Relationship of CT Radiomics and Dose Texture to Radiation-Induced Swallow Dysfunction in Head and Neck Cancer

Abstract

Recent studies have shown that radiotherapy (RT) dosimetry is associated with an increased incidence of dysphagia in head and neck cancer (HNC) patients. The purpose of this work was to develop a dysphagia model using a comprehensive atlas of swallow-related structures to determine the predictive value of imaging and dose features of HNC patients who were evaluated for dysphagia after RT. We hypothesized that imaging and dosimetry features would be significantly correlated to patient reported scores of dysphagia. CT images and dysphagia scores were prospectively collected in HNC patients. We used the 3 month post-RT patient reported Sydney Swallow Questionnaire (SSQ) that has been validated to measure pharyngeal swallow as the primary endpoint. CT images were acquired prior to RT at the time of simulation. Swallow-related structures identified as important to swallow function (e.g. ipsi/contralateral parotid and submandibular glands [iPG, cPG, iSG, cSG], middle constrictors [MPC], anterior and posterior digastrics [ADM, PDM], mylohyoid [MHM], genioglossus [GGM], hyoglossus [HM], and cricopharyngeus [CP]) were segmented. For each segmented salivary gland, high-dimensional image features were extracted from CT images using a python software package. For each swallow-related structure, high-dimensional features were extracted from the structure’s dose distribution. Major categories of extracted features included shape, first order statistics, gray level co-occurrence matrix (GLCM), gray level run length matrix (GLRLM), and gray level size zone matrix (GLSZM). Multivariate regression was performed using the Enter method to identify the predictors of SSQ. The variables considered for modeling were chosen based on statistically significant univariate relationships with SSQ (p<0.005). All statistical analysis was performed in statistical software. In 136 patients, total SSQ at 3 months post-RT was correlated with the MHM D10, GLCM informational measure of correlation of dose, and GLSZM gray level non-uniformity of dose; cSG minimum dose and D10, HM minimum dose; and MPC GLCM inverse difference of dose (r>0.30, p<0.005). In terms of CT image features, SSQ was related to cPG GLCM cluster shade and prominence, iSG GLCM difference variance, and iPG GLCM correlation (r>0.30, p<0.005). In a multivariate analysis (R=0.65, p<0.001), CT iPG GLCM correlation, iSG GLCM difference variance, MPC inverse difference of dose and MHM D10 contributed significantly to the model (p<0.05). Salivary gland image features of CT intensity heterogeneity such as the iSG difference variance as well as the dose to 10% of the MHM and homogeneity of the dose to the MPC (identified by the inverse difference of dose) were significant predictors of dysphagia post-RT. These results suggest that both salivary structure and dose to swallow muscles are important dysphagia predictors and the promise of this approach for a comprehensive RT-dysphagia model.