Use of texture analysis based on contrast-enhanced MRI to predict treatment response to chemoradiotherapy in nasopharyngeal carcinoma

Purpose To explore the clinical potential of texture analysis using contrast-enhanced 3.0T magnetic resonance imaging (MRI) for predicting the therapeutic response of nasopharyngeal carcinoma (NPC) to chemoradiotherapy. Materials and Methods The dataset comprised pretreatment T1-, T2-, and diffusion-weighted MR images from 53 eligible patients with newly diagnosed NPC. The patients were divided into two sets: the training set including 31 responders and 11 nonresponders and the testing set including eight responders and three nonresponders. The region of interest (ROI) was delineated by two radiologists for each sequence. Quantitative image parameters were extracted and statistically filtered to identify a subset of reproducible and nonredundant parameters that were used to construct the predictive model. The internal validation was performed using stratified 10-fold cross-validation in the training set and the external validation was performed in the testing set. McNemar's test was used to test the statistical difference between the performances of the extracted parameters in predicting the treatment response. Results All three parameter sets showed potential in predicting treatment response with high accuracy (T1: 0.952/0.939, T2: 0.904/0.905, diffusion-weighted [DWI]: 0.881/0.929). Supervised learning models based on parameters extracted from the T1 sequence showed better classification performance than those extracted from the T2-weighted (T2W) (artificial neural network [ANN]: P = 0.043, k-nearest neighbors [kNN]: P = 0.033) and DWI (ANN: P = 0.032. kNN: P = 0.014). No statistical difference was observed in the performance of the two classifiers (P = 0.083). Conclusion Texture analysis based on T1W, T2W, and DWI could act as imaging biomarkers of tumor response to chemoradiotherapy in NPC patients and serve as a new radiological analysis tool for treatment prediction. J. Magn. Reson. Imaging 2016;44:445–455.