Application of nonsubsampled dual-tree complex contourlet transform based wavelet texture features of CT images in identification of benign and malignant pulmonary nodules / 中国医学影像技术

ABSTRACT

Objective: To explore the diagnostic value for benign and malignant pulmonary nodules using the wavelet texture features based on nonsubsampled dual-tree complex contourlet transform (NSDTCT). Methods Texture parameters based on NSDTCT and Contourlet transform were extracted from CT images of patients with pulmonary nodules. Dimension reduction of texture features was conducted with univariate analysis and Lasso regression. The support vector machine classifiers based on these texture features for benign and malignant pulmonary nodules were constructed. ROC analysis was applied to compare the two texture extraction methods. Results For NSDTCT based features, the model based on the least number of NSDTCT texture after Lasso dimension reduction was of excellent performance, with the accuracy of 98.37% in diagnosing benign and malignant lung nodules, and the AUC was 1.00. For Contourlet transform based features, the model with all extracted texture features performed well, with the accuracy of 56.05%, and the AUC was 0.73. There was significant difference of AUC of ROC curve between the two models (Z=6.430, P<0.001). Conclusion: NSDTCT texture analysis method has good performance for diagnosing lung cancer with high classification accuracy.