Texture Analysis Based On Intravoxel Incoherent Motion Dwi For Stratification Of The Clinical Stages Of Nasopharyngeal Cancer

The objective of the study was to investigate the utility of texture analysis based on intravoxel incoherent motion diffusion-weighted imaging (IVIM-DWI) for stratifying the clinical stages of nasopharyngeal cancer (NPC). Ninety NPC patients were stratified into low and high clinical stage groups based on the American Joint Committee on Cancer (AJCC) and TNM staging system. Texture features of the primary NPC lesion were extracted from IVIM-DWI parametric maps. The Fisher coefficient (Fisher), probability of classification error and average correlation (POE+ACC), mutual information coefficients (MI), and the combination of the above three methods (FPM) were applied to select texture features. Subsequently, each texture feature subset was analyzed via raw data analysis (RDA), principal component analysis (PCA), linear discriminant analysis (LDA) and nonlinear discriminant analysis (NDA), and the misclassification probability in stratifying the NPC clinical stages was calculated correspondingly. The best discrimination of the AJCC stage of NPC was obtained with the f map and the FPM method combined with NDA classifier, which had the lowest misclassification probability of 1.25%. The best discrimination of T stage was obtained with the FPM method combined with NDA classifier of the ADC or f map, which had equal lowest misclassification probabilities of 9.03%. The best separation of the N stage was also obtained with the FPM method combined with the NDA classifier of the ADC map, which had the lowest misclassification probability of 1.25%. Significant differences were observed in the lowest misclassification probability with the four classification methods, but not with the four IVIM-DWI parametric maps in predicting the AJCC, T and N stages. Texture analysis based on IVIM-DWI may be valuable in the pretreatment stratification of NPC clinical stages, and the FPM feature selection method combined with NDA classifier may provide the best discrimination performance.