Pulmonary lesion texture classification from endobronchialultrasonogram using texture analysis

Abstract

This research aims to develop a method to help distinguish the appearance of pulmonary lesions from a high-frequency sound (Endobronchial Ultrasound—EBUS) image. According to medical information, the appearance of smooth or rough texture of a lesion can significantly indicate that it is malignant or benign. In this study, the features that are used in the classification are divided into 3 groups: group 1 consists of 22 standard features, group 2 consists of the proposed features extracted from the weighted sum of the upper and lower GLCM which consists 12 features, and group 3 is the combination of group 1 and group 2. Not all the features in each group are used in the classification, only the best features are selected from each groups using three feature selection techniques: forward selection, backward selection, and genetic selection. After the best features are selected, they are entered into eight different classifiers for the classification. The overall process of the classification consists of preprocessing, window slicing, feature extraction, feature selection, and classification. The sample input consists of 89 lesion images where 55 of them are identified by the doctor as malignant and 34 of them are identified as benign. The classification results show that the highest accuracy rate of 86.517% can be obtained by using features from group 3 with genetic selection and support vector machine.