Enhanced CT Based Texture Analysis and Machine Learning for Differentiation between Adenolymphoma and Mixed Tumors of the Parotid Gland

Abstract

Objective: To explore the application of enhanced computed tomography (CT)-based texture analysis combined with machine learning in the differential diagnosis of adenolymphomas and mixed tumors of the parotid gland. Methods: We retrospectively analyzed 21 and 19 cases of adenolymphomas and mixed tumors of the parotid gland, respectively. Regions of interest (ROI) were chosen on axial enhanced-CT images of the tumor’s maximum cross section using the Mazda software. The optimal texture parameters were selected using Fisher’s coefficient, probability of classification error and average correlation coefficients, mutual information, and a combination of the three. Receiver operating characteristic (ROC) curve analysis was performed to evaluate the diagnostic efficacy. Finally, the texture parameters were classified and analyzed using the following four machine-learning methods: raw data analysis, principal component analysis, linear discriminant analysis, and nonlinear discriminant analysis (NDA). The diagnostic efficiencies of these classification algorithms were analyzed. Results: WavEnHH_s-4, GrVariance, 45dgr_Fraction, WavEnLL_s-4, and GrSkewness were the statistically significant texture feature parameters for differentiating between parotid adenolymphomas and mixed tumors.. ROC curve analysis revealed that WavEnLL_s-4 had a balanced sensitivity and specificity, and the area under the curve, sensitivity, and specificity were 0.797, 84.2%, and 62.5%, respectively. The misclassification rate of NDA (5.0%–12.5%) was lower than that of the other algorithms. The NDA of FPM had the lowest misclassification rate (5.0%); its accuracy, sensitivity, specificity, positive predictive value, and negative predictive value were 95.0%, 95.2%, 94.7%, 95.2, and 94.7, respectively. Conclusion: The optimum enhanced CT-based texture features differed significantly between parotid adenolymphomas and mixed tumors. A combination of FPM and NDA had the lowest misclassification rate; it can contribute toward the identification of parotid adenolymphomas and mixed tumors.

Keywords

• enhanced ct
• machine learning
• texture analysis
• parotid tumor