Performance testing of several classifiers for differentiating obstructive lung diseases based on texture analysis at high-resolution computerized tomography (HRCT)

Machine classifiers have been used to automate quantitative analysis and avoid intra-inter-reader variability in previous studies. The selection of an appropriate classification scheme is important for improving performance based on the characteristics of the data set. This paper investigated the performance of several machine classifiers for differentiating obstructive lung diseases using texture analysis on various ROI (region of interest) sizes. 265 high-resolution computerized tomography (HRCT) images were taken from 92 subjects. On each image, two experienced radiologists selected ROIs with various sizes representing area of severe centrilobular emphysema (PLE, n=63), mild centrilobular emphysema (CLE, n=65), bronchiolitis obliterans (BO, n=70) or normal lung (NL, n=67). Four machine classifiers were implemented: naïve Bayesian classifier, Bayesian classifier, ANN (artificial neural net) and SVM (support vector machine). For a testing method, 5-fold cross-validation methods were used and each validation was repeated 20 times. The SVM had the best performance in overall accuracy (in ROI size of 32x32 and 64x64) (t-test, p<0.05). There was no significant overall accuracy difference between Bayesian and ANN (t-test, p<0.05). The naïve Bayesian method performed significantly worse than the other classifiers (t-test, p<0.05). SVM showed the best performance for classification of the obstructive lung diseases in this study.