Predicting Radiotherapy Response in Head and Neck Patients Using Quantitative Ultrasound

Abstract

In the ever-evolving cancer management systemic neoadjuvant, chemo, radiation and now image guided radiation therapy have become frequently used options. However, discordance between therapy and desired response continues to remain a significant challenge due to an unavailability of feasible techniques that can predict the efficacy of the treatment, prior to or during the course of therapy. In this study we demonstrate that pretreatment non-invasive Quantitative Ultrasound Spectroscopy (QUS)radiomic markers in metastatic lymph nodes can be used as response guide. Imaging of spatial molecular events by QUS could enable locoregional control with radiation therapy and optimize outcomes. The importance of this work includes the development of ultrasound imaging techniques used to predict head and neck cancer response to (chemo)radiotherapy. This study examined quantitative ultrasound (QUS)spectroscopy in metastatic lymph nodes in head and neck (HN +/- concomitant chemotherapy)and response was evaluated using magnetic resonance imaging (MRI)three months after treatment. Pretreatment QUS imaging of metastatic lymph nodes were acquired, and spectral analyses were used to determine QUS radiomic features. Second-order statistical analyses included applying a pixel-to-pixel grey-level co-occurrence matrix (GLCM)to QUS parametric maps to determine texture-based features. QUS radiomic modelling was completed using machine learning, in order to determine which parameters performed best for predicting patients according to their clinical responses. Our findings suggested significant QUS-radiomic parameter differences between complete responders (CR)versus partial responders (PR), which included the spectral slope (SS)-correlation, spectral intercept (SI)-contrast, SI-correlation, SI-homogeneity, and the average scatterer diameter (ASD)-correlation (p<0.05). The classification. Performance of univariate features indicated that the SI-contrast performed the best classification (AUC=0.741). For bivariate features, the SI-con + SI-hom demonstrated an AUC of 0.870. The SS-cor + SI-con + SI-hom combined parameter demonstrated the highest accuracy in predicting treatment response a priori with an accuracy of 87.5%.