.

Conclusion

CT-based radiomic features derived from specific ROIs in simulation images were associated with both LRF and DF events, which was not the case for canonical clinical features. Interestingly, the CTV HR was more prominently associated with DF, whereas the CTV Necks was associated with LRF. Further studies with larger, more diverse datasets are needed to refine this model and provide a path for clinical validation.