Author: Rodrigo Delgadillo, Nesrin Dogan, Benjamin J. Rich, Stuart E Samuels, Levent Sensoy 👨🔬
Affiliation: University of Miami Sylvester Comprehensive Cancer Center 🌍
Purpose: Daily Cone beam CT (CBCT) images may be useful in detecting early morphological changes during head and neck cancer radiotherapy. The aim of this study was to evaluate the performance of CBCT-based delta-radiomics features in predicting patient toxicity in the supraglottic tissue causing dysphagia, thereby enabling radiation oncologists to personalize dosing strategies.
Methods: Forty-three human papillomavirus positive (HPV+) oropharyngeal cancer patients from an IRB-approved database were retrospectively selected. All patients were treated with a definitive dose of 70 Gy in 35 fractions using the same modality and imaging platform. First and second order statistical radiomics features were extracted from supraglottic volume on daily CBCT for each fraction and averaged over five-fraction increments, representing weekly treatment. Patient toxicity outcomes were scored using treatment follow-up examinations, and in some cases, speech therapists’ exam reports. The original scoring represented none, mild, moderate, or severe toxicity. For the study, the patients were then divided into two groups: Tolerable (none, mild toxicity) and significant (moderate, severe toxicity) levels. The Random Forest model was used to estimate the correlation between radiomics features and toxicity. Logistic regression model was used for classification and area under the receiver operating characteristic curve (AUC) and a confidence interval was calculated.
Results: Several radiomics features were significant predictors of dysphagia and swallowing discomfort. Moderate to severe toxicity occurred in 16 patients. Spearman correlation coefficients were used to eliminate similar features, only retaining the most important of the correlated features. Four radiomics features were chosen as significant predictors for toxicity. These features were global skewness, geometric volume, NGTDM Busyness-VNF, and GLRLM RLN-VN. Our logistic regression outcome model yielded an AUC value of 0.73 (95% confidence interval 0.62-0.80).
Conclusion: CBCT-based radiomics features within the supraglottis show clinically relevant and promising results in predicting radiation-induced swallowing toxicities for patients with HPV+ oropharyngeal carcinoma.