Development and Validation of a Radiomics Signature for Predicting Early Local Failure After Organ Preservation Treatment of Hypopharyngeal Cancer

Abstract

To develop and validate a radiomics signature for prediction of early local failure in patients with hypopharyngeal squamous cell carcinoma (HPSCC) receiving organ preservation therapy (OPT). We used least absolute shrinkage and selection operator (LASSO) regression model to select and develop the radiomics signature derived from 801 radiomics features which were extracted by analyzing visualized primary tumors on T2-weighted and non-contrast / contrast enhanced T1-weighted magnetic resonance imaging (MRI) images of a primary cohort enrolling 92 patients with HPSCC from January 2003 to April 2010. An independent validation cohort containing 25 consecutive patients from May 2009 to December 2014 was used to confirm the prediction performance of the identified radiomics signature. Univariate and multivariate analysis were further performed using logistic regression and cox regression model to compare the performance for predicting the 1-year local control and overall survival with several potential prognostic factors, including age, gender, radiomics signature, clinical TMN stage, smoking and alcohol behavior. The radiomics signature, consisting of 17 image features, performed well in prediction of 1 year HPSCC local control both in the primary cohort and validation cohort. The areas under the curve of the receiver operating characteristic curve calculated using the radiomics signature were 0.846 and 0.780, respectively. The best cut-off point of radiomics signature score for predicitng local control in the entire cohort is 0.27, with the sensitivity, specificity, and accuracy of 60.7%, 97.4%, and 0.79, respectively. The univariate and multivariate analysis revealed only clinical nodal stage (HR: 5.6; p = 0.024) and radiomics signature score more than 0.27 (HR: 21.3; p<0.001) as significant predictors for 1 year local failure, whereas only radiomics signature more than 0.27 (HR: 2.663; p<0.001) as a significant factor for overall survival. Our results indicate that a radiomics signature according to quantitative image analysis of MRI is a novel and useful approach for predicting early local failure in HPSCC patients who received OPT. This new approach will help us select the optimal treatment modality for HPSCC patients.