Peripheral pulmonary lesion classification from endobronchial ultrasonography images using weight-sum of upper and lower GLCM feature

Abstract

This paper aims to classify a peripheral pulmonary lesion whether it is malignant or benign by proposing the new method to select a window of interest (WOI) using window slicing and the new feature called the "weight-sum of upper and lower gray level co-occurrence matrix (GLCM)" of an endobronchial ultrasound (EBUS) image. The proposed feature can be used to determine the heterogeneity of pulmonary lesion which is one of the most important characteristics of lung cancer. The proposed feature is used as input into three different feature selection methods and three different classifiers for lesion classification. In order to evaluate the classification, a set of 89 EBUS images were used as a sample set. The classifications were performed three times with three different sets of features that were extracted from sample images using the same classification process. The first set of features consists of only standard features which are mean, contrast, homogeneity, correlation, entropy, and energy. The second set of features consists of the proposed feature, and the last set of features consists of both standard features and the proposed feature. The classification results show that using genetic selection as feature selection method with support vector machine as classifier with only the proposed features as input data gives the highest accuracy rate. The statistical results show that the accuracy, the sensitivity, and the specificity are 84.27%, 82.53%, and 88.46%, respectively.